OF THE














                         8:34 a.m.


                  Tuesday, April 22, 2003
















                      Conference Room

                     5630 Fishers Lane

                Food and Drug Administration

                 Rockville, Maryland  20857





JURGEN VENITZ, M.D., PH.D., Acting Chair

Associate Professor

Department of Pharmaceutics

Virginia Commonwealth University

School of Pharmacy

Box 980533

410 North 12th Street

Richmond, Virginia  23298-0533


KATHLEEN REEDY, R.D.H., M.S., Executive Secretary

Advisors and Consultants Staff (HFD-21)

Center for Drug Evaluation and Research

Food and Drug Administration

5600 Fishers Lane

Rockville, Maryland  20857



Associate Clinical Professor

University of California, San Diego

9500 Gillman Drive

La Jolla, California  92093



Professor, Department of Pharmaceutics

University of Florida College of Pharmacy

P.O. Box 100494, Health Science Center

Gainesville, Florida  32610-0494



Professor, Departments of Pharmacology

  and Medicine

Indiana University School of Medicine

Division of Clinical Pharmacology

Wishard Memorial Hospital WP OPW 320

1001 West 10th Street

Indianapolis, Indiana  46202



Professor, Department of Pharmaceutics

State University of New York at Buffalo

School of Pharmacy

Buffalo, New York  14260

                   ATTENDEES  (Continued)





Professor and Division Chief

Pharmacology and Toxicology

Children's Mercy Hospital

2401 Gillham Road

Kansas City, Missouri  64143



Member, Pharmaceutical Sciences

St. Jude Children's Research Hospital

332 North Lauderdale, Room D-1052

Memphis, Tennessee  38105



Chair, Department of Pharmacology

College of Medicine and Public Health

Ohio State University

5072 Graves Hall, 333 West 10th Avenue

Columbus, Ohio  43210



Professor, Laboratory Medicine

University of California, San Francisco

Box 0626, C-255

San Francisco, California  94143


MARC SWADENER, ED.D., Consumer Representative

2235 Dartmouth Avenue

Boulder, Colorado  80305-5207






Uppsala University


                   ATTENDEES  (Continued)










                      C O N T E N T S


AGENDA ITEM                                            PAGE



    by Ms. Kathleen Reedy                                 7



    by Dr. Larry Lesko                                   10






    Introduction - by Dr. Peter Lee                      19


    Example 1 - by Dr. Nhi Nguyen                        26


    Example 2 - by Dr. Jenny Zheng                       42


    Committee Discussion                                 61


    Example 3 - by Dr. He Sun                            82


    Committee Discussion                                 93


OPEN PUBLIC HEARING                                     116


    Example 4 - by Dr. Mats Karlsson                    118


    Committee Discussion                                141






    Introduction - by Dr. Larry Lesko                   150


    Pediatric PPK Template - by Dr. Peter Lee           158


    Committee Discussion                                165


    Pediatric Database Analysis - Dr. Gene Williams     188


    Committee Discussion                                203

                   P R O C E E D I N G S

                                                (8:34 a.m.)

            DR. VENITZ:  I'd like to call the meeting to order, please.

            Welcome, everybody.  This is the Clinical Pharmacology Subcommittee meeting.  We have a full agenda, as you can tell.

            I'd like to open the meeting by introducing the individuals around the table, maybe starting with Dr. Derendorf, please.

            DR. DERENDORF:  Hartmut Derendorf, University of Florida.

            DR. CAPPARELLI:  Edmund Capparelli, University of California, San Diego.

            DR. FLOCKHART:  Dave Flockhart from Indiana University.

            DR. SHEINER:  Lewis Sheiner, University of California, San Francisco.

            DR. SWADENER:  Marc Swadener, Boulder, Colorado.

            MS. REEDY:  Kathleen Reedy, Food and Drug Administration.

            DR. VENITZ:  Jurgen Venitz, Virginia Commonwealth University.

            DR. JUSKO:  William Jusko, University at Buffalo.

            DR. KEARNS:  Greg Kearns, University of Missouri, Kansas City.

            DR. RELLING:  Mary Relling, St. Jude Children's Research Hospital in Memphis.

            DR. SADEE:  Wolfgang Sadee, Ohio State University.

            DR. LESKO:  Larry Lesko, Office of Clinical Pharmacology and Biopharmaceutics at FDA.

            DR. LEE:  Peter Lee, FDA.

            DR. VENITZ:  Thank you.

            Our next order of business is the conflict of interest statement.  Kathleen Reedy will read the conflict of interest statement.

            MS. REEDY:  Acknowledgement related to general matters waivers, Clinical Pharmacology Subcommittee of the Advisory Committee for Pharmaceutical Science, April 22, 2003, an open session.

            The following announcement addresses the issue of conflict of interest with respect to this meeting and is made a part of the record to preclude even the appearance of such at this meeting.

            The topics of this meeting are issues of broad applicability.  Unlike issues before a committee in which a particular product is discussed, issues of broad applicability involve many industrial sponsors and academic institutions.

            All special government employees have been screened for their financial interests as they may apply to the general topics at hand.  Because they have reported interests in pharmaceutical companies, the Food and Drug Administration has granted general matters waivers to the following SGEs which permits them to participate in these discussions:  Dr. Edmund Capparelli, Dr. William Jusko, Dr. Gregory Kearns, Dr. Howard McLeod, Dr. Wolfgang Sadee, Dr. Lewis Sheiner.

            A copy of the waiver statements may be obtained by submitting a written request to the agency's Freedom of Information Office, room 12A-30 of the Parklawn Building.

            In addition, Dr. Hartmut Derendorf, Dr. David Flockhart, Dr. Mary Relling, and Dr. Marc Swadener do not require special matters waivers because they do not have any personal or imputed financial interests in any pharmaceutical firms.

            Because general topics impact so many institutions, it is not prudent to recite all potential conflicts of interest as they apply to each member and consultant.

            FDA acknowledges that there may be potential conflicts of interest, but because of the general nature of the discussion before the committee, these potential conflicts are mitigated.

            With respect to FDA's invited guest speaker, Dr. Mats Karlsson reports that he has contracts and/or grants with AstraZeneca, Oasmia, Pfizer, and Servier.  He also receives consulting fees from AstraZeneca, Ferring, Lilly, Pfizer, and Roche; and speaker fees from Johnson & Johnson and NovaNordisk.

            In the event that the discussions involve any other products or firms not already on the agenda for which FDA participants have a financial interest, the participants' involvement and their exclusion will be noted for the record.

            With respect to all other participants, we ask in the interest of fairness that they address any current or previous financial involvement with any firm whose product they may wish to comment upon.

            DR. VENITZ:  Thank you, Kathleen.

            Before we proceed to the official business of today, I'd like to welcome a few new committee members.  Dr. Swadener to my right is the consumer representative who also serves on the Advisory Committee for Pharmaceutical Science.  We've got Dr. Shek who couldn't make it today who is the industry representative, also serving on the Advisory Committee for Pharmaceutical Science.  And Drs. D'Argenio and Davidian who couldn't make it today.

            I would also like to thank the two outgoing members of the committee, Dr. Lalonde and Dr. Hale, as well as Dr. Jusko to my left, for chairing the previous committee meeting.

            With that said, I'd like to turn over the meeting to Dr. Lesko who is going to introduce the topics for the next day and a half.  Larry.

            DR. LESKO:  Thank you, Jurgen.

            Well, good morning, everybody and again welcome back to Rockville.  This is the second Clin Pharm Subcommittee meeting.  We had our first back six months ago, on October 22-23, and as reflect back on that meeting, it was an extremely productive meeting for us.  The advice we received at that time was excellent, and we've been thinking about it since then as we move forward with the initiatives we introduced at that first meeting.

            I do want to say thanks again.  As I look around the room, I recognize everyone here as very busy in their own work, and taking time to come to Washington and participate with us is extremely important.  I think you'll find that the mission that we have for this committee is a noble one relating to some exciting times in the area of drug development generally and clinical pharmacology specifically.

            Let me talk a little bit about today's meeting.  Let me start with what's new since we last met in October, and there are really three exciting things that are new that really impact the topics that we'll talk about today.

            The first is an FDA-wide announcement that our Commissioner made back in January called Improving Innovation in Medical Technology Beyond 2002.  What this initiative entails is quite lengthy.  There are several goals, but basically it revolves around improving the process, including the drug development process, for bringing medical innovations, treatments, and devices to the marketplace as quickly as possible that would benefit public health.

            A second part of that, however, is improving the review process at FDA through a quality systems approach.  This is the goal that we are going to sort of use to couch today's topics because a quality systems approach, if I think of it in terms of goals, has the goals I have on the slide basically.

            Application of advances in science.  Well, what are those advances in science?  They're clinical trial design.  They're clinical pharmacology study designs, statistical approaches, modeling and simulation, use of dose response in PK/PD in interesting ways.

            Use of new technology.  Use of new technology embraces things like the quantitative methods we're going to talk about today.  It embraces the integration of pharmacogenetics into drug development and review.

            Rigorous analytic reviewer tools.  This is the how to do it.  What are the tools that we can make available to our reviewers to achieve this quality systems approach?  We'll be talking about one of those in the first topic.

            And finally, the overall goal of this initiative is to provide for high quality reviews.  That's translated into effective reviews, efficient reviews, and consistent reviews.

            Initiative number two that's occurred since the last time we met is an initiative under our Prescription Drug User Fee Act, PDUFA.  It's the premarketing risk assessment initiative.  We only recently had our first public meeting having to do with the risk assessment initiative, and Bob Meyer, our ODE II director, defined risk assessment as the process of identifying, estimating, and evaluating the nature and severity of risk of a drug product.  You'll see some links between this and the first topic in today's presentations.

            In that meeting on April 9th, Bob Temple described the ideal safety database that we ought to be striving for, and that included a complete characterization of the clinical exposure-response relationship certainly for efficacy, but also for drug safety.  And he made the point that this is important for making decisions about dosing adjustments, particularly in many of the clinical pharmacology studies when exposure goes up for a variety of reasons.

            He recommended a good search for individualization factors.  This obviously involves studying individual plasma drug levels, and he made the point that it's always necessary to assess polymorphic drug metabolism, and you'll see that this relates to one of our topics in this meeting.

            Finally, he touched upon pediatrics, an important topic, and made the point that these pose special issues related to dose, PK, and PD, and we'll be touching upon that as well today.

            The third initiative that's been launched since October is the one that relates to our FDA Science Board.  We had this meeting last week on the same day as our risk assessment meeting.  The theme of the Science Board meeting was integrating scientific advances into regulation with the emphasis on pharmacogenetics.  Dr. Woodcock made a presentation at the Science Board and stated that genetic contributions to variability in toxicity include differences in drug metabolism, for example, thiopurine methyltransferase, which was a topic of our last meeting, but more broadly recommended that we look at the use of genetic tests for metabolizer status to predict dosing.

            So think of those three broad initiatives, and I hope it gives you a context for the discussion that we're going to have today and tomorrow.

            We have four proposals, four topics on the agenda.  The first is a proposal that we initiated our discussion in October and we've refined the proposal and we'll present with examples today the idea of a standardized approach to quantitate the impact that changes in exposure, related to efficacy or safety, result from changes in PK caused by a variety of intrinsic and extrinsic factors.

            What we're trying to accomplish with this proposal and standardized approach relates to what I mentioned about the Commissioner's initiative, quality systems and review.  We want to achieve a rational scientific basis for dosing adjustment that's quantitative and that links to the assessment of risk.

            We have a goal of identifying individualization factors, and through our examples today you'll see some of those factors.

            And finally at the end of the day, we hope to develop a standardized method that would rely on many different tools, but a standardized thinking process that we hope would bring consistency to the label recommendations that we make in terms of dosing adjustments.

            Our topic number two is pediatrics, and again going back to October, we opened the discussion of pediatrics very briefly the last time.  Today we'll be making a proposal.  The proposal will relate to a pediatric population PK design template.  We'd like to recommend the use of this template for getting information about pediatrics during drug development.  We feel that this approach is efficient in many cases.  We feel it's under-utilized for a variety of reasons relating to perhaps a lack of understanding, perhaps related to a concern that FDA will not accept this type of study approach.  But we'd like your input on the template that we'll be presenting.

            Related to that, at the last meeting I had mentioned that we have this database at FDA related to pediatric studies.  These are studies that were done under our pediatric rule.  We felt this database is loaded with information that we could capitalize on by studying it, looking for trends, and learning something about the pediatric clinical pharmacology situation.  We'll update you on our progress.  It's been slow.  It's been difficult because we don't have access to an electronic database, and much of our time is simply gathering and assembling data.  But nevertheless, we'd like to share with you some of the things that we've learned so far, but more importantly what we'd like to do going forward and look for input on designing the studies of this database.

            The third topic, which we'll talk about tomorrow morning, is what I'll call a work in progress.  We'll all familiar with the human genome.  We've been bombarded with information about it, particularly this month on the 50th anniversary of identifying the double helix structure for DNA.  Certainly the dream of genomics is to develop new and better treatments for disease states.  But we feel there's a lot to be gained.  There's a lot of substantial improvement that could be made by integrating pharmacogenetics into the treatments that we now have and using this science as it matures for identifying more optimal doses for subsets of the population.

            We'll continue to talk about this tomorrow.  What we're going to emphasize is moving forward with the knowledge that we have on polymorphic drug-metabolizing enzymes that influence variability in drug response, especially toxicity.  It's a challenging area.  Many questions come up in the context of this, things like how much variability will genetics explain, how much of an effect on drug dose will genetics explain, how important is it.

            But I think more importantly where we're heading is to create a general construct for looking at improvement in existing therapies, existing therapies being approved drugs, to determine what criteria we ought to be thinking about that would warrant updating labels for products to optimize drug dosing using genetic information.

            Last time we talked about specifically thiopurine, TPMT.  Tomorrow we'll touch upon that, but we'll also be looking for a broad way to best program in this area and what we need to be thinking about in assessing data, assessing evidence to update labels.  So a rational scientific basis.

            Now, our fourth topic today is going to be a new topic.  We'll actually talk about it tomorrow.  We've been working pretty much over the last year in the area of drug-drug interactions.  It continues to be a major problem if you read the current literature in JAMA and the New England Journal of Medicine about adverse drug reactions and the high fraction of those that are related to drug interactions.  We have some ideas on revising our guidance on drug-drug interactions.  We have some questions on transporter based drug-drug interactions.  As was stated in our risk assessment workshop, some matters always need assessment in regulatory review, including new interactions that we may not have paid as much attention to in the past, such as interaction involving glucuronidation and transporter interactions like P-gp.

            So we're going to bring this topic forward tomorrow with some issues and questions.  We'll talk about the use and extension of a classification system for 3A4 inhibition for single and multiple drug interactions.  This is a classification system that we can see moving towards label language for bringing some consistency to how we report drug interactions in the label.

            And a big question that we frequently get from sponsors during the drug development process and we ask ourselves is when and how should the role of P-glycoprotein in drug interactions be investigated.  This is an emerging area and we're beginning to see clinical evidence that this important and we'd like to develop a path forward that's reasonable and rational.

            Each of the presenters is going to have some specific questions on the topic, but I'd like you to think about some of the broader questions that we have for the session.  For example, you'll hear many proposals during the next day and a half.  Aside from the specific questions about the subtopics of today's meeting, think about the rationality of these proposals.  Are they reasonable?  Are they feasible?  Overall, do you think these will enhance the quality of drug development and regulatory review?  Are these the priorities that we should be looking at in our clinical pharmacology program?

            We have works in progress, topics number 3 and number 4.  That means we need input as we move forward with these programs and some advice on whether these objectives are worthwhile.  And in particular, what is the best way to integrate new science and technology, whether it's genetics, whether it's P-gp transporter information, into therapeutics and regulatory review?

            Well, that's the introduction to the lineup for today.  I look forward to the discussion.  It was a terrific discussion in October.  Again, as we look around the table, the expertise of this committee is really substantial, and we're looking forward to defining and expanding upon the proposals that we're going to make during this meeting.  Thanks.

            DR. VENITZ:  Thank you, Larry.

            Now we're moving to our first topic.  As Larry indicated, we're going to talk about exposure response as a way of justifying dose adjustments.  The introduction will be given by Peter Lee.  He's Associate Director of the Office of Clinical Pharmacology and Biopharmaceutics.

            DR. LEE:  Good morning.  The first topic we're going to discuss today is quantitative risk analysis using exposure response for determining dose adjustment for special populations.

            This topic has been discussed in our previous meeting back in October 2002.  In the last meeting we talked about three main topics.  We had proposed a standardized approach to estimate the probability of adverse events in special populations using exposure-response information.  We also discussed a regulatory decision tree for recommending dose adjustment in special populations.  And lastly, we also discussed the potential application of utility functions for risk and benefit assessment.

            So today we're going to present several examples to illustrate what we talked about in the last meeting.  We're going to present examples of a standardized approach for using exposure-response information to adjust dose in special populations.  We also are going to present one example of using population analysis to obtain PK/PD information from the large clinical trials.  And in the last example, we're going to present a methodology to applying utility function for an optimal dosing strategy.

            Just to give a little bit of background information, as you know, many of the NDAs may contain up to 20 or more clinical pharmacology studies, and in these studies different intrinsic and extrinsic factors may be studied and these factors may influence the pharmacokinetics of the drug in these special populations.  Therefore, we need a consistent approach to determine the dosing adjustment requirement in these populations.

            Here's one example.  In this particular example, we have about 11 factors that have been studied in the NDA.  As you can see, the area under the curve of the drug may change depending on what factors from 0 percent, which is no change, to a 60 percent increase in the special populations.

            So the question is, how do we make the dose adjustment according to the pharmacokinetic results?  Where is the cutoff?  Do we adjust the dose at 30 percent increase of AUC, or do we adjust the dose at 60 percent increase of AUC?

            So the answer is that we had to look at PK/PD information and determine what is the clinical significance of this AUC change.

            So there are some issues related to the dosing adjustment in drug labels of NDA submissions.  Quite often we have seen inconsistency in dosing adjustment recommendations in the initial label of NDA submissions.  Exposure-response information, as is required to interpret the pharmacokinetic change, is now always available in the NDA submission.  The FDA reviewer had to conduct additional exposure-response analyses in order to interpret the AUC change.  Therefore, we feel that a standard for analyzing and interpreting the exposure-response information will be critical and beneficial to regulatory decision making in terms of dose adjustment in special populations.

            So to improve the current status, in the last meeting we had proposed to develop and evaluate a standardized approach for the reviewer and possibly for industry to quantitatively assess the impact of exposure change on either safety or efficacy that results from a change in pharmacokinetics due to intrinsic and extrinsic factors.

            This is the standardized approach that we had proposed in the last meeting.  In this example, basically we have seen an increase of exposure of the test population compared to the reference.  Using exposure-response information, we can estimate the distribution of response in both reference and test populations.  If we could determine what is the critical value of response, which is considered clinical significance, which is the vertical line here, then we can calculate the probability of a clinically significant response based on the PK change, as well as the PK/PD relationship.

            So in order to interpret the clinical significance of pharmacokinetic change, we need to have observed data in pharmacokinetics in special populations.  We also need data of the PK/PD relationship.  With this information, then we can estimate the probability of adverse events in the special populations with the response that is greater than the clinically significant critical value.

            However, in order to determine the clinically significant critical values, we need to base it on a risk and benefit assessment of the drug therapy.  Currently we're doing that on a case-by-case basis through a discussion between clinical pharmacology and the medical reviewer.  But in the last meeting with the committee, we also proposed that we can use a utility function to assess the risk and benefit of pharmacokinetic change in the special populations.

            In the last meeting, we also discussed a decision tree for dosing adjustment recommendations.  Since the last meeting, based on the recommendation from this committee, we have made some modifications of the decision tree, and this is the current decision tree.  Basically we ask a number of questions.

            First, according to our current guidance, we ask whether the 90 percent confidence interval of test over reference is within the default no-effect boundary.  A no-effect boundary could be, for example, 80 to 125.  Now, if the answer is yes ‑‑ we think it is the no-effect boundary ‑‑ then there's no dose adjustment required for the special populations.  But if the answer is no, which means the 90 confidence interval is outside the boundary, then we ask the next question, whether we have a PK/PD relationship.

            If we do have a PK/PD relationship, then we will take the standardized approach to estimate the probability of adverse events and probability of effectiveness in the special populations and ask the question whether that's a clinically significant change from the typical population.  If it is considered clinically significant, then we will recommend a dose adjustment or precaution or warning in the drug label.

            As I mentioned, there will be several examples discussed in today's meeting.  The first two examples will be used to illustrate the use of the proposed standardized approach for estimating the probability of toxicity in special populations using exposure-response information.

            And the next example will be used to illustrate the potential utility of population analyses to obtain exposure-response information from large clinical trials.  We think this is a very important topic because the large clinical trials represent a unique opportunity to obtain exposure-response information from the studies.

            The last example will be used to demonstrate a method of applying utility functions to optimize a dosing strategy.

            Today we have four speakers to present the examples.  The first speaker is Nhi Nguyen from DPE I, and she will present the first example.  The second speaker is Dr. Jenny Zheng from DPE III.  She'll present a second example of the standardized approach.  And the third presenter will be Dr. He Sun from DPE II, and he will present the population PK/PD approach.  The last speaker, our guest speaker, is Dr. Mats Karlsson from Uppsala University, and he will present an example illustrating the utility functions.

            After each presentation, we're going to ask one or two questions to the committee.  I'd like to present the questions now so that hopefully the committee can keep those questions in mind when listening to the presentations.

            The first two questions relate to the standardized approach.  We would like to ask, under what treatment circumstances, for example, intrinsic or extrinsic factors or therapeutic areas, would this standardized approach not be applicable?  We also ask a second question.  Does the exposure response differ between special populations and typical populations?  If so, how can the differences be detected?

            The next questions will be related to the population PK/PD analysis, and there are two questions related to that topic.  The first question is, what are the utility and general limitations of linking pharmacokinetics obtained from the population analysis to the response endpoints?  And the second question is what are the general considerations in using exposure response for dose adjustment in special populations, especially using the population approach to obtain the exposure response?

            The last question is related to the utility functions, and the question will be, can the presented approach of utility function be generalized to other scenarios?

            So with that, I want to introduce our first speaker of the examples, Dr. Nhi Nguyen.  She will present the first example of the standardized approach.

            DR. NGUYEN:  Good morning.  This morning you will hear a presentation on a method of analysis and how it was applied in regulatory decision making.

            This slide will summarize how we did the analysis for this NDA.

            The first step is to develop your exposure-response models.  The exposure-response models should be based on large, randomized clinical trials, trials that explore a wide exposure range and include a large number of people and are of the longest duration.

            The second step is to have an expectation for your target window of exposure.  How much benefit does one need and how much risk is one willing to assume?

            The third step is to example what happens to exposure response when various intrinsic and extrinsic factors are introduced.  Typically studies in special populations include pharmacokinetic information and are underpowered to provide good response data.  So with the appropriate assumptions about exposure response in these special populations, we took the data from the special population studies, the individual data, not just the mean data, and integrated it into the exposure-response models and then determined the probability of effectiveness and safety.

            So probability is on the y axis here and the sum of these bars equals 100 percent.  So you can see that we not only have a feel for the maximum likelihood of benefit or risk, but we also have a feel for the tails of the distribution.

            This slide summarizes how we did the analysis for this review.

            I will take one more slide to further explain the last step, determining the probabilities.  And for this example, I've chosen the QTc interval which is a surrogate for torsade, a fatal ventricular arrhythmia.  A clinical trial that examines QTc prolongation may have a distribution of baseline QTc intervals that look like this.  Modeling the data may result in concentration QTc slope distributions that look like this.  So if we want to see what happens when an intrinsic or extrinsic factor is introduced, we took the results from the PK study, and for this example I'm illustrating Cmax and overlaid it into the known concentration-QTc relationship.

            So, in essence, we sampled from each of these distributions and created a virtual patient, and we did this 1,000 times to determine empirically what happens with a concentration and achieving a specific QTc.  So by doing these simulations, we were able to test combinations of the tails of the distributions that were untested.

            So that leads me to our objective which was only to quantitate the risk-benefit of a drug.  A decision about what to do about the risk-benefit should be made with the whole review team or the domain experts.

            In developing the exposure-effectiveness model, for the primary endpoint, we chose the largest clinical trial which was also of the longest duration.  The primary effectiveness endpoint and pharmacokinetics were collected at baseline, week 2, and week 4.  This study also explored the largest exposure, and these doses have been changed for the purpose of this presentation, but let's just say that 1 milligram a day is the sponsor's recommended starting dose, with titration to 2 milligrams a day.  So this study explored a dose greater than and a dose less than the sponsor's recommended dose.

            Next we developed the exposure-safety or exposure-risk models.  For these models, we used the adverse event data from all the pivotal clinical trials.  Now, you can imagine that large clinical trials may have hundreds of adverse events.  So after discussions with other members of the review team, we prioritized these adverse events and focused on these six:  dizziness, edema, liver toxicity, palpitations, tachycardia, and vertigo.

            We also analyzed QTc prolongation because of drug properties suggestive of QTc prolongation.  For this analysis, we chose the study that had the most information on the time course of QTc prolongation.  You will note that this was a drug-drug interaction study, and the sponsor used half the recommended starting dose.  ECGs were only measured up to 4 hours post dose, so there were some study design limitations.  And the study contained 24 hours of drug concentration data.

            So now that you've seen the exposure effectiveness and the exposure risks that were assessed, let's take a look at the models.

            This is the exposure-primary effectiveness model, and the asterisk in the following slides will indicate the mean Cmax of the sponsor's recommended starting dose of 1 milligram.  These lines indicate the mean Cmax for the 1, 2, and 4 milligram dose, and you will note that the increase in concentration is more than dose proportional.  The maximum effect was 7.6 and the concentration that produced half the maximal effect was about .2 units.

            In the following slides, I show a mean Cmax line to keep the slide clean, but really we are considering the entire distribution of individual Cmax's.  So it's something that may look like this with some overlap between the 1 and 2 milligram dose.  So that's the exposure-effectiveness model.

            When you look at the risks, each blue line on this slide indicates one individual's concentration/QTc prolongation relationship.  The QTc corrections shown here are individual corrections obtained by nonlinear mixed effects modeling.

            There was a statistically significant relationship between concentration and QTc prolongation.  However, you can see that there is a lot of variability.  The starting dose in some patients results in no QTc prolongation.  However, in other patients, it results in substantial QTc prolongation.  You will also note that we have very little data around the concentration of the mean Cmax for the 2 milligram dose.

            For our analysis of other adverse events, we found three adverse events to be statistically significant and that was tachycardia, palpitations, and edema.  However, since the analysis of all these adverse events was similar, I will only present one for the sake of time.

            This slide shows the probability of tachycardia by the effective dose, and the effective dose is an adjustment of the actual dose to account for the saturable first pass of the drug.  So, 1, 2, and 6 really correspond to 1, 2, and 4 milligrams of drug.

            The probability of tachycardia was dependent on weight and dose.  So in a 70-kilogram patient, you can see that there is a very small probability of tachycardia, and this probability does not increase with a six-fold effective dose increase.  Whereas, in a 50-kilogram patient, there is about a 5 percent probability of tachycardia, and then this increases about 1 percent with a six-fold effective dose increase.

            So now you've seen the exposure-effectiveness model, the exposure and QTc model, and then the probability of tachycardia by effective dose.  So now we're equipped with the models necessary to interpret results of the special population studies.

            This slide shows results of two special population studies presented in terms of changes in AUC and Cmax.  Ketoconazole with a half a milligram of drug resulted in a 13-fold increase in AUC and a 7-fold increase in Cmax, and grapefruit juice with 1 milligram of drug resulted in a 7-fold increase in AUC and a 6-fold increase in Cmax.

            So the next step is to see how these data integrate into the known exposure-response relationship.  This is the same figure you saw earlier, only it's smaller, of the concentration effectiveness.  When a half milligram of drug is given, you would expect to see an effectiveness around 3.  Taking ketoconazole with a half milligram of drug increases concentrations about 7-fold, reaching the Emax of about 7.6.  Taking 1 milligram of drug with grapefruit juice results in about a 6-fold increase in concentration, and no additional effectiveness.  Now, hypothetically if ketoconazole were given with 1 milligram of drug, we might expect to see a similar increase in concentration.

            If we look at the concentration/QTc relationship, a half milligram of drug would result in about this amount of QTc prolongation.  Taking ketoconazole with a half milligram of drug pushes patients from this amount of QTc prolongation over to this amount.  Taking 1 milligram of drug with grapefruit juice results in about a 6-fold increase in concentration, and the concentrations are then off the figure and we do not have QTc data there.  And then hypothetically again, if ketoconazole were given with 1 milligram of drug, we might expect to see a similar response.  So in this situation, we would not be able to make any conclusions about what happens at these higher concentrations on QTc prolongation because we do not have data there.

            Now, I also want to remind you again that there is a distribution of Cmax's and slopes.  So really we are looking at data that looks like this.  So if we want to consider the worst case scenario, we have to consider both of these distributions, and the results of some of those simulations will be presented in the next slide.

            Now, if we look at the probability of tachycardia, taking a half milligram of drug with ketoconazole in a 50-kilogram patient barely increases the probability of tachycardia, whereas taking grapefruit juice with 1 milligram of drug increases the probability of tachycardia about 1 percent in the 50-kilogram person.  In a 70-kilogram person, you can see that this slope is pretty much a straight line and there is not much of an effect on the probability of tachycardia.  Then again if ketoconazole were given with 1 milligram of drug, we might expect to see a similar response as that with grapefruit juice.

            So to summarize the data integration, ketoconazole with a half milligram of drug results in a 13-fold increase in AUC and a 7-fold increase in Cmax.  And this translated into a 4-unit effect, and we did simulations to determine that 5 percent of the population may have a prolonged QTc greater than 32 milliseconds.  Realize that these simulations are determined from the data in the ketoconazole study.  So we could present this data in other terms, such as change from baseline or percent or the time above a certain threshold QTc.

            And then the probability of tachycardia with a half milligram of drug and ketoconazole was barely increased or affected.  Grapefruit juice with 1 milligram of drug resulted in a 7-fold increase in AUC and a 6-fold increase in Cmax, and this translated into no additional effectiveness, and we were unable to conclude anything about the effect on QTc because we did not have data at those higher concentrations.  And then the probability of tachycardia increased about 1 percent in the 50-kilogram patient, and there was negligible effect in the 70-kilogram patient.  And then again if ketoconazole were given with 1 milligram of drug, we may expect to see a similar response as that seen with grapefruit juice.

            So at this point we would present the table to the review team and weigh the effectiveness and risks and realize the assumptions of our models.  One is that it is the higher drug concentrations, not the intrinsic or extrinsic factor itself, that alters response.  We recommended to conduct an appropriate QT study, one that explores a wider concentration range and one that collects QT data over 24 hours.

            DR. VENITZ:  Thank you, Nhi.

            We have time for questions.  Go ahead.

            DR. SHEINER:  I gather that the various parts of the various models were gathered from different data sets sometimes.

            DR. NGUYEN:  For the effectiveness, we used the largest clinical trial, and then for the safety and risks, we used the largest clinical trial and the other pivotal trials.

            DR. SHEINER:  I guess the question is this.  You've got several models that are translating from A to B and then B to C and so on.

            DR. NGUYEN:  That's correct.

            DR. SHEINER:  And the question is, were enough of them gotten from the same set of people so that you could look at things like correlations?  Does it turn out, for example ‑‑ not that it should ‑‑ that the people with the high concentrations essentially ‑‑ in other words, your model is kind of assuming that this association that you see, there are no correlations.  So it's not necessarily true that somebody who has, let's say, a raise in level and doesn't have toxicity will also not have efficacy or something like that.

            You've got a set of relationships that translates from concentrations before you add ketoconazole, let's say, to afterwards, and then you map from concentration to effect, but there is no part of this thing that says, well, when you raise the concentration to the ketoconazole, maybe the relationship of concentration to effect is not the same.  I'm not saying that it is, but you don't have any evidence to say one way or another.  Is that right?

            DR. NGUYEN:  Yes.

            DR. VENITZ:  Any other questions?

            DR. FLOCKHART:  One thing directly to that point.  There is some data ‑‑ but this might be addressable ‑‑ to suggest that ketoconazole itself can affect cardiac repolarization.

            DR. NGUYEN:  That's right.

            DR. FLOCKHART:  You might have data on that from the control arms of the smaller trials.  If they show no effect, that would be reassuring.  It doesn't completely get to Lew's point because it's still possible that the concentration-effect relationship is different in the presence of ketoconazole.

            DR. CAPPARELLI:  Just a clarification so that I understand the terminology.  When you say probability of tachycardia, you're speaking of sinus tachycardia in this case, not torsade de pointes?

            DR. NGUYEN:  That's correct, sinus tachycardia.

            DR. CAPPARELLI:  I just wanted to be clear because I think one question that I had is, did you take a similar approach to heart rate that you did to QTc?  Because your sinus tachycardia is going to be relative to where you start from, at least the risk.

            The one thing that I think was brought up as a question is, are there extrinsic factors that we need to think about in these models?  Clearly, strictly from a PK standpoint, I wouldn't expect a 50-kilogram person to have a different response based on weight.  So I think there clearly is an extrinsic factor that's linked to the 50-kilo patient rather than the 70-kilo patient.  But I'm not certain that it's gotten here.  So I have some questions about the classification scheme based strictly on weight, and maybe linking to where their baseline heart rates would be of help from that standpoint.

            DR. NGUYEN:  Actually for the tachycardia analysis, we would have preferred to analyze it by heart rate, but the data were collected like that.  So it was sinus tachycardia, yes or no.  So we did a logistic regression.

            DR. CAPPARELLI:  Was there an age effect or other disease effects that you looked at in terms of heart failure?  It's kind of difficult to look at this and see where you expect a large change in concentration such that the 70-kilo person at the highest dose is going to have much higher concentrations than the 50-kilo person at the lowest dose.  And yet, you're seeing this differential PD response.  Without understanding what's causing that, I think it becomes very difficult to extrapolate from this aspect of the analysis.

            DR. NGUYEN:  Probably the 50-kilogram person did receive more of a dose, milligram per kilo, than the 70-kilogram person.  But the analysis ‑‑ they were given the same dose.  So we didn't have concentration data to analyze data by concentration and probability of tachycardia.  We only had dose data.

            DR. SHEINER:  Just a comment.  I think we're getting at the fundamental problem that what you want to do is extrapolate to new circumstances, people having these other co-factors.  You want to get some reasonable guess as to what's going to happen, what's dangerous and what isn't.  Yet, you're extrapolating from observational data based models, which is the hardest thing to do because you don't know where causality resides in those models.

            One of the solutions in the past is to just not do it, and I don't think that's the right solution.  But I do think there is no easy solution, and we have to be quite careful about things and recognize that we're talking about outer boundaries and recognize that we're talking about sort of worst case scenarios or maybe even best case scenarios.  We can't be sure.  We have to somehow get comfortable with the increased degree of uncertainty that arises in this activity.  But as I say, I think we should do it because the alternative is even greater uncertainty.

            DR. DERENDORF:  There are a lot of straight lines in your concentration-effect and dose-effect relationships.  Is there enough evidence that you have linear relationships between those parameters, particularly when you use them to extrapolate?

            DR. NGUYEN:  Which one are you referring to?

            DR. DERENDORF:  The concentration-QTc prolongation plot and then also the one below the dose-tachycardia plot.  You just have straight lines there that suggest that concentration and effect are linked that way.  Do you know that?

            DR. NGUYEN:  No.  I mean, that was the data that we had.  So like for the QTc, they measured ECGs at 0, 1, 2, and 4 hours post dose, and they had 24 hours of concentration data.  So that straight line is the relationship between the concentration and QTc.

            DR. DERENDORF:  You think it is or you know it is?

            DR. NGUYEN:  Well, that's what it was in that population.  They could have gone with a higher dose range, and then we could have seen what type of model the relationship is.  So I don't know.

            DR. DERENDORF:  Then the other question that's related to the previous question that I really am puzzled with is this 50- versus 70-kilo situation where you have a 6-fold dose and a 70-kilogram person doesn't have probability versus a one-sixth of a dose in a 50-kilogram person.  So there would be quite different exposures and very different risks.

            DR. FLOCKHART:  It just means to me that the relationship between dose and weight isn't a simple linear relationship.  You're right.  There may be something else involved, but that's not uncommon.

            DR. SADEE:  I have a comment also on the variability from one patient to the next.  You're looking at two interactions.  It's maybe one of the classic examples in pharmacogenetics where you have a variety of different genetic variance from one patient to the next, and actually that could affect the interaction between the two drugs so that you may have specific cases in the single patients that are totally different in their exposure than others.  So there would be no way of extrapolating from that because you're disturbing the very relationship with the dose response that you're looking at.

            DR. FLOCKHART:  My difficulty here is the FDA is faced with the problem of trying to make a rational prediction.  We can sit as academics and ding them all over the place for it.  You know, you can't do this and you can't do that.  But the reality is you have to try.  And I think Lew's point is salient.  We have to try and include the error.

            DR. KEARNS:  And that's the point that I think I want to make.  Back to your ECG slide.  It's not to be critical.  It's quite exciting to see 20 percent of people have a response that way and then one outlier at the top who really had one.

            But I was struck by the recommendation that you showed on your slide which was, okay, we did this.  Now we go back and recommend a trial with more concentrations, wider range.  And as Dr. Sheiner was kind of getting at, there's a lot that's riding on this extrapolation.  I think from a public side, there's always the question of how much additional time is that going to take.  From a medical side, there's always the question about that one or two outliers.  Are those the people who die in that trial because they have some hERG channel defect that's not recognized at the outset?

            So I think trying to go back and do the diligent thing is to wire up the model as best you can.  For instance, if that was a pediatric population and you looked at baseline QTc's, you'd see quite a different dispersion based on age for no treatment than you would in an older free-living population.

            So is the applicability of the model approach ‑‑ can we go across populations?  It depends I think.  But to jump to the study and to add the patients, to add the concentrations, to maybe add the risk until you've taken all the flies that you can out of the ointment could be premature.

            DR. VENITZ:  Thank you, Nhi.

            Let's move to the next presentation.  Dr. Jenny Zheng.  She is a pharmacometrics reviewer in the Division of Pharmaceutical Evaluation III.  She's going to give us a second example.

            DR. ZHENG:  Good morning.  Today I'm going to present another example to illustrate how the dose-response relationship was used for recommending dose adjustment.

            In our review process, it's very often to see the pharmacokinetics of a drug is influenced by intrinsic factors such as age, gender, impaired renal and hepatic function and extrinsic factors such as drug-drug interactions.  In this situation, we have to ask the question what is the clinical significant of the changes in concentrations.

            Currently the decision will be made based on the clinical assessment based on the clinical experience and totality of the evidence.  But the disadvantage of that approach is it's not a quantitative and standardized approach.  The assessment could be pretty subjective.  In other words, the decision may not be the same based on who makes the assessment and from what perspective.  Therefore, we propose from a clinical pharmacology perspective to use the exposure-response relationship to bridge the response and exposure data to quantitate the influence after changes in the exposure.

            The example I'm going to present will focus on the drug concentration increase and the safety assessment.  This is drug Z.  It's a noncardiac drug.  From both preclinical and phase I studies, it shows that the drug caused QT prolongation and this QT prolongation is concentration dependent.

            The phase I PK studies showed three factors increased the drug concentration.  In an age study, it shows that the mean steady state maximum concentration was 100 percent higher in elderly subjects as compared with Cmax in young subjects.  The renal study demonstrates steady state Cmax in severely renally impaired subjects was 50 percent higher as compared with healthy subjects.  And drug interaction studies showed ketoconazole increased the steady state Cmax by 60 percent.

            Knowing the concentration increase in this situation, the question raised is, should dose be adjusted in elderly, renally impaired subjects or when co-administered with ketoconazole?  To answer that question, we need to understand the effect of increase in drug Z concentration on the QT prolongation which will rely on the exposure-response relationship.

            The exposure-response relationship was obtained from several phase I studies.  They were all placebo-controlled crossover studies.  The doses included in the study were a clinical dose and two times of the clinical dose and three times of the clinical dose.  The higher dose is important here to provide the wide range of the concentration which is the key for obtaining exposure-response relationship.  From all these phase I studies, blood samples were collected for drug measurement.  Also the QTs were measured.

            The results of the analysis are shown in this slide.  The QT prolongation is represented as delta QTc, which is the QT change from the baseline.  So the association between the delta QTc with the concentration was described by a simple linear regression model.  The linear mixed effect model was used for analyzing these data.  The dashed lines represent individual regression lines.  The solid line represent the population regression line.  The wider band of the lines indicates that the inter-subject variability is quite high.  The estimated slope ranged from 1.5 to 7.6, indicating that for some of the subjects, the delta QTc change is sensitive to the concentration change.  In some of the subjects, the change is not quite as sensitive.

            An outlier analysis is a very important part of QT assessment.  We want to know how many subjects would experience the delta QTc longer, for example, 10 milliseconds, 20 milliseconds, 30 milliseconds, or 40 milliseconds.  Unfortunately, the phase I study usually included a limited number of subjects which limits its ability for that type of analysis.

            For the phase III study, even though hundreds of subjects may be included in that analysis, the QT measurement is not as intensive as the phase I study.  So it's difficult sometimes to capture the outlier from the phase III study.  On the other hand, if you're interested in the special population, even a phase III study may not provide sufficient number, for example, severely renally impaired subjects.

            So in order to make an outlier comparison between the population, a simulation exercise was conducted here.  Most specifically, the phase I data, the concentration data was modeled assuming the logarithmic distribution in the PK parameter.  Using that model, 2000 maximum concentration was simulated for young subjects, elderly subjects, for renally impaired subjects.  The same approach is used to simulate 2000 Cmax for when ketoconazole is co-administered with the drug Z.  So we have 2000 concentration in each special population, special situation.  Then we used the exposure-response relationship as described in the previous slide to predict delta QTc.

            The results of the age effect are presented in this slide.  The data is presented as the percent of the subjects who would have the delta QTc longer than 10 milliseconds, 20 milliseconds, 30 milliseconds, 40 milliseconds.  These results indicated that about 2 percent of young subjects would have delta QTc longer than 40 milliseconds and for the elderly, the percentage will increase to 7.3 percent.  So for delta QTc longer than 30 milliseconds, in young subjects it's about 8 percent; in the elderly, it's about 19 percent.  So a similar trend could be seen for a delta QTc longer than 20 milliseconds and 10 milliseconds.

            This slide presents the results for renal function.  As you can see, most subjects with severe renal impairment would have longer QT prolongation than the normal subjects.  For example, for the normal renal function subjects, 2 percent would have delta QTc longer than 40 milliseconds, but if you have severely renally impaired function due to the concentration increase, there will be 5 percent of the subjects who would have a delta QTc longer than 40 milliseconds.

            The results in this slide show ketoconazole increased the percent of subjects who experienced a certain extent of delta QTc.  Like if you're taking the drug alone, 2 percent of subjects would experience delta QTc longer than 40 milliseconds.  But if you take drug Z with the ketoconazole, the percentage will increase to 6.2 percent.

            The percent of subjects with delta QTc longer than 40 milliseconds is summarized in this slide.  You can see that the risk of having a delta QTc longer than 40 milliseconds is higher in elderly subjects, in severely renally impaired subjects, and when the drug is co-administered with ketoconazole.

            Examination of the creatinine clearance indicated that the age effect might be partially attributed by reduced renal function.  So in the age study, creatinine clearance was 50 percent lower in elderly subjects as compared with the young subjects.  So it's believed that age effect would be reduced if the renal function effect was corrected by dose reduction.

            Since the consequence of the worst event could be very severe, the question was asked in the review team, what would be the effect of ketoconazole in subjects with severe renal impairment?  Not many subjects would belong to this group, even from a phase III study.  So in order to make that assessment, a simulation was conducted.

            First, steady state Cmax in severely renally impaired subjects was simulated as I described earlier.  In the second step, the Cmax ratio of drug Z at presence and absence of ketoconazole was obtained from the crossover study so that the ratio actually characterized the ketoconazole effect.  From that study the ratio ranged from 1 to 4.  So the combined effect for both factors was simulated by just randomly multiplying the maximum concentration from step 1, which is the maximum concentration for severely renally impaired subjects, and the ratio from step 2 which characterized the ketoconazole effect.

            The results are shown in this slide.  As you can see, 19 percent of subjects who are severely renally impaired would experience a delta QTc longer than 40 milliseconds when co-administered with ketoconazole.

            This slide just simply summarizes all of the factors, the effects.  It summarizes young subjects.  It's the percent of subjects with a delta QTc longer than 40 milliseconds.  For young subjects, it's about 2 percent.  In the elderly, it's almost triple the percentage, up to 7.3 percent, and more than double that percentage in severely renally impaired subjects.  And when drug Z is co-administered with ketoconazole, the effect could be very dramatic if the two factors are combined.

            That analysis leads to the conclusion that the increase of concentration by age, severe renal function, and co-administration with ketoconazole resulted in increased number of subjects with a delta QTc longer than 40 milliseconds.  The effect is more significant when two factors are combined.

            Based on that analysis and the consideration of the nature of an adverse event, a dose reduction was recommended in severely renally impaired subjects.  A dose reduction was also recommended when drug Z is co-administered with ketoconazole.

            DR. VENITZ:  Thank you, Jenny.  We have about 5 minutes for questions.

            DR. DERENDORF:  I think it's the same issue as in the last case.  You're assuming that the exposure-response relationship that you got from your phase I study is a constant and it doesn't change in elderly or in severe renal impairment.  So the calculations that you're making are all focused on exposure, and then at the very end, you convert that into expected ‑‑

            DR. ZHENG:  I don't quite understand your point.  Actually the delta QTc versus concentration, that is the relationship between the effect versus concentration.  I don't think we make any assumption with a constant concentration.

            DR. DERENDORF:  No, not constant concentration.  But the relationship between the exposure that you have in your different cases and the outcome ‑‑ you take that linear relationship that you have from your phase I study where you have concentration versus change of QTc and apply that to all of these cases assuming that this relationship holds true for all of these.

            DR. ZHENG:  Okay.  So you have the problem with the extrapolation from the young healthy subjects to the elderly population.

            DR. DERENDORF:  I don't see any evidence that it holds.

            DR. ZHENG:  In one of the phase I studies, they included not only the young subjects but the elderly.  So we do look at the relationship there.  We don't see much difference in terms of the slope, the relationship.  So that's one of the information we could have.

            In terms of the drug-drug interaction and the severely renally impaired subjects, yes, we don't have data to show that they are going to have the same relationship.  But you are right.  That's the assumption we have to make for these type analyses.

            DR. CAPPARELLI:  Just as a follow-up on that, because this is a recurrent issue ‑‑ I mean, this particular QT prolongation looking at drug concentration effects ‑‑ has there been a systematic look at several of these drugs looking at especially, say, with renal impairment where you're going to have changes in electrolyte abnormalities and looking really at sort of a population dynamic model to identify the covariates?

            So I think as Hartmut was saying, we're going forward with the assumption that the exposure-response relationship is totally uncorrelated with the changes in ‑‑ the disease states that are causing the changes in PK.  I think this is actually a great example where maybe in some across-study evaluations, one could actually look at some of these populations not only at the variability in response in subpopulations, but maybe the electrolyte differences in your renal failure patients may change that slope entirely.  So adding these effects as we go along in the chain, it's nice along the way to test some of these assumptions.

            DR. ZHENG:  I think if we could have enough information, definitely that's a good thing to do.  But I think here, unfortunately we just don't have that much information.  So the focus here is simply the effect of concentration on the delta QT.

            DR. SHEINER:  Of course, the answer is get more information.  But the answer to the problem when you don't have enough information is not that you have to make some assumption and go with it, but that you have very carefully display.  It is sort of what I was indicating before.  You have to carefully display the limits of your knowledge.

            So if I look at, for example, the last page and the last several slides you showed, you've got these bars that are just heights, the amount of change with the elderly or renal failure, and there are no uncertainty intervals on them.  Yet, this is exactly what you need to pay a lot of attention to, it seems to me, in this kind of a situation so that you can have a rational dialogue with other people.

            And where do the uncertainties come from?  And we have techniques whereby you say, well, look, I have this linear relationship between QTc change and the concentration, but I could put in some uncertainties about, let's say, whether it applies to other populations.  Then I can actually build that into my projections, and I can see that instead of having 30 percent of people above QTc of 40, it will be anywhere from 10 to 50 percent, or whatever the numbers are.

            That's the point.  You've got the computers to do it.  It doesn't cost money.  And that's the way I think to deal with the problem that there are so many assumptions that have to be made, sensitivity analyses and honest uncertainty intervals which involve model uncertainty as well as data uncertainty.  And then everybody is talking about the same thing.  It may well be that the conclusion stays the same.

            DR. LEE:  Dr. Sheiner?

            DR. SHEINER:  Yes.

            DR. LEE:  To follow up, if we don't know the true relationship in different populations, how do we build into the model the uncertainty due to population difference?

            DR. SHEINER:  Well, let's say we'll talk about the average slope.  Let's say you're willing to assume it's linear, but it's the average slope that's different in different populations.  So then you just talk to a bunch of people and you say, how big do you think it could be, and you just build that uncertainty in.  Now it spreads out all of your predictions, and it means there's a larger fraction of people who have low values, but there's a larger fraction of people who have high values.

            So it's a matter of assessing the risk.  It's what's the probability based on everything we know, including all the uncertainty, that the value will be greater than this.  And that will be your most educated guess.

            The point is it will be everybody's most educated guess, and anybody who says I don't think that will happen, you'll say, well, you're pointing to the 40 percent that's still below the line because we have uncertainty.  And I understand you're betting on that 40 percent, but we're worried about the 30 percent.  So that's the way we're going to go.

            The point is you're never going to get the answer from doing the numerical calculations.  All you get is an honest statement of what you know and that everybody can agree on.  I think that's the big thing, is that everybody can agree this is the state of our knowledge.  Therefore, if you're going this way and I'm going that way, it's because we're valuing different outcomes differently, and so the expected value comes out differently.

            Another example here of a place for an opportunity for this is in the discussion ‑‑ well, actually, I'll let it go.  But I think you get the idea.

            DR. VENITZ:  Let me just follow up to that, Jenny.  Whenever you do an outlier analysis, which is really what you're trying to do, worst case scenario, what are the few that have a large change in QTc, distribution assumptions are very important in terms of what your final outcomes are.  I look at your simulation slide.  You're talking about a logarithmic distribution.  I'm assuming you mean a log normal distribution.

            DR. ZHENG:  Right.

            DR. VENITZ:  How did you then actually simulate the changes due to the disease states or the drug-drug interaction?  Did you just change the mean or did you change variances as well?

            DR. ZHENG:  Actually I just changed the mean because the model is fitted to the raw data.  For example, the young subjects ‑‑ we modeled that.  So we know the mean for that group.

            DR. VENITZ:  But what about the variance?  I guess what I'm worried about, whenever you look at outliers and you have a change in variance, in other words your elderly or your renal population are probably more variable than your young reference population even in terms of kinetics.

            DR. ZHENG:  We used the same model to model the data for young subjects and the data for elderly.  I mean, the same compartment model.

            DR. VENITZ:  But in terms of your parameter variability, did you use the same variance in your ‑‑

            DR. ZHENG:  No.  The data ‑‑

            DR. VENITZ:  So you used the actual data variance.

            DR. ZHENG:  Yes.  I used the real data variance.

            DR. VENITZ:  Just to follow up on that, I'm assuming when you looked at your slopes, you assumed that the slopes followed normal distribution or log normal distribution?

            DR. ZHENG:  I did that analysis using NONMEM.  So it's an additive model.  So it's normal distribution.

            DR. VENITZ:  Well, based on what I've seen or based on the previous example, that may not be a good assumption.  It could be that you just have a few outliers and have very steep slopes, but the rest of them have a fairly shallow slope.  Whenever you do an outlier analysis, just as a general rule, the distribution assumption of the variances that you make really determine what your final outcome is.

            In addition to that, I would reinforce what Dr. Sheiner said, and that is, I was missing the fact that you didn't really give us an idea of the uncertainty ‑‑

            DR. ZHENG:  Right.  I think that's something I should have included.  Probably I don't have enough information to speak to severely renally impaired subjects, what the relationship will be.  But I do have the information about uncertainty of the parameter estimate.  So I agree 100 percent.

            DR. VENITZ:  Just look at your three slides where you tell us what happens for the young individuals.  Let's say the QTc of less than 10 is 41, 44, and 4 and 42.7.

            DR. ZHENG:  Right.

            DR. VENITZ:  Those are three different simulations.  So you just do that a couple times and you know how much ‑‑

            DR. ZHENG:  Right, yes.  The uncertainty of that estimate should have taken into that exercise.  I agree.

            DR. VENITZ:  I think we have one more question.

            DR. JUSKO:  I imagine you're using the best available metrics on evaluating the exposure-response relationships.  But you might consider using the availability of these data to examine additional possibilities.  For example, if you look at absolute changes in QTc, there might be the possibility that a change of 10 or 20 in the elderly is a bigger problem than a change of 10 or 20 in the young subjects.  And perhaps something in relation to baseline values should be considered.

            Secondly, you're using Cmax as the exposure index, and it would seem to me that in addition to that, the duration of time that a person has an abnormal QTc interval could be an additional hazard that could be factored in in exploring bigger sets of data as you may be doing.

            DR. VENITZ:  Okay.  Thank you.  Thank you, Jenny.

            Before we go on a break, just an announcement.  For those of you on the committee who haven't handed in your lunch orders, now is the time to do it or you're going to starve.

            With that, we're going to reconvene at 10:10.


            DR. VENITZ:  I'd like to reconvene the meeting please.

            All right.  While Peter is posting the questions that the FDA is asking the committee, are there any additional specific questions to the two presenters, Dr. Zheng and Dr. Nguyen?

            DR. KARLSSON:  Yes.  Just a question regarding this last presentation.  The elderly showed quite a change when you looked at the distribution; 7.3 percent would be above 40 milliseconds.  Maybe that would be mitigated by the renal impairment dose adjustment, although I guess even in the elderly population, it wouldn't be that many that's below 30 mls per minute in the elderly population.  But I guess you could look at that through simulations as well.

            But another question is, when looking at the percentage of a particular subpopulation that's outside, is it only the percentage within the population you're looking at, not the size of the population at all?  Because I guess the elderly population is very large in absolute numbers compared to maybe severe renal impairment or ketoconazole.

            Did I make myself clear?

            DR. ZHENG:  Actually could you repeat your second question?

            DR. KARLSSON:  Well, if we're looking at dose recommendations, is it only the percentage within the population that's interesting?  Is it not also the size of the population as such?

            DR. ZHENG:  The simulation I did is 2,000 subjects.  So it may change if you change the sample size to ‑‑

            DR. KARLSSON:  No.  In essence, what I mean is that the elderly population is maybe like 80 million people in the U.S. and the severe renal impairment population is maybe 1 million.  I don't know.  2 million.  So that would come into play as well when making dosing recommendations.

            DR. ZHENG:  Okay.  Yes, I think at the time we make a decision, we should consider the population who use the drug, the impact.

            DR. LESKO:  Mats, I'm not clear how you would consider it, though.  Would you consider it in the context of saying that equal changes in a population that's larger number would get more weight in a dose adjustment scheme?  Or how would you think about it as far as that issue goes?

            It's like saying because the elderly population is so large, there's a greater overall risk to public health than there would be with patients with severe renal function.  But it would seem in labeling a product, I'm not sure that would be taken into account for dose adjustment.  Or if it is, I'm not sure how it would be.

            DR. SHEINER:  It would be if you were thinking about what dose sizes to make, for example.

            DR. LESKO:  Okay, from a manufacturer's standpoint.

            DR. SHEINER:  It's more convenience.  If 90 percent of people are going to use this to make them safe and 10 percent ‑‑ you know.  They're the ones who are going to get out their pocket knives and hack the thing in half.

            DR. VENITZ:  Peter, do you want to review the questions for the committee one more time?

            So those are the questions that we are asked to discuss with regard to the approach that we just two examples of using exposure-response information as a way of predicting probabilities of, in this case, adverse events as a way of deciding about dosing adjustments or not.


            DR. SHEINER:  Did you want discussion on that?

            DR. VENITZ:  Yes.

            DR. SHEINER:  Well, I think we're back to where we were sort of in the very beginning.  It's a good thing to do, but if it's not done with a little extra care, then maybe it's not a good thing to do.  So I think it really comes down to that.

            As I was saying to Jenny at the break, if you do a well-designed, even clinical experiment in which you know exactly the question you want to ask, you've got adequate data by good design, and you analyze it, in a funny way the statistics are relatively unimportant.  The signal to noise is usually pretty high, and it's usually pretty clear what the result is.  Yet, that's where most of the statistics that most of us have seen have been applied, in making sure that type I error is controlled.  And there's nothing the matter with that.  It's a good idea.  But it's not really where you need it.  And it's not that you need sure inference here because you can't get sure inference when you're this uncertain.

            But what we need is we need to have a good way of displaying what we know so that everybody is looking at the same thing and understands the uncertainties.  It seems to me what we didn't see were two ways in which I feel that that needs to be done.

            One, as soon as you generate a simulation model, you have to show me that that simulation model can simulate the data it was derived from.  There are lots of different ways of going about convincing me of that.  Some of them treat the data it was derived from as though they were new by leaving it out and then making the thing and remaking.  There are many, many different techniques.  But the fundamental idea is show me that the sorts of conclusions that you want me to draw about extrapolations are at least verified on the data that you built the thing from when you apply them to those data.  So that's number one.  I want to see a lot of that.

            And then I want to see a real honest uncertainty in my simulation.  We all understand we're not talking about anything that's sure here.  But I want to know how big the uncertainty is and I want to have some way of knowing where it came from.  I personally really want to see model uncertainty as well as data uncertainty.  In fact, I'm more concerned about model uncertainty than data uncertainty.

            What do I mean by that?  I mean if you have 100 patients from whom you've generated the data set, I understand the next 100 patients are going to have somewhat different numbers, and so you're going to get somewhat different conclusions.  And we all estimate that uncertainty, and it's not that tough to estimate.  And sometimes it isn't that large because we have a fair number of patients.

            What's really uncertain is whether or not the relationship we discovered on this population is going to apply to that population.  There we have no data if we haven't studied that population, if we're extrapolating to it.  So there we need just some reasonable guesses.  How different have populations been with respect to this kind of thing in the past with similar sorts of things?  This is where the science comes in.  This is where the judgment comes in.  But you can build those model uncertainties in, and I can get to see how big they are.  That's kind of like a robustness test.  It's kind of like a way of saying how much will conclusions vary if I vary my assumptions.

            Assumptions there will always be.  I'm not against assumptions.  What I'm against is making assumptions look like facts.  It turns out that the things we don't know anything about we put the least uncertainty on, and that's very peculiar.  We choose a form of a model.  So it's a bi-exponential.  And then we say, boom, that's it.  No questions about that.  And that's the thing we know the least well.  What we do know well is the data we observed, and that we say, aha, that's got noise.  So it's kind of like backwards.  I want to see the model uncertainty.

            This is not to be critical.  I believe in modeling and I believe in trying to be quantitative about conclusions.  But without that kind of thing, you won't ever get people around a table to agree on what you know, and if you can't do that, they won't agree on where you ought to go.

            DR. VENITZ:  Any other comments to the first question?  Can we think of any specific examples or circumstances, therapeutic areas where this approach may not be applicable?

            DR. KEARNS:  Yes.  I think one glaring one ‑‑ and Dr. Sheiner again speaks of model uncertainty.  As I might understand it, it would be in the context of the facts of an experiment that we saw examples of.  But as Dr. Flockhart mentioned, for QTc studies where a patient may ingest a medicine that can have effects on its own, that's not necessarily part of model uncertainty.  I don't know that you could predict the rate of co-ingestion of those drugs.  And I would argue that with some combinations that are available, the relationships that you so nicely shared with us could look quite different.

            So are there treatment circumstances that the approach might not be applicable as it was presented?  Yes, and I think that's one example.

            DR. VENITZ:  What about the second question?  I think that's something that we talked about last time.  Differences in the exposure-response relationship between the typical and special populations.

            DR. DERENDORF:  Yes.  I think that there are some examples in the literature where there are clearly differences in exposure-response relationship.  If you think of benzodiazepines, for example, with the sensitivity and all the patient changes, so for the same concentration, you'll get a different response.  But there's actually very little hard data available in the literature because it's hard to study.  If you want to do it right, you have to do a complete PK/PD study, and just focusing on exposure is simply easier and therefore it's done more frequently.  But I think that's what we need:  more clean PK/PD studies in different populations to see how much variability and how much systematic change we have in the exposure-response relationship.

            DR. VENITZ:  Larry.

            DR. LESKO:  There are actually two levels of uncertainty that we're dealing with.  The first ‑‑ and I think it was the first example.  We were talking about exposure-response relationship across various special populations.  The second example illustrated a different problem and that was the assumption of exposure-response relationships between healthy volunteers and then patients.  That's just the fact of the way, at least currently, drugs are developed.  So we have to find ways to think about that, and it would seem there are two things I thought about.

            One was in the pediatric decision tree or in the pediatric rule, we make the assumptions, or at least we ask the questions, about disease progression being the same in adults and kids and whether or not the mechanism of action in the exposure response is the same in adults and kids, and then we proceed down a path of logic that requires perhaps a dose being changed based on simply pharmacokinetic differences to achieve the same type of exposure.

            It gets to the question, though, is my base assumption that exposure response is similar in the absence of hard data, and then I look for reasons, perhaps mechanistic reasons, why it wouldn't be, or do I look and say, well, let me assume it's different and find mechanistic reasons that it should be the same?  For example, in the pediatric adult area, you might ask the question, does a beta receptor's either density or sensitivity change and is it safe to assume that with a beta blocker I'm going to have similar exposure-response relationships?

            It just seems to me that there's a way to think about it mechanistically if one understands the way the drug is working and the changes that are occurring in the special population.  Like in the QTc, for example, if renal patients have altered potassium levels, then we know that affects sensitivity in terms of drug effects on QTc, and that could be kind of a rationale for including some assumption about heightened sensitivity or something like that or a change in the exposure-response curve.  But in the absence of that kind of mechanistic information, it would seem we have to go with the assumption that these exposure-response relationships are the same.

            I mean, does that line of thinking make sense?

            DR. VENITZ:  That would be the way I think.  My default position is there is no difference between my typical population and the special population unless I have either hard data to show that it is, which is rare, or I have mechanistic reasons based on the pathophysiology of the disease of that special population and/or the mechanism of action of the drug to suspect that it is.  Then I either have to question the need for additional studies to show whether it exists or not or build it in as an uncertainty in my model.

            DR. KEARNS:  And Larry, I think another answer to your question that you posed is it depends on the surrogate chosen to assess effect.  For example, if we look at studying a proton pump inhibitor in a child, there's convincing physiologic evidence that the maturation of the proton pump occurs very early and that the children respond to those medicines in ways that are very similar to adults.

            But if you go pick a surrogate far from the tree of effect or drug action and you apply it and say, is gastroesophageal reflux in a 6-month-old the same as in a 46-year-old, and then try to make arguments about bridging, you'll find that the pillars that you've constructed the bridge out of are not worth traversing.  So it depends on how close your surrogate is to where the medicine works.

            DR. VENITZ:  Something else I think we're going to talk about in a minute that I would also consider ‑‑ and I'm pretty sure you do that in your briefings with the medical reviewers ‑‑ is what are the consequences of being wrong.  In other words, what's the utility of whatever assumptions you may not be very certain about?  Sometimes that severity or that consequence may be relatively inconsequential, and then it really doesn't make a difference.  Forget the fact that you have statistical uncertainty associated with it.

            DR. LESKO:  There are many ways these kind of data are handled for purposes of dosing adjustment, and maybe that's one of the reasons we're trying to arrive at a standardized approach to doing it.

            It would seem the safest way of doing it is to simply adjust the dose based on an area under curve change.  The question then becomes what is the threshold level for that area under curve change to trigger that.  And that's where the difference of opinion occurs because you don't have a method on the table that allows one, as people have said, to discuss this in a quantitative way.

            So there may be, in essence, a lot of likelihood of not optimal dosing by doing it that way, either making dose adjustments when you don't need them or not making them when you should based on people's interpretation of the data without a methodology to discuss.

            DR. VENITZ:  But in the examples that you've shown, the endpoints, as far as I can understand them ‑‑ QTc.  That's a surrogate of fatal arrhythmia.  So you're worried about a potential fatal consequence.  There's a high, in my terminology, negative utility associated with it.  On the other hand, things tachycardia or palpitations would rank much lower on the totem pole of my concerns.  But I'm not sure how you quantitatively incorporate that short of using utility functions.

            DR. FLOCKHART:  This is an editing, small point.  If you're going to talk about changes in AUC of a compound, I think particularly when you're talking about the QT ‑‑ but this may be representative of other things ‑‑ the area under the exposure curve is not necessarily the main thing.  The time of exposure to a drug is not the trick.  Parameters like the rate of rise to Cmax can be very important and the QTc max at a given dose can be very important.  There are dis-relationships, blocks between the time-effect curve so the time of the concentration Cmax is absolutely not necessarily the time of the effect Cmax.  It can be later.  So it's possible there would be situations where a parameter other than a change in the PK AUC would be the appropriate parameter.  It could still be a PK parameter, but it might be Cmax itself or it might be the rate of rise to Cmax.  And that would be a drug-specific question.

            If, for example, you looked at quinidine, quinidine has a very poor relationship to the QT interval.  If you were able to talk about torsade, what really matters there is the rate of rise, how quickly you get to Cmax.  And if you get to a very nasty Cmax very slowly, it's not a terribly dangerous thing it looks like, but if you get to the same Cmax very quickly, then it can be a very dangerous thing.

            That's a drug-specific question, and I would caution about always using AUC.  I mean, you can think of examples related to Greg's example too.  Above a certain point, changes in the AUC of a proton pump inhibitor do nothing.  They're meaningless.  I guess that just emphasizes the point that you need to know the pharmacodynamic relationship first.

            DR. SHEINER:  True as what you say is, I quake at the notion that things as uncertain as area under the curves, which are essentially integrals and consequently smooth out error, and how you're telling us we're going to have to take derivatives, which augment error ‑‑

            DR. FLOCKHART:  Well, it's taking a smaller part of the data.

            DR. SHEINER:  We may never be able in sort of a naturalistic setting to estimate a derivative with any kind of accuracy.

            It would work at the level of a preparation.  If you had a preparation that was rapidly absorbed and one that wasn't, and that was pretty consistent, then you'd know that you'd have more danger from one than another in that sort of circumstance.  But that we'll ever discover who are the people who absorb more rapidly by sort of surveying the world and then trying to put it together across several models ‑‑ and I'm the mad modeler.

            DR. FLOCKHART:  But, Dr. Sheiner, shouldn't that come out of some models?  In other words, you would be able to see in a large population study whether people who get fast absorption get a QT longer.

            DR. SHEINER:  Well, I don't know the rate of rise because I don't know when their level was drawn.  I put it on the graph at a certain point because that's what they told me, but I can have two levels that are 10 minutes apart and they're really 2 hours apart.

            I guess what I'm trying to say is that ‑‑ and I was just being facetious, but I think we do need to temper the kinds of conclusions we hope people to draw from sort of messy clinical data.  I'm just mentioning that derivatives are really hard.

            DR. VENITZ:  But that's the empiricist talking. As a pharmacologist, I say maybe I can understand something about what's responsible for orthostatic tachycardia and it may well be that my rate of change in concentration or my Cmax is much more physiologically important, and I know that without having to do an empirical study.

            DR. SHEINER:  Right.  What I'm saying is the implications of what I'm saying, to be serious rather than just making trouble, would be if that's the kind of thing you want to know, if you're not sure about it, then you need to do a very well-controlled experiment.  You're not going to be able to learn that from the same kind of data you might be able to learn that area under the curve was the determinant.

            DR. VENITZ:  Either that or you have some mechanistic understanding how the drug concentration leads to a response.  That's the point that I'm making.


            DR. DERENDORF:  Well, I think this discussion shows that it's really impossible to even try to have a standardized approach in terms of a parameter like a bioequivalence approach, that you have a single criterion that would summarize it all up.  I think each drug, each class of drug is different, and each situation is different.  I'm not sure if we can find a standardized approach as we're asked to.

            DR. LEE:  I guess by standardized approach, we mean a standardized conceptual approach, which means we always like to calculate the probability of an adverse event, rather than saying that we're going to standardized an Emax model as the method to be used or the magical AUC or Cmax.  So, again, we're trying to standardize the conceptual approach.

            DR. SHEINER:  I think it's really worthwhile focusing on the positive here, which is this is a difficult problem and the very fact that people involved in regulation are acknowledging that it's worthwhile to try to be quantitative about things that are extremely uncertain and to try to come up with a better way to be more quantitative about a problem where there will never cease to be disagreement about any particular case because you'll never nail anything down close enough ‑‑ you'll be saying this is what we think we ought to do in terms of dosage recommendations.  And it will be based upon an information base which would allow a rational person to say, no, you don't need to do that.  That's where we're going to wind up, and to wind up anywhere else would be so prohibitively expensive that it would not justify the effort.

            So I'm extremely encouraged.  It's not as it has been in the past, hands being thrown up and we can't do this well enough, so we won't do it at all.  That's not the right attitude.  And that you're seeking advice on how to do this difficult thing I think is a very good thing.

            But I do think that some kind of standardization, for example, about turning things into probabilities and utilities in an honest way so that everybody can be on the same page ‑‑ they can all understand what you know and what values you're applying.

            DR. LESKO:  A lot of the context for the discussion in the case studies so far have been what we've seen, what has come in in an NDA, but is there a way to translate the methodology we're talking about, let's say, to a drug development program in order to get studies designed that would provide for information that would reduce some of the uncertainty that we work with in the absence of some formal recommendation to do studies a certain way?

            In other words, let's say a standardized method evolves over time, and let's say that that could perhaps evolve into a guidance on study design that would provide for information that would be better apt to provide the information that we're asking here in terms of dose adjustment and quantitation of risk.  Is that a logical follow-through on the path we're on in the minds of people?

            DR. VENITZ:  But is the uncertainty and the consequence of the uncertainty that we currently have so large that we really need to do a whole lot more experimental work, short of what you're doing right now, which is on a case-by-case basis, evaluate whether the information is sufficient for you to assess the risk-benefit, and then as in Jenny's case, recommend to the sponsor that they would have to do a larger study to look at high exposures?

            DR. LESKO:  I guess I was sort of asking the question ‑‑ in Jenny's case, for example, QTc was obtained for the first 4 hours.  Would a study design that looked at a longer period of time ‑‑ wait a minute.  Was that your case?  Well, it was one case.  Sorry, Nhi.  I should know these data.

            There was one case where the QTc was obtained for only 4 hours and blood levels were obtained for a longer period of time.  And would a different study design have provided a better basis to make the recommendations that people were trying to make with the analysis of the data?  That's sort of where I'm heading with that.

            DR. FLOCKHART:  I think the answer to that is obviously it depends.  4 hours might have been long enough for that drug, but there are other drugs ‑‑ haloperidol comes to mind ‑‑ where that would not have been enough.

            To go back to my point, I think actually it is, Dr. Lee, very generalizable.  I think there is a generalizable conceptual approach.  My point about bringing up just sticking to the AUC was just to be educated about that.  I suspect that the AUC would very often be a valuable parameter, but you have to be open to using other things when that's biologically and pharmacologically appropriate.

            DR. VENITZ:  The only thing I would add is, as you've heard the committee talk about last time, as well as this time, I think there's a lot of favorable sentiment.  As Dr. Sheiner likes to point out, it's better than what we currently have.  It beats the competition.

            The one thing that I would reinforce, though, is that it's very important to communicate it appropriately, and that has to do with all the assumptions that are being made.  Are they verifiable to some extent or not?  Do you want to err on the conservative side or on the more liberal side?  So that the people that deal with the clinical pharmacology reviews interact with the medical reviewers.  They may not understand the technical side of it, but they're the domain experts and they can follow those kind of thoughts.  So it's really a matter of risk communication in my mind more than it is the actual process.

            DR. KEARNS:  And to pick up on a point that Dr. Flockhart mentioned too, it has to be driven by biological or pharmacological plausibility.  To use an approach, a guidance across the board can create information that is not factual.

            For example, I had the occasion to look at a new molecule just last week with a sponsor to talk about a study design.  Of course, they had received some input about that study design, which included multiple ECGs over time that was coincident with the sampling time for the pharmacokinetics.  When I inquired as to the preclinical data about the ability of the molecule to prolong QT, about the only way that I could be convinced it could happen is if the structure could be inserted in the chart of the alphabet and somehow got between the letters Q and T.


            DR. KEARNS:  So what will we see when we do the trial?  We do multiple ECGs, in this case, on children.  What happens if we see a relationship come out of that that can be described by a host of models with all the appropriate variability nested in?  Will we have proven something that wasn't shown by prudent preclinical testing, or will we be finding ourselves in the midst of yet another epi phenomenon that has implications about how the drug might be used?

            So I think one has to use caution in making sure that when we do these things, we have good reason to do it based upon what we know.  I'm not saying that we will always know everything up front.  We clearly, clearly don't.  It's an imperfect science in an imperfect world. But to just cast it out there as indiscriminate use of an approach carries with it some liability that might not serve the public at the end of the day.

            DR. DERENDORF:  Just a follow-up comment to what Dr. Sheiner said.  I fully agree that it is worthwhile doing it and it is a good thing to do it.  But the standardization ‑‑ really my point was that it stops at the point where we say each drug is different and the more you know about the exposure-response relationship for that particular drug, the more we can use it to make predictions and the better they will be.  That's a trivial conclusion, but I think that's where the standardization ends.  Then from there on, really each case is different and needs to be dealt with individually.

            DR. VENITZ:  Would it be helpful, as far as the internal workings are concerned, to come up with a list of questions that you typically consider when you go through this process and for the committee to have a look at them?  I'm not sure whether the approach is something that can be unified, but maybe the kind of questions that should be asked every time you do this can be found consensus on.  Does that make sense to the committee?  So maybe at a future meeting, the questions that you would ask, what surrogate markers do you have, what relationships do you have, do you use areas or Cmax, those kinds of things that you go through every time that you have to review an NDA based on your experience.

            DR. SHEINER:  I think you can go a little further.  I think there are sort of best practices.  Maybe that's the way to think about it in doing this kind of thing.  Since I don't know anything about anything in particular, I've been dwelling on the generals of showing clearly what you know and what you don't know and somehow checking your models against your current data and so on.  So I think there are best practices in this and I think there are some things you can say in general, although I agree with you, when you get down to putting the labels on the x axis and the y axis, then suddenly you're in the domain area and you've got to talk to the right folks.

            DR. VENITZ:  Any further comments by the committee or any further questions from the FDA staff?  Larry.

            DR. LESKO:  Yes.  Maybe this is a deeper question and there isn't time to discuss it, but it does lead us down the path if we develop a standardized approach, the question that I have, in terms of labeling, comes into my mind.  Right now we put in the label descriptive information, for example, in the clinical pharmacology section about a change in an area that describes the, let's say, drug interaction or a special population change, and then if it warrants, a change in the dosage and administration section as to what to do about it.  But if you have more data in hand, i.e., the likelihood of a risk or the probability of a risk or the probability of an increase in risk or other things that might come out of a standardized approach, the question would be to what extent would this information be helpful to prescribers or would it be a distraction to the prescribers and how can we enhance labels.  Because we now know there are certain pieces of information that go into labels that at least the consumers, public and physicians tell us are not helpful to them and they can't interpret, and drug interaction seems to be one of those areas we frequently hear about.

            So we're thinking of ways of improving labels in terms of consistent language, the scope of information that goes into it, and with this kind of standardized approach, it could lead to some interesting ways of revising labels to convey different information to prescribers and patients.

            DR. VENITZ:  Any comments?

            (No response.)

            DR. VENITZ:  Okay.  Then let's move to the third example for today, which is going to be presented by Dr. He Sun.  He is a pharmacometrics reviewer in the Division of Pharmaceutical Evaluation II.

            DR. SUN:  Good morning.  I will try to discuss some general questions in my discussion, and we may switch the specific detail in numbers to general issues.

            These will be the questions I'm going to ask after the presentation, but I just put it up front to get some initial feelings.

            The first question is, if we get adverse reaction data from clinical studies, these data can be treated as either a continuous variable or a categorical variable.  Then, what should we do?  Do you have any preference, and why?  I will show some examples to illustrate it further for this.

            The second question is, in phase III clinical trials, lots of subjects, but we may not have observations in special subjects.  Therefore, the population PK approach may give us the opportunity to either simulate or predict the exposure parameter for the population who don't have exposure observation but do have response observation.  So what's the limitation and utility of this approach?

            Now, if we have a PK model based on the above approach, we get some kind of conclusion on side effect versus drug concentration relationship.  How do we make a dose adjustment recommendation for subpopulations?

            There's some limited information here to present what data distribution pretty much looks like.  On the slide here, in this corner it shows what the data distribution may look like.  It can be dense data from phase II trials or it can be sparse data from phase III clinical trials, or some kind of a combination with an unbalanced situation.

            Now, the safety information can be either a single critical key adverse reaction parameter which is a continuous variable, like QTc variable or high blood pressure and so on.  But it can also be a categorical variable like pain or "yes or no" for liver toxicity and so on.

            But these two actually are switchable.  Let's say blood pressure.  You can set up a cutoff marker that says if above such and such, it is abnormal, below such and such, it is normal.  So the continuous variable actually can be changed to a categorical variable.

            Now, a categorical variable, although it can be a "yes or no" situation, but for the group with "yes," you can also give a score of 1, 2, 3, 4, 5 or from 0 to 10.  So it becomes some kind of a continuous variable.

            So these two actually have no clear cut.  That's why I ask this question in this presentation.  If you have this situation, which one do you prefer?  Of course, this will also change your data analysis process.

            So you can also have a combination of both with multiple ADR observations.  But for phase III trials, pretty much we have this kind of situation:  mixed types, multiple ADRs, and unbalanced.  Then that is why population approach can play here.

            Let me first show some data sets.  Then we go back to see what analysis process we can apply.  This data set is used just to illustrate the question or the process I mentioned before.  Forget the exact numbers and the true terms.  Sometimes I have to modify this.

            Let's say we have two clinical trials, very big size, 1,500 evaluable treatment patients.  And we also have multiple dose levels from X to 4 times higher.  And the patient plasma drug concentration was measured, although for some are dense and for some are sparse.  Therefore, the total data set is kind of unbalanced.

            Then we have endpoints for safety measurement.  This can be some blood chemistry variables which usually is a continuous variable at the beginning, but a clinician can define some value as a cutoff point shows this variable as normal/abnormal to claim at such situation there's no ADR and others will be ADR.

            The ADR can be also present as a "yes or no" situation for some, like headache, liver toxicity, phototoxicity values.  And this variable again can be changed to different scores for the extent of headache, like mild, moderate, or severe.

            Now, PK results.  Let's say drug-drug interaction causes AUC to increase by almost 300 percent.  The Cmax changes by 150 percent.  AUC and Cmax may also be changed by age, gender, or so on and so forth, even between ethnic groups.

            The safety results.  We will not focus on efficacy in the presentation.  We will only focus on safety parameters.  Safety parameters usually are very, very small in percentage and very sparse.  So there are some cases where you never have any so-called "maximal effect" for side effect terms.

            The efficacy results.  Let's make this discussion a little simple.  We see efficacy has no such exposure-efficacy relationship detected although we see there's a demonstration of clinical efficacy in total.

            Now, with these data sets, let's come back and see what process we usually can do.  First of all, of course, there are managing/editing data processes we can use.  For this part, we start to have a question:  shall we treat the data as a continuous variable or should we treat the data as a categorical variable?

            Then we can conduct a population PK analysis based on exposure data such as building a base model, add variability, add covariate, and so on and so forth.  Now, there's one problem:  if we cannot find a significant covariate in the PK model, then the next step for predicting E for the new population or new individuals will havew a little problem.  But let's say we have the model built and the model validated.  Then we can go to the next step and determine individual exposure or subpopulation exposure parameters.  There are two parts here.  We can do post hoc for the subjects who are already included in the study, or we can do a simulation trial to determine exposure parameter for the population who was not really in the trial or the observation was not available in the particular patient.

            The next step will be to derive secondary exposure parameters, such as AUC, Tmic, effect compartment concentration and so on.  And another important factor here I want to emphasize is that the accumulative exposure parameter can be estimated and determined, like say what if after multiple dose or long exposure situation.

            With these exposure parameters available through the above processes, we can determine exposure-response relationship for individuals or for special populations.  We have lots of methods here.  We can use classification method.  I will give you some examples later on.  We can do classification or regression tree analysis, logistic regression for binary data, and so on.

            Then we consider the accumulative exposure time, then perform statistical analysis on response data.  Now, this again correlates with what do we do with the data?  If our data is a continuous variable and we have odds ratios with uncertainty built in, we can do statistics.  But sometimes if the variable purely is categorical and is divided to either above the mean or below the mean, the statistics will be hard.

            Now, with all this situation, the next step, finally we will make a dose adjustment.  What are we going to do especially if we have multiple variables?  Let's say the exposure-response depends on age, gender, body weight, and blah, blah, so on.  If we take all of this condition together for making a dose adjustment, it may be too complicated in drug labeling.  Shall we only consider the one which is critical, or shall we consider the one which has most frequently occurred, or some other method?  I want to hear some discussion on this.

            Let's see some results.  If we're dealing with this process, what result can we get?

            Classification.  The first part we can see is to divide the whole population into some equal populated segments.  Let's say every 25 percent subjects from low exposure to high exposure.  Or we can divide this whole population into equally distanced segments, the percentile; that is, the first 25 percent in concentration, the second 25 percent, and so on and so forth.  Then we count what's the frequency of ADR.  For example, the results can be presented as total ADR is 18 percent if AUC is greater than the mean and only 5 percent if AUC is less or equal to the mean value.  That makes the whole discussion for this kind of classification results.  Of course, there are lots of pros and cons.  I really want to hear a discussion later on.

            Now, the second one is that we can do a classification analysis based on severity.  Let's say we reclassified R values, the response values, as different class or different scores, as severe, moderate, or mild, and so on.  Then we see the example.  Severe ADR occurs if Cmax is greater than 10 but only mild ADR is apparent if Cmax is less than 2, although the frequency probably between these two has no significant difference.  But in this situation, we see it looks like 10 is some kind of cutoff value to avoid severe ADRs.

            Then we can throw this data into a computer to search for the best maximum split, maximum split distinctions between R values as by a classification tree or regression tree.  One result I present here, for example, the first split on ADR frequency was at AUC equals to 871.  So when the split occurs on ADR, it was 23 percent if AUC is greater than 871 and would be 2.5 percent if AUC is less than 871.

            Next, we will do modeling work.  We can do modeling work for the same data set for different ADR or the same ADR parameters.  When we do modeling work, there are several ways.  I do not want to discuss further on this part.  I only want to show the possible ways.  We can base on purely statistical models to do the modeling work, or base on some kind of physiological-based, meaningful models.  There's a lot of discussion on the pros and cons for each.  But let's go to the next one.  Our model can be a linear model or a nonlinear model.  Of course, there are uncertainty parts when building nonlinear models, adding fixed effects and random effects, again, with this model.

            So two examples.  We can do a simple regression analysis based on so-called logistic regressions.  So, for example, we can get a result with even a 95 percent confidence interval for the logistic regression results for odds ratios.  For example, we can see the odds ratio for acute tissue rejection increases 23 percent if AUC 0-24 decreases by 10 percent.  That's one way to present this data.

            Then we can treat all the data as a continuous variable, do as the next few examples for modeling work.  We can get an equation that says the HDL drops below normal on day 95 if the concentration, average concentration, is greater than 10 micrograms per ml on day 95 for patients with high body weight and low initial HDL at baseline.  So there's a covariate effect built in and it also has this kind of a drug concentration curve profile.

            The back pain we treat as 0 to 10 scores, some kind of semi-categorical variable.  It's nonlinearly correlated with plasma drug daily AUC and the number of treatment days and the dose regimen.  So three factors.  The result here found was b.i.d. actually has less incidence of back pain.  T.i.d. will have more, although the total daily doses are equivalent.  And the number of treatment days significantly correlate with or predict back pain scores.

            QTc prolongation.  Now, we can find some models.  The relationship between QTc and the drug concentration can be described by an Emax model.  Then you can find some E-R parameter like E0, Emax, ED50, and so on.  But these parameters can be, as we discussed before, correlated with either gender or age or some other subpopulation variables.

            Phototoxicity.  Now, we only have two variables, yes or no.  We get a "yes" value on day 10 if C average on day 10 is greater than 8.  This will occur only in female subjects.  So this can be due to either data limitation or this really is a true result that female subjects are more sensitive to the drug in terms of having phototoxicity occurring.

            Liver toxicity.  We can get some kind of frequency linearly correlated with C in the initial few hours of time of exposure.  It may not correlate exactly with the Cmax, but it correlated with the initial range of concentration average.

            Blood chemistry.  Now, blood chemistry actually is a continuous variable, but we can treat the blood chemistry variable as a categorical "yes or no," normal or abnormal, or give them a score from 0 to 10.  Now, if we have a score from 0 to 10, we can get some kind of correlation with plasma concentration, either AUC or number of treatment days.

            So these are examples I want to show.  So one single variable can be treated by different ways, but in one study we have lots of different variables, and different variables may be treated by different ways, and use different data sets as different base for information.

            And there are others, probably we never find any relationship, never can find a cutoff point.  Descriptive.  Here I give some examples.  Some we just can have a definition of normal/abnormal values, but may not see any difference between different subgroups like these four examples I show here.

            So now we come to the end.  We have all the information, by all different methods, with different bases, from different data sets, so on.  We definitely already used population PK, did it two times.  One is predicting exposure parameter for subjects who do not have observation in exposure but do have a response measurement, and in the second part, use a population approach, nonlinear mixed effect modeling, to show whether the E-R relationship depends on some other co-variables.

            So with all this population of subjects and information, now we will make a dose adjustment.  See, for example, we can do this:  the average upper therapeutic limit is probably around 10 for Cmax and 871 for AUC.  Remember, these two variables are gathered from previous toxicity analysis.  And the female subjects seemed the most sensitive to phototoxicity, and the concentration average should be less than 8.  Then we see a b.i.d. dose regimen is preferred because it reduced one of the particular toxicity results.

            Now let's go back to my questions with all the data we have seen in the examples.  First is what are the utility and general limitations of linking PK obtained from population analysis to response endpoints?  And what are the general considerations in E-R based dose adjustment for special populations?  And should we treat this data as a continuous or categorical variable?  What's the preference?

            Again, I really want to hear a discussion more focused on the general concept and ideas based on the experience you have and let us know the pros and cons for each situation rather than focusing on the numbers because I have modified the values somewhat to make the presentation smooth.  Thanks.

            DR. VENITZ:  Thank you, He.

            Any questions about Dr. Sun's presentation before we delve into his proposed questions?

            On one of your slides, you mentioned Cint.  Can you tell me what that meant?

            DR. SUN:  This is the initial concentration exposure.

            DR. VENITZ:  Oh, initial concentration.  Okay.

            Do you want to pose the questions and then let the committee bat it around?

            DR. SUN:  Okay.

            DR. VENITZ:  The first question regarding the limitations of linking PK from a population analysis to response endpoints.  Do you want to elaborate on that question what specifically you had in mind?

            DR. SUN:  Okay.  This question was the utility and general limitations of using population PK for population for this analysis.  As I mentioned, we have two places we can use nonlinear mixed effect modeling work for doing data analysis for this data.  The first part is in clinical phase III trials, we may not have a concentration exposure measure for every subject, but we do have a response measure for every subject.  In this situation, if we have sufficient information, use the population PK, get the model, then we can predict or estimate concentration or other exposure parameter for the population we see in the phase III trial.  Or in some situation, patients only have one or two trough measures and we want to determine the total exposure and the time of exposure.  So this is one place population PK can be used.

            The second part is after we have E data and R data, either categorical or continuous variable, now we can use mixed effect modeling to see whether these two variables are related to each other based on some covariate factors.

            So that's my question.  What are the utility and general limitations on this if we do it this way?

            DR. SHEINER:  I can't decide if what you're asking is, is there a manual for how you treat any given set of data to come up with the conclusions that you're going to find most believable.  We can't address that.  So when I look at that first question, your description of what you might do sounded sort of like something I might do.

            The only thing I can say, the only serious general limitation about which even good data analysis cannot help you is the problem of confounding.  Both the PK and the responses are endpoints, are outcomes, and whenever you try to relate outcomes to outcomes, you have the problem that you can't tell which way causality runs.  And you base that conclusion, if you do, on external information in the way of science or other things.  You can't tell it from your data.  So that's the limitation.

            Now, that doesn't mean we don't proceed every day, based on observational data, to make the most important decisions in our lives.  We do.  But we have to understand that in a regulatory context, there are other forces operating.  You want to be cautious in certain ways.  And that's the serious problem.

            Most of the other stuff, it seems to me, that you brought up were technical issues.  And I'm not sure that we really want to spend ‑‑ even though I'm a real techno-wonk when it comes to model-based analyses, I don't think you want to hear me dilate on that.

            I think the basic thing there is if you get different conclusions when you treat your data as categorical versus continuous or when you use one kind of a model or another, then there's something wrong.  So you ought to get all the same conclusions.  What happens as you turn data from continuous, if it's got a lot of information, to categorical, you're losing information.  So some things will drop out.  Some things will appear no longer to have relationships that did appear before to have relationships.  That's got to happen as you limit the information in your data.

            But other than that, I think you want to basically use the data representation that's most relevant to the people who are going to use it and that keeps the information, et cetera, all the good rules of modeling.  But I don't think we can get into too many details.

            So maybe if you have particular instances where you think, looking at data in different ways, the same data in different ways led you to very different conclusions, I think that's something that I might be interested in hearing about.  Otherwise, I don't know what we can say in general.

            DR. LEE:  Can I rephrase the question a little bit?  The reason we're making this presentation is because phase III studies actually present a unique opportunity for us to look at exposure and response, especially the safety endpoint that we don't frequently observe in a phase II study which is too small to capture a rare adverse event.  That's why we wanted to ask the committee whether a population approach would be a good approach to look at exposure response in the phase III studies.

            However, there may be some limitation in terms of study design.  For example, we may not have enough plasma samples or maybe the sampling time between the PK and the pharmacodynamic endpoints is different.

            So this is the type of question we're trying to ask the committee, whether internal study design, whether there's any limitation to conduct such type of a population PK analysis.  And if there are certain limitations, can we recommend to the sponsor to design the study differently in the future so that we can get a better quality PK/PD relationship out of the phase III studies?

            DR. SHEINER:  Let me just say one more thing about that.  So you're talking about wanting to use this confirmatory study for learning purposes.  And there are certain kinds of learning data elements that don't interfere with your design that would make your life a lot easier.  Whether they're worth it or not, you can only say afterwards.  But measuring things serially, whether it be toxicity or efficacy or both, rather than just the 6-month and 1-year endpoint, or whatever it is that was the primary thing; measuring compliance, that is to say, what drug did they actually take; measuring PK.

            I think, by the way, my guess is that adherence is more important an influence on outcome than PK is for most cases.  But when they say measuring PK, you want to measure that in the case where adherence is assured so that you have those as two separate variables.  And so on.

            Basically the idea is measuring biomarkers, whether they be adherence or chemicals, you know, along the causal path from the prescription to the effect, and measuring them serially over time.  That's the best you can hope for without changing the design radically.  And if you want to be able to do these kinds of analyses, that's the kind of data that you need.

            But techniques for dealing with missing data, techniques for dealing with other problems that arise, with mixed kinds of data, both continuous and categorical, and so on, those I don't think are essential.  They exist.  They make your life a little tougher or a more interesting, depending on where you come from.  But they're there and you should be able to get the information out of the data.

            DR. SUN:  He was asking whether I have an example regarding how the data can be treated either continuous and categorical.

            DR. SHEINER:  And reach different conclusions.

            DR. SUN:  Yes, reach a different conclusion.

            Let's say this situation.  If you treat data as a "yes or no" situation and you divide the concentration distribution to be above the mean and below the mean, do you will find the only conclusion you can get is the frequency of "yes" when concentration or AUC above the mean is 23 percent.  If lower than the mean, it will be 5 percent.  That's all you can get.

            If you treat it as a continuous variable, you can get some kind of sigmoid models, correlation between scores of adverse reaction versus the concentration.  Then from the curve, you can pick up ‑‑ say you want to limit less than 10 percent of subjects has a score less than 2 ‑‑  a concentration.  So this becomes a different decision.  And the curve becomes a smooth curve.  You pick up a point at which you limit two factors.  Percent of subjects reach a score of XYZ.  Compared with the first one, you only can get a result if above the mean will be such and such, if below the mean will be such and such.

            Then in labeling, it will be different.  In labeling, when make a dose adjustment, say due to drug-drug interaction, if the population Cmax change, still somewhat below the mean values, for the overall population, or you can get a feeling what the frequency of side effects due to drug-drug interaction will be.     But in the second situation, if you have a continuous curve, you can estimate when concentrations switch a kind of 10 percent, what's the percentage of patients will have a score of 2 or 3 increase by such and such.

            So when we recommended these two suggestions to clinical or to the labeling committee for NDA review, these two really makes different.  That's why the question comes.  Any parameter really we can treat as one of, or we can switch between the two.  And what really we want to do?

            DR. DERENDORF:  I think as was said earlier, whenever you move from continuous to categorical values, you throw away information.  I think it comes down to a compromise that you have to make with the information that you have and how you want to communicate it.  If you make it too complicated and include everything you know, nobody is going to use it.  So you have to find a way to focus on the important things, but still come up with an accurate conclusion.

            You have a great example that you can overdo it and make it too simple.  One of the conclusions you have here, the total ADR is 18 percent when the AUC is above 1,200 and 5 percent when it's below.  That may be true for the data set that you have, but it's totally useless for someone who wants to extrapolate it for a certain situation because obviously you can have very, very low AUCs that wouldn't have any ADRs or you could have very, very high where the probability would be much higher.  So that would be a misrepresentation of the actual information that you have by making it too simple.  So I think there's the trick that you have to find the right balance.

            DR. LESKO:  We're in the world of safety here, and it would seem like in the clinical trial design, there's going to be prespecified endpoints of safety that the sponsor provides.  It also seems to me that there's going to be a convention for a safety biomarker, let's call it, that will be continuous or categorical.

            For example, if my concern with the drug is heart rate, I'm going to look at that maybe as a continuous variable because that might be the way it's measured and the way it's analyzed.  On the other hand, if I was looking at a hematological toxicity like neutropenia, I might be concerned about grading the severity of that.

            So I guess the question becomes, given there are certain prespecified endpoints in terms of severity and frequency, and given that there's a conventional way of presenting data as continuous or categorical, what would be the motivating factor to change continuous to categorical? What would the benefit of that be in terms of the pragmatic aspect of it, in terms of dose adjustment?  Because you think about a drug that is going to be used therapeutically and being monitored by these same endpoints and doses being adjusted by those same endpoints.  So it's sort changing something from the usual to the unusual, and you'd have to say there's a reason to do that.  And I guess it wasn't clear to me what the reasons for me to do that would be.

            DR. SUN:  I will give one example.  Liver toxicity, one situation we saw in one NDA.  If you see the frequency of liver toxicity based on concentration above a cutoff point, below a cutoff point, you don't see a difference.  So it looks like all concentrations show ‑‑ if you only define liver toxicity as a yes or no parameter, you don't see a correlation between these two.  But if you use actually the measurement of blood chemistry values, which is used as an indicator for liver toxicity, you're able to correlate concentration with the blood chemistry variable.  So clinically it may be easy to see, well, 9 subjects or 10 subjects have liver toxicity, but we don't know correlation with concentration.

            But on the other side, you can see actually when concentration or AUC increases, the chemistry value have some trend increase.  Then you define a cutoff at that point that says on the curve where is the cutoff, then where is the concentration's cutoff?  So there are two ways to present this data.  The same data set we can do two ways.

            My own experience in the drug labeling is that it looks like the first way is easier, the second way somehow, like you mentioned, has to be communicated to others in different professions as well.

            DR. FLOCKHART:  I would just emphasize that point.  I think obviously you lose power with a categorical variable, but if you have a change that is picked up by the categorical variable, as well as by the continuous one, in general it communicates better.  And we have a humongous problem with communication.  That's an understatement.  So I think when it's possible to state something in stark, clear yes or no terms to somebody who's practicing medicine in the area of drug interactions or recommendations within a label ‑‑ we're all aware of the vast majority of the label is just dust anyway.  So when you can simplify it, that helps.

            DR. VENITZ:  Can I just make a general comment?  Looking at some of the, I guess, labeling language or some of the statements that you reviewed with us, He, the only time they're going to be useful for a practitioner is if they actually draw blood levels because otherwise the fact if I'm above some certain level or some certain area, something happens or doesn't happen, it won't help me as a practitioner.  The only thing that I want to know is can I change the dose and how do I change the dose.

            So when you translate the information that you get from what you just reviewed for us and translate that into labeling language, you really have to target doses not concentrations, unless part of the therapeutic management with this particular agent requires dose titration based on plasma levels.  Otherwise, I don't see how that is useful information for the practitioner to translate into practice.

            DR. SUN:  You prefer dose-response relationship rather than concentration-response.

            DR. VENITZ:  The only thing that the clinician can change is the dose unless part of the management requires taking blood levels, and then depending on the blood level, I can make certain adjustments.

            DR. LEE:  Can I say something here?  I think what we propose here is we're not going to use a different approach either as population analysis or regular PK/PD analysis.  So what we're presenting here is can we use population analysis to get a PK/PD relationship.  But once we get a PK/PD relationship, we're going to follow the standardized approach to estimate the probability of an adverse event for special populations.  So that's what we're trying to propose here.

            DR. VENITZ:  The statements that he reviewed for us, at least most of them, have some statement about levels or areas that are too high, too low.  And I'm saying from a practitioner's point of view, unless part of the management with this particular agent requires drawing blood levels so I can actually measure levels, it's useless.  I need to know what to do with my dose because that's the only thing I can change short of changing the drug itself.

            DR. SUN:  But on the other side, let's say, same dose level due to drug-drug interaction change the concentration.  This information will give us an idea if the concentration changed by such a degree, what's the probability in the ADR will be.  So in terms of decision on this side, we still have to rely on this rather than only rely on dose.  Although at the end, we can see ‑‑ the end and the level and we see in terms of drug-drug interaction, you have to deduct the dose by 50 percent, but a 50 percent deduction was got from concentration dose-response relationship.

            DR. VENITZ:  I don't have a problem with your conclusions.  I'm questioning the usefulness of incorporating the conclusions as they are in a label to convince a practitioner to change a dose.  That's all.

            DR. SHEINER:  Getting back to this issue of dichotomous versus continuous.  There's a big difference as to whether you translate ‑‑ and I think this is what David was getting at ‑‑ from continuous to dichotomous, let's say, on the way in ‑‑ that is to say, you change the data and then analyze the data that are now dichotomous ‑‑ versus on the way out where you take this complex model that deals with all the variables in their full complexity and then draw a very simple conclusion based upon simulations of that that says half the time it's going to be bigger than this if you do that.  There's no problem with the latter, and that's very important for communication, to make things simple, to talk about the key issues.

            The problem with the former is that you're losing information.  The places you see it worst are in these clinical scores where you have 15 or 20 questions that you ask and then you add up the number of yeses and that's a number.  They're clearly not combined optimally.  You don't know whether one or another question would have more information than other things.

            What they do is they allow you to do your analysis simpler.  The only time that in my mind it's justified to simplify the data before you analyze them is when the loss of noise exceeds the loss of signal and you can afford some loss of signal and you're losing more noise by dichotomizing, let's say.

            And I think most of us actually are probably ‑‑ I don't know.  It would be interesting to take a survey of how many bits of biological information people believe there is in something, let's say, that's three significant figures like a serum sodium.  Is there really the 12 or so bits that it takes to represent a three significant digit number or is it really only three or four bits?  The sodium really 140 to 143 is all the same, et cetera.  I think there's probably a lot less information in these continuous variables than we think there is because of the high feedback systems that we're dealing with.

            But the main point, it seems to me, is the three questions that I always say you've got to ask before you do anything.  So this one is, what's the question?  If you want to write a label that says that as you increase the dose by these increments, the probability of this toxicity will go up by those increments, then you have to have a model that somehow represents continuous probability versus continuous dose.  If you don't want to write a label that says that but only says don't give the drug to people who have values beyond these, then you don't need that.  So first figure out the question.

            There I think maybe is where one could sometimes schedule a meeting to talk about that.  What should recommendations for dosage changes in special populations look like?  What kind of statements ought we be trying to make?  And that will determine the data we want to gather and how we want to analyze it.  I'm not sure that's all settled, but I'd like to hear at some point what the agency does think about what constitutes a complete set of statements and what degree of precision and what kinds of words you use and what kinds of things you want to be able to say.

            DR. VENITZ:  Any comments, Larry or Peter?

            DR. LESKO:  No.  We've sort of thought about that last statement that Dr. Sheiner brought up.  I think it's a very valuable statement, but I think we need to think about it and put the story together and bring it before the committee.  But I think it would be a very interesting discussion to have, what elements of a good label should there be in terms of a probability of risk and an intervention of some sort and how do you present that in a consistent way across special populations or something like that.

            DR. VENITZ:  A couple of more comments, I guess more general comments, not specific to your question.  But how to translate information that you would gather from this kind of analysis and translate them into labeling language.  How is the drug being administered?  Is it titrated on some kind of effect?  That would make a big difference in terms of how some of those things would translate into a dosing recommendation because you may not have to make a dosing recommendation because you're going to pick it up as part of the normal way of therapeutic drug monitoring.  And it might not be a drug level.  It might be some surrogate marker, some biomarker.  Obviously, what are the available ways of adjusting the dose?  Do they have the dosage forms to accommodate that or can you not?

            And then along the titration route, do we know anything about intra-individual variability that would allow us to assess for a given patient how likely they're going up and down, and is our dosing algorithm going to pick that up?

            Those are questions that aren't really addressed in what you're talking about, but I think they're very relevant for translating this information into recommendations in the labeling.

            DR. LESKO:  Just to add to that in thinking about maybe a future discussion with the proposed labeling rule that the agency has out, there's going to be some revamping of the label such that certain information gets to the top of the label out of the individual sections.  And a question could be raised as to what criteria in the context of what we're talking about would warrant raising information to a more prominent part in the label.  If that is done for catching the attention of a prescriber, it gets to that language and how you present the information in translating it into therapeutics.  So I think the whole thing sort of flows as a future issue.

            DR. SUN:  Yes.  We do see situations where a label can be ‑‑ let's say it's an IV formulation or all doses are available.  You can put a table.  It shows about every year with a different dose.  But when the formulations only have 15 and 30 milligrams, you only can do is categorize them to two classes.  That is really true.

            Let me summarize.  As will be ‑‑ if we translate from continuous to categorical, we lose information.  But on the other hand, maybe categorical is easier to communicate.  And it also depends on the outcome, what are you going to do before you handle the data.  Before you even start an analysis, how do you use this information.

            Thank you very much.

            DR. VENITZ:  Any other comments?

            DR. DERENDORF:  Maybe it's way out there, but it seems that one problem is not so much the information and the data analysis, but it is really the ability of the user to do something with it.  I think maybe one of the reasons is that we're limited in the label to just a written document, and with modern technology, a lot of that information can be presented to someone in a palatable way like in a little computer simulation where all the information is entered, and then there's a recommendation or an assessment that comes back.  That may be something to consider in the future.

            DR. RELLING:  I'm new to this, but you talk about special populations and you've presented different examples of covariates.  is there some sort of list that you have of what you consider to define special populations and what you consider covariates that should always at least be asked about?   We heard about ketoconazole and something else, age or renal dysfunction.  Are you letting individual studies drive these things?  Are you going through some algorithms to define what you should look for a priori?  How are you deciding what you'll even include or think about looking at?  Because there are all kinds of examples where we're making dumb mistakes like looking at ketoconazole but not itraconazole.  How are you going through this stuff?

            DR. SUN:  The first part you're asking how to define the special populations.  In the legal terms, they define special population as disabled patient/subject, blind subject or some others.  In our clinical pharmacology term here, we refer to subpopulation like a pregnant patients, pediatric subjects, or other ethnic groups rather than the legal term defined those special populations.  That's the first part.

            The second part.  When we do clinical trials in phase III trials, we do have some inclusion/exclusion criteria.  A majority of the time it's how the drug in the future will be given to patients in the clinical setting.  So most likely we will include other subjects who potentially will take this drug.

            Did I address your question?

            DR. RELLING:  Not really.

            DR. LESKO:  Let me add to it.  First of all, the answer to the question is somewhat drug dependent, but there is a standard range of assessments that are expected within the drug development program.  And most of these are revolving around the changes in pharmacokinetics.  So, for example, there are demographic factors.  We would want to know age, gender, and race and the effects on pharmacokinetics and whether that's pertinent to dosing adjustments that might be necessary in those subpopulations or special populations.

            We then move next to intrinsic factors, as we call them, and they're predominantly disease states that handle drug disposition.  So renal disease and hepatic disease is a standard study that's in most NDAs.

            And then finally the issue of extrinsic factors and predominant amongst those are co-administered drugs.  So there's a heavy emphasis, ideally mechanistically driven, to look at clinical drug-drug interactions that are likely to be important in the clinical setting.

            So the categories of special populations are defined by these demographics, the intrinsic, and the extrinsic factors.

            Now, depending on the drug used, there may be additional special populations that would be looked at, but I would say the ones I just mentioned are the standard covariates that you would be interested in.

            DR. RELLING:  Do you require those to be addressed a priori in the trial?  And what makes the difference whether you decide ‑‑ when are you going to start looking at genetics?  Where do you start drawing the line of saying you've got to look at something besides a creatinine and a bilirubin or whatever it is you are asking for?

            DR. LESKO:  "Require" is kind of a harsh word.  I'd like to think of it as recommendations.  By and large, everything I just said is contained in guidances to the industry that say what the expectations of the agency are, and if those types of analyses aren't available, the burden of proof then is on the company to say why it's not important in the case of that particular drug.

            When you get into what I would call evolving areas or evolving covariates ‑‑ and you mentioned genetics I think as one of them ‑‑ we have now appearing on sort of the drug development scene the ability to look at changes in DNA sequence that would describe what we typically have called phenotypes, poor metabolizers and extensive metabolizers.  It's a logical extension, to me at least, to begin to ask the question, given the availability of a test to measure a genotype, at least answer the question somewhere in drug development, is that an important covariate.  It's no different in my mind.  The fact that it's genetic doesn't make much difference to me in a sense because it's a covariate that could be identified that may have a significant effect on exposure and then subsequently response.  So it would seem to me you either need to have some information that would say I need to worry about it or I have information that would say I don't need to worry about it.

            If one thinks about the size of special populations, certainly a genotyped group, i.e., a poor metabolizer for a 2D6 substrate, is much larger in size than would be, say, a patient population defined by their renal function.  So I think we're moving in that direction as the science evolves and as our ability to identify these covariates becomes more commonplace and available.

            DR. SADEE:  It appears to me that these models all still assume exposure in terms of how much drug is there.  So the alternative is to look at special populations that have a clearly different dose-effect relationship.  So what I'm struggling with, in terms of trying to understand how to simplify this and how to still get reality in there, is if you have two different populations with the outlier, the toxicity occurrence, is because somebody is genetically or environmentally predisposed in a way that cannot be predicted by the model, that it's just looking at exposure, how much drug is in the body.  And if you try to merge the two, you're obviously making an error.  On the other hand, it's clear that even for those patients, the more drug you have, the more likely it is that there would be an adverse effect.

            So I think we're talking about two different models here.  One assumes we have this relationship between how much drug is there and the effect, and the other one would be there's a completely different relationship.  Those two have to be merged, I suppose, without then incurring too large of an error.

            DR. LESKO:  Yes.  I think that sort of brings home a very interesting point, and we kind of talk about it but don't know exactly what to do with it.  And that is the paradox of drug development, as I call it, that drug development revolves around population signals and population data, whether it's the efficacy signal, the safety issue, or the pharmacokinetics.  You're looking at a population average.

            Yet, as Dr. Sadee has mentioned, when you're treating patients, you're worried about the individual situation and how you can best optimize a dose in the individual.  When we talk about individualizing or optimizing dose, I'm sure we're talking about it in the context of a subpopulation or special population defined by something I could measure or observe as opposed to the individual that genetically may be predisposed.

            I think it's easy to focus on pharmacokinetics and much harder to focus on the factors that influence receptor sensitivity or the things such as long QT and things of that sort because they're a little bit more complex, especially in the polygenic nature of being more complex.

            DR. VENITZ:  Any further comments?

            (No response.)

            DR. VENITZ:  Then it looks like we're going to get an early break.  We're going to break from, I guess, right now until 12:45.

            We have nobody signed up for the public hearing.  So we are starting at 12:45 with Dr. Karlsson's presentation, the example number 4, in using exposure response to recommend dosing adjustments.

            So enjoy your lunch and we'll be back at 12:45.

            (Whereupon, at 11:40 a.m., the subcommittee was recessed, to reconvene at 12:45 p.m., this same day.)






















                     AFTERNOON SESSION

                                               (12:50 p.m.)

            DR. VENITZ:  Can we reconvene the meeting, please?

            Our next presentation is Dr. Mats Karlsson, the faculty of Uppsala University of Sweden.  He is going to talk about optimizing dosing strategies for defined therapeutic targets.  Dr. Karlsson.

            DR. KARLSSON:  Good afternoon.  I'm really grateful for this opportunity to get some insight into the American regulatory process.

            So I'm going to talk about optimizing dosing strategies for defined therapeutic targets.  I'm going to talk about target definition in relation to dose finding.

            We have done some work, mainly simulation work, but then also applications to a few real drugs under development, and I'm going to focus more on those as examples of what we've been doing.

            The dosing strategy alternatives that there are are, first of all, the single, one dose fits all always, which of course is very convenient for patients, clinicians, and producers, if it's appropriate, but often variability and pharmacokinetics and pharmacodynamics will lead us down the individualization route.

            There are two types of individualization. First, we can individualize based on patient characteristics, observable, or feedback individualization based on some measurement or observation of the patient, or we can have a combination of these two.

            So the next thing is the defined therapeutic target.  Anybody who is involved in decision making on dosing strategies have to have some implicit target concept.  This might not always be quantitative.  It might not always be stated, but there has to be some target.  And if we were to quantitate it and actually spell it out, we could base it on, most reasonably, the weighted balance between beneficial effects and side effects.  You could also consider other endpoints like only one side of the coin or drug concentrations or biomarkers, especially when it comes to individualization in subpopulations, as we have been discussing today.  And, of course, the target might differ between different patient subpopulations.  We want to treat more severe disease maybe more aggressively.

            We need not only to know the target, but also seriousness of deviation from the target, and we can use an all or none criteria which we recognize from pharmacokinetics as the simplest concept of therapeutic window where all concentrations within the therapeutic window are equally desirable and all concentrations outside equally undesirable.  And the pharmacodynamic correspondent is the responder/nonresponder concept here.

            But we usually think that biology works in a more graded manner and we might want to actually value the deviation from target in a more graded manner.  Of course, what's important there is the clinical picture, but we might approximate that with various statistical distributions.  And also the seriousness of target deviation may vary between patients, although I haven't seen any examples of such applications.

            When I'm talking about the target concept and the penalty function, of course together these form the utility function, and I understand that that was the topic of talks at the last committee meeting.

            So one scenario for selecting a dosing strategy would be that it's based on some implicit criteria on the target and penalty function, and then that more or less stops there.

            Another scenario would be the same thing, but then take it a step further.  If one has a population model for the dose, two-target variable, one could actually estimate based on the decided dosing strategy and the model, what target and penalty function does that correspond to and then assess whether it seems to be a reasonable target, and if it is, stop there.  Otherwise, revise your dosing strategy.

            A simple example that we came across when collaborating with a drug company was a drug under development for a disease which had two harmful events.  The frequency of one was decreased by the drug and the frequency of the other was increased by the drug.  Of course, what exposure, what dose you would choose would depend on how you evaluate these two against each other.  If this effect is deemed more harmful than this, you would choose a low dose.  Otherwise, the opposite, the high dose.

            So we did some calculations.  So the black line here corresponds to what I was just saying.  If we weight the adverse event high compared to the event that is beneficially affected by the drug, then we would choose a very low dose, and if they're equally weighted, we would choose a much higher dose.

            In this case, the project team had already selected a dose of 1, which corresponded to a weight of 3 to 1 for the two events.  And when we presented the project team with this, they said, well, this seems to be a reasonable weight.

            However, there were two sub-diagnoses that had actually different PK/PD relationships, but they had decided to go with the same dose in both subpopulations.  So, of course, that meant that the weighting was different for the two subpopulations.  So in one it was 4 to 1 and in the other 2.5 to 1, and again the project team said, well, that's reasonable because in the sub-diagnosis, when the harmful event occurs, it's more serious than in the other.

            Then finally for renally impaired patients, a dose of a quarter of that for the main population was selected and it corresponds to the same weight between adverse event and beneficial event. 

            So this is a way to rationalize a selected dosing strategy after the effect. 

            Of course, one might have more use of defining the target and penalty function beforehand, and I'm certainly no expert in this area, but what seems to be wise is to ask a clinician because they're really the ones who are sitting on the information here although maybe not used to formulating these type of functions in quantitative terms.

            If there is a drug first in class, consult preclinical phase I data to assess what tolerability issues there might be.

            Consult literature and marketing which might have done surveys in patients and clinicians of what are deemed important features of a drug therapy for a certain disease.

            And then develop a few alternative targets and penalty functions, and apply it to historical data, maybe make up a bank of hypothetical patients to be ranked.

            And then again ask a few clinicians about the developed utility functions.  Most likely they won't agree, which might be a source for revising the utility function, but even so, you might not always get full agreement between different clinicians, and just as Lewis said earlier today that we need to incorporate uncertainty in all our aspects, maybe we need to include uncertainty or variability in the utility function as well.

            So if we actually have a defined utility function, then we can proceed in a more rational manner.  If we have the utility function, we have a population model for the dose-to-target variable, we can estimate the best dosing strategies given different constraints such as we're going to give everybody the same dose.  We're going to individualize, but only with two doses and based on a covariate, or we might individualize based on feedback, et cetera, and then select the dosing strategy based on target fulfillment and practical considerations.

            So what we would do in more detail for the first step, if we want to optimize the one-size-fits-all dose, would be to, based on the utility function and population PK/PD model, we would actually maybe not need the full PK/PD models that we usually use if we're only considering steady state concentrations' relation to effect.  We might only need the model for clearance.  Covariate models are essentially superfluous because they're not going to affect our dosing, and we would, based on these models, simulate a large number of hypothetical patients.  We would obtain a prediction of each individual's deviation from the target for a certain dose, and then obtain the optimal dose by minimizing the overall loss.  In this case, we can do this simply by just repeated simulations, trying different doses, but we could recast the problem as an estimation problem and estimate the dose instead.

            This actually is just a pictorial slide showing the same thing.

            So if we want to actually do individualization, the questions become more and it's more problematic of how to do it best.  I'm only going to focus on this first question for the case when we want to do feedback individualization.  What dose strength should be made available?

            We came into such a problem when collaborating with a company that had developed a drug.  It was planned to go into phase III as a fixed dose size, everybody getting the same dose, but in light of high variability in PK and PD, partly because of polymorphic 2D6, they were contemplating maybe doing individualization.  And the question was, what would we gain by that?  Would it be sufficient?  And they wanted the gain to be measured on a responder scale and the overall responder rate.

            We went ahead and did the estimations based on one-dose-fits-all or a feedback individualization with two dose strategies where we estimated the lower dose size, the higher dose size, and the fraction of the patients that would be preferentially treated with the lower and the higher dose.  I won't go into technical details here, but we used the $MIXTURE function in NONMEM.

            DR. SHEINER:  (Inaudible.)

            DR. KARLSSON:  No.  This is feedback individualization without ‑‑

            DR. SHEINER:  This was feedback on the response?

            DR. KARLSSON:  Yes.

            DR. SHEINER:  So actually somebody was observed to be a responder or not a responder, and then the dose was changed.

            DR. KARLSSON:  Yes.

            We had built population PK and PD models for both the satisfactory effect and for the side effect previously.  These were more elaborate models with continuous and ordered categorical type data, but they could be easily reduced to the dichotomous question of were they responders or not responders.

            And when we estimated the single dose size to be given, it was very close to what the project team actually had come up with, and it resulted in 47 percent overall responders. 

            The best two-dose strategy was two doses, one lower and one higher, a 4 and 18, with about 60 percent gravitating towards the higher dose.  This was predicted to increase the overall responder rate to 63 percent.  This maybe was one piece of the puzzle that made the company actually in the phase III to go with both fixed dose and individualization in parallel.

            We did simulation studies on similar problems, and just to relate two observations there, one was that although a particular dosing strategy may not be the most optimal for one utility criteria, it may be near optimal across all relevant utility criteria, and therefore may be superior to other dosing strategies.

            Also, another observation that all-or-none type responder definitions, like in this example, seems to favor individualization to a higher degree than gradual utility functions.  Although this was obtained from a single example, it seems reasonable that if you have these very harsh, steep benefits of changing maybe somebody just a little bit on a continuous scale into having a utility of 0 to have one of 1, that that is more sensitive to individualization.

            Moving on from feedback individualization to individualization based on covariates identifiable up front when the patient is to be started on a therapy, we will face questions such like what is the best covariate to base dosing on?  What should the number of dose sizes be, and what covariate intervals should each dose size be applied to?  Just illustrating here, the covariate here might be organ function, body size, age, et cetera.

            If we want to go with two dose groups, the parameters we need to identify are what is the optimal cutoff value, what is the dose in the higher group and the dose in the lower group?  So that's three parameters to actually estimate.

            If we wanted to have three dose groups, that's five parameters, et cetera.  The problem becomes more complex.

            We would proceed to estimate those parameters in a very similar fashion as before, but with some differences.  First of all, in this case it's, of course, very important to have covariate models for the covariates that we are intending to be using for dosing decisions, and we also need to have distributions of these covariates in the target patient population.  We can obtain those from simulations, but more relevant maybe from empirical databases from previous studies.  What we're estimating are then the dose sizes and the cutoff values for them, where to change dose size.

            We also had an example of this in relation to drug development.  I can actually name this.  This is NXY-059, a drug under development by AstraZeneca.  We had a publication coming out in Clinical Pharmacology and Therapeutics in January this year.

            This is a drug to be used for stroke.  It's acute dosing, a 72-hour infusion with a 1-hour loading infusion.

            The project team was worried about too high variability if one were to give everybody the same dose in particular since this is entirely ‑‑ not entirely, but mainly renally cleared compound, and they were worried about the end of loading infusion and the maintenance infusion concentrations.

            So we tried to see what individualization could do in terms of bringing down variability to a reasonable level.  The target that was set was a free concentration of 100 micromolar.  The penalty function used was quadratic loss in log domain which means that half the target concentration is as bad as twice the target concentration.

            We had a pop PK model developed from the first patient study which showed clearance being highly dependent on creatinine clearance and volume on body weight.

            We used empirical covariate distributions from previous phase III studies of stroke patients.

            And for the loading infusion, we considered one dose, the same to all, or two-dose groups either based on creatinine clearance or on weight.

            For the maintenance infusion, it was clear that we could not give everybody the same dose, but two to four dose groups were explored and dosing were to be based on creatinine clearance.

            An additional constraint was made, which said that the therapy has to be fulfilling the following criteria, namely that 90 percent of the patients have to be above 70 micromolars and less than 5 percent above 150 micromolars.

            As it turned out, actually these criteria could be met by just giving one loading dose of 2,400 units to everyone, but for the maintenance infusion, it was necessary to give three different infusion levels, depending on the creatinine clearance, with the cutoffs at 80 and 50 mls per minute.  And you can see the dose units there.

            This dosing design was implemented in a phase II study, and the target fulfillment was acceptable, with more than 92 percent above the lower limit but more than 7 percent above the upper, which is slightly more than what was desired.

            So actually before that, we had done some simulations to look at a priori dosing based on covariates, and we took dosing on creatinine clearance as an example.  The standard approach often used when individualizing doses based on renal function is to use predetermined cutoff values for the renal function and quite large dose decrements, often a factor of 2 or higher, when going down to lower renal functions.

            We wanted to explore what would be optimal approaches to renal based dosing and we wanted to see what drug characteristics and what other factors influenced what would be the optimal approach.  So we did simulations where we changed the drug characteristics of hypothetical drugs, where we changed the creatinine clearance distribution in the target population, and also the utility function shape.

            And two factors came out as by far the most important.  For the selection of what would be the optimal cutoff value in the patient population, the median creatinine clearance in the patient population was most important.  And if only two dose groups are to be used, the cutoff should be ideally positioned close to that median value regardless of other drug characteristics.

            For the dose ratio, the ratio of the high to the low dose, the one factor that again by far was the most important was the strength of the covariate relationship, and for renally cleared compounds that can be expressed as the fraction excreted unchanged.  So when the fraction excreted unchanged is 1, the higher to the lower dose ratio would be around 1.7.  And for all other situations, the ratio would be lower than that ideally.

            These two pictures are quite in contrast to what the practice is today which is using cutoff values below the median value and often dose ratios higher than 1.7.  This might have practical and other reasons, but it's also to be maybe recognized that this has an impact on the target fulfillment.

            This is just a picture showing some of the gains that could be made from doing individualization based on fraction excreted unchanged and the number of dose levels.

            This is actually a picture very similar to the one that got us involved in this area.  It was when again we collaborated with a drug company who had already decided upon a dose individualization scheme where they had actually selected a low cutoff value and a large dose decrement.  So what in effect they were doing was they changed around the doses, but they didn't manage to lower the variability in exposure.  It was different types of patients that were in the tails of the distribution, but the overall variability in exposure was not reduced by the individualization.  So doing this might actually not be the most simple task.

            So to summarize with respect to target definition or utility function definition, this is something that certainly can aid data collection.  If one knows what parameters are the most important and therefore go into the utility function, that will aid both data collection and modeling efforts, being able to maybe take both quantitatively and qualitatively better data for those variables and also do modeling more focused on those.

            To improve communication within project teams or maybe between project teams and those outside.  My experience is that many of the important factors for the utility is something that resides with only a few persons within the project team and many of the others that are to contribute to the dosing decisions are not particularly well informed about the weighted balance between effects and side effects or between different types of effects.

            And of course, it's important to appropriately value the drug compared to other drugs.

            And the last point.  This was a slide prepared for a meeting in Europe last December that had the title "Getting the Dose Right."  If you ever want to say that you got the dose right, obviously you need to know what you're aiming for.

            Separate from defining the utility function, which in its own right has a lot of benefits, dosing strategy estimation might have additional benefits to motivate the choice of dose or dosing strategies and to obtain conditions for optimal individualization and thereby assess the maximal potential value of individualization to justify doing it or, oppositely, maybe to justify not doing it before because the benefit of doing it isn't large enough.  And if you do know the optimal individualization strategy, then you can directly offset any practical consideration that simplifies dosing against a decrease in target fulfillment.

            I know that some people don't believe me when I say it's easy, but compared to doing population PK/PD modeling, compared to defining utility functions, which are the two more difficult tasks, I think this is very easy.

            Thank you.

            DR. VENITZ:  Thank you very much, Dr. Karlsson.

            Any questions or comments by the committee?  Let me ask you, in your simulations, you used a symmetric loss function.  Right?

            DR. KARLSSON:  We used both symmetric and non-symmetric loss functions.  Obviously, if you use non- symmetric so you penalize the high concentrations or adverse events more, then you're going to gravitate towards lower doses.  If you are non-symmetric in the other direction, you're going to penalize the low doses, so you're going to get a higher overall dose.

            DR. VENITZ:  So your conclusions with regard to the effect of mean creatinine clearance, that was based on a symmetric loss function.  Right?

            DR. KARLSSON:  We actually did explore various loss functions also there, and across a range of reasonable loss functions that we thought are reasonable, it was pretty stable towards that.  But obviously if you have a very asymmetric loss function, you will tend to get other values.

            DR. VENITZ:  And my guess would be that's the reason why your recommendation is different than what's currently done because what people build in is a loss function where they're very worried about overdosing, less worried about under-dosing.  So the easiest way to account for that is by adjusting the dose in people that have renal failure.

            DR. KARLSSON:  Yes.  I think that's true, that they are usually dosed to an average AUC that's actually below what's seen in the main patient population.  So if that is what is decided, then certainly that's the case.  For the example we had here, where actually we're operating at one point with different loss functions for people with lower renal functions because if there was a concern about renal function, then we wanted them to have a lower target as well.  So there is certainly the possibility of incorporating all these types of considerations.

            I think the reason why we see more lower cutoffs and large dose decrements is probably the way drug development pursues with more healthy patients to start with and then inclusions of larger and larger.  So the initial dose levels are based on those with higher renal function and then the other ones are added as a tail more towards the end.

            DR. SHEINER:  I just wanted to sort of raise a point in the questions.  You were romping through your slides.  So you looked at individualization based on that drug that you told us the name of, that slide you just showed a moment ago where it had the 90 percent and the 5 percent?

            DR. KARLSSON:  This one?

            DR. SHEINER:  No.  At the bottom of the slide, it had selection of dosing strategy.

            DR. KARLSSON:  Okay, yes.

            DR. SHEINER:  Yes, there.

            Correct me if I'm wrong.  I think one important thing to realize is that those numbers down at the bottom, of course, might not be attainable by any strategy.  But to discover whether they are or are not attainable by some strategy, you fix on a model for the process.  Then those percentages there have got to do with variability among patients.  So this doesn't contain that uncertainty that I was talking about.

            The model that says how frequently you'll get toxic at a given level, that very model is itself uncertain because it's based on assumptions and it's based on data.  It would be very likely with not a great deal of data ‑‑ it would be impossible to attain a 90 percent probability of being in the right range or whatever it is if you incorporate that kind of uncertainty because that uncertainty says, I don't know how the world works.  So how can you possibly get 90 percent certainty on anything?

            And there's nothing wrong with this.  This is the right way to proceed.  Then you want to look ‑‑ presumably you did ‑‑ at the sensitivity of those two various assumptions that went into your model.

            But this whole business of being clear with people about what uncertainty you're bringing in when, when it's appropriate and when it's inappropriate ‑‑ it would be here inappropriate to bring in model uncertainty when you're trying to find the strategy because it's got to condition on some state of nature.

            So I just wondered what your experience was with dealing with all of those, I think, somewhat subtle issues with people who are not used to speaking this language.

            DR. KARLSSON:  It is true that we actually didn't ‑‑ at first, when we saw these, we thought they were too stringent criteria.  With normal variability, this would be very hard to achieve, and as you say, it might not actually be possible.

            When doing these calculations, we did not take the structural model uncertainty into consideration, but we did do simulations looking at the uncertainty in the population PK parameters, what the impact of that would be.  And it wasn't actually as large because they were relatively well defined.  But this is, as you might guess, done with the point estimates.

            In general, I do find it often difficult to discuss these matters with the project team I think maybe because it's not usually done, talking about quantitative models in this sense.  The type of utility function that does seem to be used is the responder/nonresponder criteria.  So it's easier if it's simpler, but again, the discussions around the responder criteria was where to put the cutoffs and people differed in their opinion of where to put the cutoff, which is the same thing.

            DR. LESKO:  Mats, I want to bring you back to a little more maybe pragmatic question.  But clearly this can be done in the context of what we talked about this morning in having a data set in front of us and then looking at an approach like this to adjust ‑‑ or make a decision based on adjust dose.

            However, as I listened to you speak about the project team where there's a need to define a target and the penalty function, this method requires a weighted balance between the effectiveness and the safety.  I guess in some cases that's very clear, depending on the nature of what the effectiveness and safety is.

            If you think of risk as sort of an overriding issue in drug development as opposed to efficacy ‑‑ in other words, the risk of an adverse reaction, limiting approval, limiting dosing, limiting the label ‑‑ does the notion that a safety consideration would drive the relative agreement that you would have in a project team on the utility function ‑‑ in other words, I'd give more weight to a safety side of the drug's effect as opposed to the efficacy side.  As an example, if I had a QTc issue, that would seem to weigh heavily in terms of my utility function consideration even against the most promising efficacy that I might be speaking about.  So it would be the driver, if you will, in trying to get an agreement on what weights to assign to the utility functions.

            Can you speak to that a little bit?  How does that work?  And we sort of talked about this last time.  You said ask a clinician, but we have actually asked clinicians after our October meeting and we do get quite different views of how a utility function would serve the purpose of what we're talking about.  So I wondered if you can sort of pursue that a little bit.

            DR. KARLSSON:  I wouldn't have any hopes or belief that utility functions within the very near future would drive the decisions so that you would just define a utility function and then forget what went in there.  I think a more reasonable way is to actually come up with some initial utility function and then use that in parallel maybe to illustrate different consequences of decisions.  Then that would maybe show up where utility functions would actually fail because it wouldn't take something into consideration that is of importance, which would point to how they need to be refined, what needs to be considered in them.

            When it comes to the safety, I think in some situations where there are tolerability issues of maybe not so severe nature as QTc prolongation, it's easier to incorporate them.  Obviously, severe side effects are ‑‑ by its nature, you're not going to see very many of them, and you're going to not want to expose patients or volunteers up to the range where you get a very good handle on what the function is at the dose levels.  So you're going to have a large uncertainty in the models at that range, and you need to take that into account I think.  So you're going to have a much larger uncertainty on the upper end than on the lower end probably.

            DR. SHEINER:  I know you know my response to this, Larry, but I just wanted to say it keeps on reminding me of the old data analysis argument about the Bayesians and frequentists and the issue of I'm trying to fit a model that's too big for my data, and so I'm going to fix some parameters.  The Bayesian, of course, hearing that, says well, okay, you may not think you'll be acting like a Bayesian but fixing parameters is the same as saying they have a prior distribution that's got point mass at the value that you said.  That can't ever be as good as giving it a little bit of wiggle room.

            Well, it's the same kind of thing here with utility.  Of course, if you have trouble convincing people that they ought to sit down and do a utility function, then there's some implicit utility function that's dominating usually by the most aggressive person or the person highest up in the organization who happens to be at the table.  And you don't know what it is.  If you believe in decision theory, it's got to be there somewhere, and it's what's overriding everything else.

            Just listening to the talk and saying, wouldn't I like to be in a room where the discussion was about how much do I believe that these data really do support this notion of what's going to happen, how much weight do I put on versus how much you put on, the various side effects.  It seems like such a rational and sensible discussion to have rather than, well, I think we ought to go with 25 milligrams.  What do you think, Joe?

            DR. VENITZ:  Any further questions for Dr. Karlsson?

            (No response.)

            DR. VENITZ:  Thank you again for your presentation.

            Peter, do you want to pose the questions for the committee?

            DR. LEE:  I think the question is very simple.  Can this methodology be generalized to other scenarios?  That's pretty much the question that we have.

            DR. VENITZ:  The question for the committee is, can this utility approach be generalized to other therapeutic areas?

            DR. SHEINER:  Well, why isn't it just about as general as it gets?  This is one of those things that really is ‑‑ you know, it applies to drugs and automobiles and airplanes.  How can you make any decision if you don't know what you're deciding about and what your values are and what the likely state of nature is?  That's all we're really talking about here.  Then after that, it's computation.

            DR. VENITZ:  I obviously second that, but I'd also like to maybe point out an approach of how we can convince other stakeholders that this is something useful.  And it goes towards identifying utility values generically for certain kinds of adverse events, just like you would assign generically speaking for certain efficacy, life-threatening disease versus quality of life changes, and get agreement on that regardless of what drugs you're looking at rather than what you presented are specific to that drug where a decision was made this is how I define my utility function relative to my target, which makes it vary case by case.

            But I think if you want to get consensus, let's start on the safety side.  We agree certain adverse events are going to get certain utility values associated with it regardless of what causes it.  By the same token, on the efficacy side, certain disease interventions get certain utility values assigned regardless of whether it's a drug treatment, device treatment, or whatever it is.  Then at least there's some common acceptance I guess on what the ranges are.

            Otherwise, you're really getting into this swamp of, well, everything is a case to case.  Show me the drug and that's the only way I can come up with a utility function, which defeats the purpose in my mind at least of using utility functions.

            DR. LESKO:  It just strikes me what you said would lead to agreement particularly on the safety side.  I think you can identify certain safety signals that people would agree are bad and are most serious, and then weigh those, in turn, against a range of efficacy or benefits that one might get from a drug and against the loss of efficacy if you were to somehow misappropriately adjust the dose.

            I think you can define boundary conditions in a sense with regard to certain safety signals, with regard to certain categories of drugs for efficacy.  For example, hepatic toxicity, QTc.  I think you'd get general agreement these are bad.  We focus on those extensively irrespective of what the efficacy side is offering.  By the same token, a disease state where efficacy is of extreme importance, there may be a different view of the safety signals.  But you need some anchor points, it would seem, and they might be those boundary conditions, and then the gray areas fall in between somehow.

            DR. VENITZ:  But if you look on the adverse event side, in oncology there is already a consensus of how to categorize and how to rank order certain adverse events according to organ function.  It's unanimously accepted by the whole oncology community.  Now, they don't necessarily use it in the context of utility.  They use it to define dose limit of toxicity.

            But why can't we do that as a general approach to adverse events?  And you would do that regardless of the underlying cause of that.  Whether it's a specific drug or other drugs, it's irrelevant.  To get out of this discussion where everything is case by case, the moment you do that I don't think the utility approach is going to work.  It might work on specific drugs, but it wouldn't work across the board because you've got nothing to compare to.

            DR. KARLSSON:  Although I agree and I like the idea.  Maybe one complicating factor is those types of grading 1 to 4, which I assume you're talking about, are based on outcomes, aren't they?  Whereas, if you have like a biomarker for a safety event, that's something a bit different and maybe more difficult to value.

            DR. VENITZ:  Maybe.

            DR. FLOCKHART:  I'm just thinking about something I know more about which is the QT interval.  So the question would be, is a given QT prolongation, a prolongation of 10 milliseconds, by one drug the same as a QT prolongation of 10 milliseconds by another drug?  Unfortunately, the answer is no because drugs do more than one thing.  They have more than one mechanism of effect on the QT interval.  For example, many antipsychotic drugs that affect the QT interval are anticholinergic as well, so they have effects on the heart rate, and that might be a little bit protective, make them a little bit tachycardiac.  In situations where they got more drug, they would not only get more IKr blockade, but they would get more mu 1, M1, receptor blockage.

            So because of that, I think hepatotoxicity gets even more complicated.  If one drug causes X amount of change in the LFT's, is that the same as another drug causing the same?  It's a very complicated thing.

            Nevertheless, I think the effort is probably noble.  It's worth venturing along that path at least to find out how different things are.  I'm intuiting that QT interval drugs would be different.  I don't really know.

            DR. VENITZ:  But I think the current ‑‑

            DR. FLOCKHART:  You need an outcome, though. You need an outcome, and that's a very important point.

            DR. VENITZ:  As far as QTc is concerned, the current assumption is that QTc prolongation is bad regardless of what causes it.  The relationship then really is what you're talking about, how predictive is a biomarker of a clinical outcome?  Because the biomarker itself is not bad.  It's the clinical outcome.  You're worried about the fatal arrhythmia.  So in addition to your utility, now you have to look at what are your uncertainties involved in linking that biomarker that you're measuring to the final outcome.  In other words, for a given change in, I don't know, 40 milliseconds in QTc, how many people are likely to develop fatal arrhythmias.

            DR. FLOCKHART:  Yes.  You can put a model on it, and you can also model in things like there are a significant number of drugs that prolong the QT interval for which torsade has never been reported, and we have them on the website.  You could model that in as well.  That's a chance that a drug in that class would not do it.

            DR. VENITZ:  But from my perspective, that's not a utility.  That has something to do with how your biomarker relates to the clinical outcome.

            DR. FLOCKHART:  Right. 

            DR. VENITZ:  Because really what you're assigning utility to is the bad outcome and fatal arrhythmia in ‑‑

            DR. FLOCKHART:  Well, it is affecting the utility, though, because it alters that number. 

            DR. SHEINER:  (Inaudible.)

            DR. FLOCKHART:  Dr. Sheiner was pointing out that it affects the expected utility, and he's right.

            DR. VENITZ:  Any other comments or more specific questions of either Dr. Karlsson or the FDA?

            DR. KARLSSON:  I was just going to add to that that maybe if you do start with a utility function that incorporates many different aspects of the drug therapy, you will find that actually in the end it's only really a few of them that are going to be important and you could reinspect those and the assumptions that go along with those particular events.

            DR. SHEINER:  Let me just ask.  Some of these examples were actual examples from your interaction with the industry.  I guess we can't draw too much of a conclusion from anecdotal experience, and also these folks asked as opposed to having it pushed at them by a regulatory agency.

            Is it your feeling that this exercise was appreciated, informed people, and had some consequence in terms of the way in which the development plan went forward?

            And finally, do you have any information on whether or not the regulatory agencies with which these companies dealt ‑‑ how they responded to the justifications offered through this means?

            DR. KARLSSON:  I don't think any of these drugs have gone to the regulatory authorities yet, although I might be wrong.

            For the example that I presented, the AstraZeneca one, they were very helpful and really appreciated all our efforts.

            In the other cases, I think it was more add-on and maybe it was not the core of the project team that was really wanting to have this information.  It was more a side effect of having the possibility of doing it, and they thought maybe it was nice to know, but to what extent it influenced their decision I'm not sure.

            DR. DERENDORF:  A follow-up question.  How many times do you think this is done a posteriori so that the decision is made first and then you need a justification for it?

            DR. KARLSSON:  I don't think that's done very often.  I don't know, but in my experience outside the examples I've been involved in when I've actually done this, I haven't heard these kind of discussions going on with people in industry.

            DR. LESKO:  Just following up on the same question.  Maybe this isn't a fair question, but I sense, at least in our regulatory agency, a stronger desire to understand dosing strategies than there has been in the past.  It's partly related to a lot of things. 

            The question I sort of have is a follow-on to one that was asked, and that is, is there anything in your experience in working with the method, within the context of drug development, that would cause any concern on the part of a sponsor with regard to what a regulatory agency might ask of this kind of methodology?

            It would seem like it would be received rather positively because it provides a fair amount more than we normally would see in terms of a dose justification or a rationale for dose selection.  Therefore, my sense would be it would received positively.

            But on the other side of that coin is anything different is different, and what are the issues that somebody might think about with regard to if I presented this, the regulatory agency might do X, Y, or Z?

            DR. KARLSSON:  Well, I've heard both arguments, both that it's good to do your homework and to be able to know your drug so you can argue well.  I've also had the response that let's not do it because we're waking the sleeping bear, or whatever your expression is, that maybe they wouldn't think about this if we haven't done it.

            DR. DERENDORF:  Well, but for internal decision making, I think it's always helpful to do it.

            DR. KARLSSON:  I think to some extent, though, it goes hand in hand.  If you don't want others to know it, you don't want to know it yourself.


            DR. VENITZ:  Any final comments?

            (No response.)

            DR. VENITZ:  Thank you again for your presentation.

            Then we are going to start with our second topic and the last topic for today which is the pediatric database.  I think Dr. Lesko is going to give us an introduction to that topic.

            DR. LESKO:  Thank you, Jurgen.

            This is going to be a rather brief introduction.  It's to lay the groundwork really for the next two presentations.  The focus of this now, switching gears, is pediatrics, and we're going to be talking about two topics.  The first is going to deal with a template or standardized approach, if you will, for pediatric studies that utilize sparse samples.  The second is to discuss with the committee some ideas that we have for mining the pediatric database that we have as a result of the pediatric rule.

            Let me start with that.  The pediatric rule is something many of you are familiar with who were actually here at our last meeting because we discussed it in some detail.  We shared with you the types of studies that we've received under the pediatric rule.  Basically the rule encompasses the idea that we want to bridge the adult data to the pediatric situation, usually not directly but through the initiation of some studies in the pediatric population, and depending on the nature of the adult data and the nature of the assumptions that we make in doing that, the types of studies that are conducted for pediatrics would be either efficacy, safety, or pharmacokinetics.  We presented last time a pediatric decision tree that sort of guided the thought process on what studies to conduct.

            Basically the idea with the pediatric rule is to avoid large-scale studies.  It's considered inefficient to redesign the efficacy and safety trials when there's a preexisting database in adults.  Although the challenge from a clinical pharmacology perspective and a clinical perspective is to decide what studies are, in fact, appropriate to conduct and will provide the most information.

            The pediatric rule, as you know, is intended to speed up access of drugs for children, to do that in a cost effective way without reinventing the entire efficacy-safety spectrum.  I think overall I think most people feel this has been generally successful in meeting the goals that were intended for it.

            But some questions always need assessment within the context of the pediatric decision tree and the pediatric rule.  Is it reasonable to assume a similar PK/PD relationship as exists in adults?  This sounds very familiar to the questions that we were discussing this morning in terms of different populations, and probably the underlying principles are very much the same.

            We feel there's a need ‑‑ and we've begun to look at this ‑‑ a need to develop standard methods for answering this question for specific drugs and drug classes particularly where we've seen submissions in pediatric patients in drug classes over a number of drugs.

            The other question is what is the appropriate dose.  And to answer this question, we generally rely on pharmacokinetic studies.  They can be of two types:  the full exposure, sample-rich type study design, or the sparse-sample, population PK approach.

            The overall goal of these studies is straightforward, to achieve a dosing that's intended to achieve exposure similar to adults.  Generally PK studies aren't the only studies conducted in pediatrics.  Frequently a safety study or in some cases a small efficacy study is also required.

            But it's probably safe to say that a sparse-sample strategy has been under-utilized in pediatric studies.  We have some ideas why that might be.  It might be that there's an insufficient understanding of this approach.  It might be that there's some concern that the regulatory agencies won't accept such an approach.  But it seems like there's room for opportunity here to do these types of studies that call for pharmacokinetic information in a sparse-sample strategy that is based on good principles.

            So the first thing we wanted to talk about today was a discussion of a standardized ‑‑ and I use "standardized" very generally, but a standardized study design template that would be useful, for example, in communication between investigators or between companies and the agency to agree on a sort of starting point for designing both acceptable and informative PPK type studies.  So that would be the first thing we're going to talk about, and Dr. Peter Lee will talk about that primarily.

            The next topic is a follow-on to what we talked about in October.  We had made the point and we shared with you the specific drugs in October for which we have a database on pediatric studies.  We raised the question about if this were your database, what would you want to learn from the database and what would you look for.  What would be the questions you would have?  We didn't spend a lot of time with that.  We have more time today.  But we've begun to look at this database and you'll hear from another individual from our office, Dr. Gene Williams, who is on detail to the immediate office of OCPB to specifically look at what we can learn from this database.

            This is a work in progress.  We began to assemble the database, in an organizational way of age groups, the PK data that we have.  We've begun to look at individual drugs in the database, elimination pathways, the clinical endpoints that were studied.

            I'd have to say this has not been easy.  Unfortunately, this database is not in a form where you can just press a few buttons and pull it out.  So there's a fair amount of up-front work that goes into assembling the database.  I think in recognition of that, it's important that we understand where we want to go with the questions that will derive from this effort to organize the database once we start moving in that direction.

            So Gene is going to do an overview of some of the research objectives that we have, with the goal of generating knowledge from this database.  What we think we'd like to do is look at underlying mechanisms where there are exposure differences between kids and adults, look at breakpoints perhaps on age, look at specific elimination pathways.

            The reason we want to do all this is it's sort of common sense.  We don't feel we should be asking the same questions now of pediatric studies that we asked three years ago before we had 50 or 60 studies in house.  So the idea is to learn from the database and then use that information and knowledge to revise our pediatric context or pediatric decision tree for the future.

            So the goal of this effort:  to improve or revise our pediatric decision tree based on identifying better studies that we need to conduct in the future ‑‑ and by "better," I mean more informative studies ‑‑ or perhaps reduce the number of studies in the areas that the data would allow.  So for that, Gene Williams will present an update. 

            And Peter is first.

            DR. VENITZ:  Lew.

            DR. SHEINER:  Can I just ask you a question, Larry?  On your third slide where you have "some questions always need assessment," and you say, "is it reasonable to assume a similar PK/PD relationship as adults," I understand that that's not particularly the focus of today.  I'm sensitized to this because of some work recently that I've been doing with a topic that's related to this with Novartis. 

            It's an interesting question.  What would constitute evidence that children are like adults with respect to PD now, not the PK part?  I just wanted to put it on the list of many things that you have to think about, that maybe we ought to address that at some later point, or maybe you might want to have us address that at some later point.  It's a vital question.  Obviously, if what it takes to establish the similarity relationship is more work than it takes to clear a drug to start with, you're done.

            DR. LESKO:  Yes.  I think it's a vital question.  We have a few instances, a few drugs in hand where we actually have information both on the adult where that happened to be done in the NDA and it was also done in the kids.  We're going to focus on a few of those examples to try to answer that question.

            We've often turned the question around and asked how much of a difference would be important in the PK/PD relationship and would that warrant necessarily a different dosing strategy in kids.  These are open questions, but we do not have a lot of opportunity to look at this issue based on, at least, what I've seen of the database to date.  Maybe Gene will comment a little more on that.

            DR. SHEINER:  Right.  It does relate to that.  But it's also a question of style in the following sense.  If the data are sort of anecdotal, are they really data?  So the issue becomes ‑‑ when you say I think that the relationship is the same, and then I say to you, okay, here's some evidence, and you say, well, that's not really evidence because that's physicians' opinions, let's say.  So the question is what constitutes evidence that it's the same, given that you're inclined to believe it.  That's the issue I was getting at.

            DR. LESKO:  Right.  That's a good question whether it's in the statistical domain or whether it's in the modeling domain.  I think we need to talk about that, but I think that's a good open question for another discussion.

            The point of bringing that to the committee might be when we've done our analysis of the database as we have it and share with the committee what we've learned from the data that we have and maybe what we might want to know from data we don't have at the moment but might recommend somebody begin to look at, not necessarily in addition but maybe as an alternative to the studies that are being conducted today.

            DR. VENITZ:  Peter?

            DR. LEE:  What I'd like to present to the committee today is a proposal to develop a pediatric population PK study design template.  I believe we have sent a copy of the full proposal to each committee member a few weeks ago, so hopefully you would have had a chance to look at the proposal already.  But what I'd like to do in the next few minutes is to just go over this on the key points in the proposal.

            The objective of the pediatric PK study design template is the following.  First, to provide a consistent approach, like Larry mentioned earlier, to design and evaluate a pediatric PK study.  We'd like to develop a computer-aided pediatric study design template which will take the user-supply study design and automatically estimate the study performance and study power.  So it will be making it easier for the user to determine which type of study design is appropriate for their drug and for their design.

            We also want to select case studies from the FDA database to test the template and refine the template.

            Finally, through this template, we hope to promote a wider use of population design in pediatric studies.

            I think Larry has mentioned this pediatric study decision tree.  He also presented this decision tree in the last meeting.  Basically we use this decision tree or propose to use this decision tree to determine what type of study will be necessary to bridge the adult efficacy and safety data to pediatrics.  Depending on the answer to many of these questions, we may end up doing or recommend doing a efficacy study or clinical study, a PK study or a PK/PD study, or all of the above, or safety studies.

            But today I just wanted to focus on this particular box here which is we will recommend doing a PK study as a bridging study between adults and pediatrics.

            So once that decision has been made, then we will have to use the PK study for dose selection in pediatric populations.  So this goes back to the dose adjustment in special populations, the same decision tree we talked about in the morning today.

            Based on this decision tree for dose selection, we have to answer several key questions.  First, we have to ask whether there is a clinically significant difference in pharmacokinetics between adults and pediatrics.  Secondly, we have to ask what is the pharmacokinetic parameter in pediatrics, and we had to use that pharmacokinetic parameter to adjust the dose in that population.

            So based on this decision tree, we had to conduct the pediatric population PK study to get two information.  First, we had to identify whether there's a clinically significant difference ‑‑ not just any difference, but clinically significant difference ‑‑ between adults and pediatrics.  Secondly, we had to accurately estimate the parameters in pediatric populations without any bias.

            There are, of course, many factors that may influence study performance, population PK study performance.  For example, the study design which may include a number of subjects, demographic information maybe, the number and timing of samples.  Also, the study conduct, such as the compliance of the patients, the variability of dosing time and sampling time and in missing dose and in drop-off.  Of course, the pharmacokinetics itself and the variability of the pharmacokinetic parameter also influence the study performance or how we design the studies.

            In the proposal, we also bring up several important points to be considered during designing a population PK study.  We believe that when we design a study, we have to take into consideration the objective of the study which I just mentioned to identify the clinically significant difference and to estimate the PK parameter in pediatrics.

            We also believe that because the pharmacokinetic parameters are very different between drugs, also there are many different varieties of population PK designs that will provide sufficient study performance to achieve the objective that we just mentioned.  There is no one-size-fits-all design for population PK.  Each design should be looked at on a case-by-case basis, but using a consistent approach which we believe a clinical trial simulation will be a good practice to estimate the study performance and study power.

            In our proposal, we also bring up several other points to consider.  For example, we had to look at a number of factors that may influence study performance, such as dosing time, sampling time.  Compliance is an important factor.  Of course, the number of subjects, the number of patients, and how the sampling time and dosing time is distributed with time.  We also need to consider if there's an unbalanced design in terms of number of subjects as well as the number of samples between different populations.

            This is a flow chart of the proposed study design template.  It consists of a module where the user can input their study design protocol.  So this is where the user can enter the number of subjects, the number of patients, when do you take the samples, and what is the variability of pharmacokinetics, and so on and so forth.

            So this will be your input parameter where the user can put into the module.  And the template will take this information and automatically generate or translate the study design template into a simulation code.  With that simulation code, the software will generate a study performance indicator.  In this case, it will be the power to determine whether there's a clinically significant difference in pharmacokinetics between populations, also what is the accuracy and precision and bias of the parameter estimations.

            So the input and output of the proposed study design template includes the following.  The input will be a study design, the pharmacokinetics in adults, but also the variability of the demographic or patient population you will include in the studies, and also the study design variables.  The other input parameter will be the criteria for evaluating the study performance. 

            But the output from the template will be the estimated study performance which is related to the two objectives of the population PK study.  One is the power to identify ‑‑ I want to emphasize ‑‑ the clinically significant difference, and secondly, the precision and bias of parameter estimations.

            The clinical trial simulation that we propose to be used in the study design template is pretty standard.  It includes the following steps.  First, it will generate demographic variables and pharmacokinetic parameters based on the user input information.  It will simulate a study design as well as the study conduct, and it will generate population PK data.  Once the population PK data is generated or simulated, then we will conduct a population analysis based on the simulated data.  And finally, we will repeat the process perhaps a few hundred times to a few thousand times to estimate the power of the study as well as the precision and accuracy of the parameter estimations.

            Like I mentioned earlier, there are two main objectives to measure the study performance.  The first one is to identify a clinically significant difference in pharmacokinetics.  Based on the decision tree or dose selection that we had presented in the morning and early in my slides, the first option is to just look at a 90 percent confidence interval and see whether the 90 percent confidence interval of a PK parameter is within a default boundary.

            But the second option which we might prefer is to determine whether a clinically significant difference in pharmacokinetics will exist.  In order to determine the clinically significant difference in pharmacokinetics, of course, we had to know the PK/PD relationship and we have to assume that the PK/PD relationship that we have perhaps from the adult population is similar to those in the pediatric population. 

            For example, in the simulation we can assume that there's X percentage which is considered a clinically significant difference in the body weight normalized clearance.  With that difference, we want to ask the question whether the population PK study that we tried to design will be able to capture this clinically significant difference.  So with that we can calculate the study power of the population PK studies.

            Now, the second criteria for study performance we propose is the precision and bias of PK parameter estimations.  Precision and bias can be presented in terms of a percentage prediction average.  Precision will be the standard deviation of the prediction errors and bias will be the average of the prediction error.  The prediction error is defined as the percentage difference between the true value and the predicted value because we're doing a clinical trial simulation, so we know exactly what the true value is.

            So the output again are two information to measure the performance of the study.  One is whether the study has a power to identify a clinically significant difference.  The second is precision and accuracy of the PK parameters.

            So this is basically a general description of the proposal.  Again, the detail is elaborated in the actual proposal itself. 

            I guess we'd like to ask two questions to the committee.  The first question is, are the proposed objectives for pediatric PPK studies reasonable, considering the decision tree for the dose adjustment?  So we have mentioned two objectives for the population PK studies.  We also talk about criteria for study design performance.

            The second question is, is the proposed pediatric PPK study design template reasonable?  This is related to the clinical trial simulation approach that we propose, as well as the factor to be considered, a study design factor.

            We also would like to ask what feature we should include in the pediatric study design template so that we can make it more user friendly and more useful for both the clinician and the reviewer who might use it to design pediatric population studies.

            That's it.

            DR. VENITZ:  Thank you, Peter.

            You have the questions for the committee.  Any comments by the committee?  Hartmut.

            DR. DERENDORF:  I have a question for clarification.  On your decision tree, right on top you said it's reasonable to assume similar response to intervention.  Then you have yes, and reasonable to assume similar concentration response in pediatrics and adults.  What's the difference between the two?

            DR. LEE:  I think one example was asthma.  For example, it may be a different endpoint that can be measured in a pediatric population than an adult population.  That will be the first block of questions.  So if we can answer that question yes, then we go to the next block.  But if the answer is no, then we have to go to the clinical studies.  So if the answer is no to the first two questions in the top block, then basically either the clinical endpoints are different between the two populations or the disease is totally different or disease progression is totally different in the two populations.

            Now, the second question on the right-hand side of the block is that once we decide that through our clinical opinion the disease and disease progression and endpoints are similar, then we ask in our database do we know there is a PK/PD relationship and can we assume that a PK/PD relationship is similar between pediatrics and adults.  For example, for a proton pump inhibitor, we know the PK/PD relationship of gastric acid versus drug concentration.

            Now, can we assume that the relationship we have seen in the adult population is the same as the relationship we will see in the pediatric?  If the answer is yes, then we can just rely on PK information to determine the dose selection in the pediatric population.

            DR. JUSKO:  A comment and a couple of questions.  This, in general, seems like an excellent idea.  Of course, we always want to utilize information we know ahead of time to anticipate the study design and changes that will occur in a new group for study. 

            One small point that seems to be totally missing is dosage form and bioavailability.  Young children are typically getting liquids and chewable tablets, and in order to anticipate what will happen in the kinetics and dynamics in young children, you may need a comparable database in adults with a similar dosage form or at least make an adjustment for perhaps faster dissolution or absorption.  So at some point some reminder of that question may need to be added.

            Then there's a certain vagueness when you talk about dose adjustment or dose selection because in pediatrics there's always dosage adjustments.  Children are seldom given the same 500 milligram tablet that adults are.  So something needs to be said about what do you mean by dose adjustment.  Are they getting certain milligrams per kilo already or certain dosage sizes depending on weight ranges?  There's a great deal of flexibility already inherent in selecting dosages in children.

            DR. LEE:  I think the answer to the first question is I think we are talking about dose selection because we know that we're going to adjust the dose anyway. We're going to give a different dose.  Normally the dose is given on a per body weight basis or sometimes we will look at the body surface area.

            DR. LESKO:  It depends on the age group.  Sometimes the dose is adjusted based on average exposure to the drug, and depending on the age, it may be down to a milligram per kilogram basis, something like that.

            But I was thinking of the first question on the response that Hartmut asked and I don't know if Peter clarified.  But I was thinking basically you're asking the question is the response the same that you'd be interested in in kids as you would be in adults.  That's sort of like a two-part question.

            So the first question is, is the response, let's say, FEV-1 in an asthmatic patient the appropriate response?  And that sort of gets to the heart of the mechanism of action of the drug and the progression of the disease similarity.  Often data to support those assumptions or answers is not available, but if you do agree that, yes, I believe that's true, then the question is, is there a concentration related to that response that you previously agreed is similar to the adult?  Then that puts you further down in the decision tree.

            DR. CAPPARELLI:  I just had a question and a comment as well.  In regards to the scope of this, is this looked at as part of a safety study as well, or is this really structuring a population PK study with the only endpoints being population PK?  Because I think one of things that's missing in the objectives is getting at the question that we've been asking how do we assess these potential age-related exposure-response relationships.  And the design that may be very robust for estimating the PK parameters of the population may not give you all of the estimations in the individuals that you may want to do some of that exploratory analysis.  So I think that it needs to be very clear along those lines.

            It was brought up before, and I spoke with Dr. Lee before a little bit about the concept that, really, kids are different.  The question is whether or not we can predict those differences a priori.  So saying are there differences, there are always differences.  The question really is that based on our knowledge of modeling, our knowledge of pathophysiologic changes, developmental changes, can we predict those well and then go from there.

            I also just wanted to amplify Dr. Jusko's comments about the dosage formulations.  A lot of the exposure issues are going to be based on what size doses you have.  So when you say clinically significant differences in PK, it's really what sort of exposures we're  going to get out of the available dosage forms.  There may be a modest change in PK, but because of where you're left with your cut-points, you may end up having big changes in dosage exposures which again it needs to get back to, I think, what we're interested in and at least getting the exposures as comparable when we don't have the information in terms of differences in exposure-response relationships.

            DR. RELLING:  Is it implicit that it's only worthwhile to do these pharmacokinetic studies in children if there aren't good a posteriori methods of dose adjustment based on more readily available clinical measures like blood pressure, like immediate response to anesthesia, like pain relief?  Is it any part of your interest in the pediatric rule that pharmacokinetic studies be performed for drugs for which there's a narrow therapeutic range or small therapeutic index and there's no other good way to adjust doses?

            DR. LESKO:  No.  I think the pharmacokinetic studies are routine in these types of situations.  I don't think dose adjustments ‑‑ I'm trying to think if I have any experience with dose adjustments being made without the availability of PK information, for example, being based on observed responses in kids relative to adults.  Is that sort of the question?

            DR. RELLING:  My point is that I feel that there are a lot of studies being done by pharmaceutical companies, which they claim are being done to fulfill FDA regulations or requirements or suggestions, which are done for medicines that don't need to have pharmacokinetic studies done.  They're done for anesthetics that could be easily titrated based on the response of the patient to the oxygen saturation.  They're done for narcotics for which the drug is going to be adjusted based on pain response.  And I feel like a lot of resources are being expended on these studies for unclear reasons.  So I'm trying to figure out is this a suggestion that's being made for all medications that would ever be used in children regardless of the therapeutic index and the ability to adjust doses based on other parameters besides PK.

            DR. LESKO:  Yes.  I'm not sure "all" is the right word, but I'm thinking of the implications in labeling the drug product with a starting dose in pediatrics.  You need to have some information, it would seem to me, to begin dosing, and that generally has been the pharmacokinetic studies to recommend some changes in that initial dosing strategy.

            This isn't to say this is the only objective measure that one looks at in the pediatric area.  There are always other studies, in particular safety studies and, in many cases, small efficacy studies as well.

            DR. KEARNS:  I want to thank Dr. Relling for her insightful question because everybody wants to know the same thing about children and that is how much do I give, do I need to give a different amount as the child gets older, and will it work like I want it to.  That's really crux of all this.  I won't belabor all my soapbox points about this issue, which are many, but I want to make two points about this recommended approach.

            First is the issue that Dr. Sheiner said we should grapple with at a later date, but it's at the top of the decision tree, and that is, do we believe that whatever condition occurs in a child it is substantially similar?  And that's the language in the law, "substantially similar" to what it is in an adult. 

            For the last four or five years, I have seen people at all levels of the agency grapple with this as though it were a very large, mean animal, and at the end of the day, people rather than slay it, seem to run from it and invent ways to try to avoid it.  I think that's tragic because what that has produced is an incredible consumption of resources, not to mention the exposure of children to clinical trials that is a needless exposure.  That doesn't get talked about enough, and I'll stop talking about it now because that's probably another whole day.

            Now, with respect to "substantially similar," what comes to my mind is something very similar.  I'm starting to sound like one of our politicians who uses the same word over and over.  There's an article in CP&T, Art Atkinson.  It's near the front of the journal this last month on biomarkers.  It talks about the goodness or badness of biomarkers on how far away they fall from the trunk of the tree of drug effect.  The same thing could be looked at with respect to this issue in children.  Drug action is obviously something we can, at times, determine whether it's similar.  If we can't talk about action, we can talk about drug effect, a physiologic response in an association with a drug dose or a concentration or an amount.  If we can't talk about that, we can then talk about disease response and lastly disease progression.

            Sadly, I hear people put disease progression at the top of the discussion list as they look at that box because if we were to get a bunch of pediatricians in a room and ask them if they agreed that the progression of GERD was the same in adults and children, they would never agree.  They would never agree about asthma.  They would never agree about leukemias, other malignancies.  Pick a disease.  No one will agree at the end of the day.  Which means then you punt the ball, and the ball is punted in terms of time, dollars, delay, and for all the reasons that people line up on the opposite side of the pediatric argument to say it's bad things to do, it gives them fuel for their fire.

            So I think it is incumbent upon the agency and those of us who you've elected to advise you to, at some point in time, grapple with what is substantially similar so that any well-designed pharmacokinetic approach can get on the right track and do what it's intended to do.  So I applaud Dr. Sheiner for suggesting that and hope we can talk about it.

            With regard to population pharmacokinetics which is, Peter, I think central to your presentation, my question is always the same, and that is, when we use a pop PK approach in a pediatric study, are we aiming to explore relationships with age or, alternatively, are we aiming to define them?  I think we certainly can use pop PK, appropriately designed, to explore them. 

            But keep in mind that for those drugs where age is an important covariate with respect to metabolism or perhaps pharmacodynamics or response, an exploration and a definition can be very, very different with respect to impact because at the end of any pediatric program that's conducted, we are trying to do things on the quick, on the cheap, and on the small.  That means the generalization that we want for an entire population of patients that represents about 15 percent of the population of the United States is predicated on a fraction of the numbers of subjects and things that we would do in adults.

            So what you're proposing is very important, has incredible potential if done correctly, but we've got to be mindful of knowing how that tree starts and making sure that at the end of the day the people at level of the review divisions and the Office of Pediatric Therapeutics understand the power of this tool and how it can help them as opposed to what's going on now in the area of PPIs ‑‑ I hate to harp on this, but it a plays a nice tune ‑‑ where all the things that people around this table know, if you go into the little, bitty room downstairs on the third floor and you listen to the recommendations to a sponsor, you would have thought you woke up in the stone age.  The magic is still very much there, and this approach has to be used to make the process better.

            DR. SHEINER:  I'm glad we're moving off in a direction away from the techno-nerd thing.

            Mary, at the risk of maybe setting you up as a straw man, let me just think about what you just said.  I'm going to take away the word "pediatrics."

            Why wouldn't everything you said apply to a drug that's easily titratable no matter who it's intended for? 

            Now, I think Larry's response to that was, well, we need a dose in the label somewhere.  It's got to be based on some kind of evidence that people can refer to.

            So I think, if I can rephrase your question, what I think you're saying, to put it in sort of a Bayesian context, is there's something about having studied dose-effect relationship in adults that for a drug that may be isn't too toxic and is easily titratable with respect to effect might mean we don't have to do any more than that for children, which means, in a sense, that there's a prior somehow on the doses that you ought to be trying in children because, after all, the implication of an easily titratable drug that's not very toxic is that you don't have to get the first dose very right.  So you're saying I'll tolerate a much wider range.  And there's something about having studied it in adults that gets me close enough to that range that if I were to work it all out with a utility function and put a prior on what I think I can extrapolate from adults to children, I'd find I don't have to do a study at all.  I'd find my net benefit to society would be better served by going ahead and approving it for children than doing any more studies.  I think that's how I'd try to put it in a quantitative context.

            All I'd like to say about that is we could do that.  We could put it in a quantitative context so you don't have to sit there and feel a little uneasy about what other folks are suggesting and Larry doesn't have to feel a little uneasy about what you're suggesting.  We can put it all together and, just as we were talking about earlier with Mats' presentation, take a look at it.  What values do we need to place on doing ‑‑ what negative utility do we need to place on doing studies in children in order to make it worthwhile to extrapolate, given our sense of uncertainty from adults to children, and conditional on this drug, its safety profile in adults, et cetera?

            What I'm trying to say is I think that the same way of trying to be quantitative about this could answer that question.  I think it's just like the drug companies ask themselves now.  Am I better off going right to phase III and skipping phase II?  And there's a decision analytic argument that might say that that's really, in terms of net present value, et cetera, a good idea for this drug in this circumstance given what we know about it and its competitors and the fact that it's quite similar chemically and so on.  And all that can be worked out and it doesn't have to be an opposition of people not really understanding each other or thinking that they're coming from different places.  They're not.  They may value some things differently, but I doubt very much.

            And that's a worthwhile exercise, it seems to me, to do as we think about going to this in the future to try to approach this whole issue from the way in which we're saying people might approach the much simpler and therefore easier metaphor of a dosing decision, this whole issue of do we do a study at all, because it's really all the same.  It will get us into the habit of thinking that way in a place here where we've got some time to think rather than having to act right away.

            Peter, I wanted to get to your question.  It's really I guess my question about your question.  It's really a question for Larry, and it's this one.

            You called it a template and maybe words are important here.  But it sounds to me an awful lot like a piece of software.  At the sort of highest level, does the FDA want to be writing software for people who are then going to probably feel obliged to use that software because the agency says this is what you should do to design the study that you're going to do to come to us with?  I mean, the agency I think has generally been pretty careful about that and has had best practices and guidances and suggestions and all that kind of stuff.  But here's something we made and it's for you.  That's a whole new line, isn't it?  And do we want to go down that line?

            DR. LESKO:  Yes.  I think you're right in pointing that out.  I don't think we're talking about, at this point, the software that we're either going to develop and advocate, advance for drug development.  I think what we're talking about is a template that is based upon software that a reviewer might use in conjunction with a discussion with a sponsor that would prompt for the critical information that would go into making a robust study.  It's intended to sort of be a starting point for discussion or designing such a study that would be, at the end of the day, generally acceptable to the people that need to accept it in terms of its review and utility.

            I think the problem now is we don't necessarily see a consistency in advocating the design of these studies across different opportunities to do so.  This represents a way of channeling the discussion into the critical areas that would lead to usable results.

            DR. SHEINER:  I like it.  It's a good metaphor, but I think you may be getting too specific too fast.  It's sort of like a guidance, a statement about what are good things, what you want to see, what kind of principles you apply, and it's got a lot of wiggle room in it.  The thing about software is it hasn't got any wiggle room at all.  You put the inputs in and it's deterministic; out comes the answer.  So you have got to be pretty sure that's the algorithm you like.  So I'm not sure I would start there.

            But taking the metaphor of designing software to say that ‑‑ and that gets us to focus on all the key issues.  That's, I think, a good idea so long as at least you're contemplating maybe not going the whole route and going into the software business.

            Then I think fundamentally the questions are right, the first one being what's the minimum evidence I can gather that I ought to bother anymore.  Is there any difference?  But I don't think you'll get that from the same study.  As I say, I think this is essentially sequential.  That's the difference that I'd have between the way you put it.  It's not going to be one study.  There's something you do to figure out whether I need to go any further, and that may be quite different in design ‑‑ although I haven't thought this all out ‑‑ than what you do when you say, oh, I guess I better go further.  I better pin down the key PK parameters in this population and how it varies with disease state and other things.  My suspicion is that those will be two different activities.

            DR. LESKO:  I think the problem with the studies that have been done ‑‑ and it's probably why we haven't seen very many of this sort ‑‑ is the believability of the outcomes because these studies are not as well understood as obviously a sample-rich study design and the issues that go into analyzing a sparse-sample study using NONLIN or something like that.  Is the information reliable and how do I know that?  It's having to explain that to people who have to make decisions over and over or having designs that would lead to an acceptable result is sort of where we're heading here.

            That being said, if we have an optimally designed study to get at the questions we're asking about differences in pharmacokinetics for the purposes of dosing changes, can this method be confirmatory enough to stop there?  I think Greg used the word "exploratory."  I don't know if that was a suggestion that these studies at best can be exploratory for the purposes of designing another study or would they be confirmatory enough to say I know what the difference in clearance is between this drug as a function of age and maybe more age groups if I can do a sparse-sample approach, and thus I can recommend some different dosing for these age groups based on the study design without having to do a full study which in some cases limits the number of age groups that can be looked at.

            DR. KEARNS:  But, Larry, I think there are some examples on the books where it does work.  The whole program on montelukast to me has been an incredible success story because a very careful pop PK approach was taken.  We went down through all the pediatric populations now down to 6 months.  From my opinion, the appropriate variability was considered and the parameter estimates seemed to hang together, and when the data were taken the next step into showing proof of concept with respect to effect, the effect was there.  Consequently, the labeling of the drug has been changed multiple times.  It likely will continue to be changed based on that approach.  It worked.  It was done right.  There was a need to do it so we know the dose.

            But you're correct in that many other companies have not followed suit, so to speak, and for reasons that I don't completely understand.

            On the other side of the issue too, logistically ‑‑ and certainly Dr. Capparelli knows this because he's kind of in the business ‑‑ for the most part, if you have a pediatric study, a PK study, and you go to the trouble of obtaining repeated blood samples, you're obtaining them through a catheter.  If you're analytical method is such that you don't need a lot of blood, the bother in getting eight samples is no more than the bother in getting three or two.  IRBs anymore, at least pediatric ones, do not allow you to stick children several times.  So there are many times when a pop PK approach could, indeed, be used and it would be perfectly valuable and valid.  But a traditional approach is very achievable, more so than most people think.

            DR. DERENDORF:  Just as a follow-up, the other technique that is coming on strong is microdialysis where you can take as many samples as you want without taking any blood out.  So I think that will change the ability of doing studies in children.

            DR. SHEINER:  Just a quick response, Larry, to your question.  The essence of having a credible confirmatory analysis is controlling type I error, and controlling type I error involves essentially saying what you're going to do before you do it because you can't have feedback from the data to the analysis.  That doesn't mean you can't get valid conclusions from doing that, but you can't control type I error if you do that.  So I think that's the only issue.

            If you do a well-done analysis, then you know how uncertain you are when you're done.  That's an issue of design.  That is to say, given the assumptions we're willing to make and the data that we get, the sparse data that we gathered, do we wind up with sufficient precision on these things to make the kinds of statements we want to make?  That's an issue that unfortunately there isn't a lot of theory for because they're complicated designs and they're complicated analyses, and so you have to do it through simulation.

            But the key thing would be specifying beforehand exactly what models you're going to use, what procedures you're going to do, et cetera.  All of us in the business of doing extensive modeling have always felt a little anxious about that because you can't take away from me my ability to look at the data and decide what model I ought to use, but you have to take that away from me if you want to control type I error.

            So I think there will be an interesting issue there of how you balance that and whether, in fact, controlling type I error is as important as you sort of said it is by bringing it up.  I don't know.  I think I'm probably with Mary and Greg on this one, that we've got a lot of priors behind us and I don't think I need to pin it down to a fare-thee-well.

            DR. LESKO:  One of the inputs into the model is the variability within the kids or within the age groups that are being studied, and I'm not sure how that's been handled.  I can't recall the montelukast or, Ed, some of the studies you've done.  But the variability associated with the ‑‑ what you would expect with your different age groups ‑‑ how was that generated and how important is that in terms of designing these studies?

            DR. CAPPARELLI:  Well, in terms of looking at simulations and sort of real data that come out, the variability goes up as you go down the age group.  So from a pragmatic standpoint, one starts with at least an adult value and goes up from there.  But clearly there may be thresholds, some of them drug-specific and age-specific, in terms of dealing in HIV where we've got drugs that have major food effects, we've got formulation effects, and as soon as you cross certain thresholds, the variability is going to drop down.  But it adds complexity both on the design standpoint and the analysis standpoint when you have observed doses where you've got your compliance and you see no drug.  And this is in a CRC setting, but it's just the way that it behaves in this population.

            So there clearly are needs for evaluating distributions much more intensively, especially if we're interested in sort of the outlying regions, which I think most of us are.  But experience is that it's greater.  It's just variable how much more.

            DR. SHEINER:  I've got to ask about that.  I always thought the opposite.  Once you line up kids by some maturational marker, whether it's gestational age or whatever it is, I thought they're all newly minted coins and they all look the same.  In fact, I think I remember Bill Jusko saying that when you get real old, the variability goes up because you've run a longer race and you're sort of stretched out there at the finish line.  That's what I always thought the case was.  There was more variability in old folks than there was in little babies, again lining them at the right maturational level.  Obviously, a 3-month premature is not a term baby. 

            DR. KEARNS:  No.  Actually we're finding the opposite.  It's quite interesting because if you look at what Dr. Sheiner said, if you look at a 3A4 substrate in a healthy adult, there's 20-30-fold variability in the processing.  You look at it in a 3-year-old.  There it's about the same.  If you look at it in the 3-month-old, it's about the same.

            The problem is that as that little beast travels into adolescence and adulthood, the shape of the acquisition curve, if you will, changes, not to mention changes in body composition which are quite evident.  So there are a couple of moving targets that make it a particular challenge which, in designing a pop PK study, trying to estimate what your real variability is when you're up at the front, is not always an easy cookie to get.

            But there's got to be a way to do it.  I think what we hopefully will see, as we see the database, is some of the information that the agency is collecting is beginning to show us where these patterns might be, if you will, or breakpoints might be, and that makes things a bit easier to deal with.

            DR. DERENDORF:  And if enzymes change like that, what makes us assume that receptors don't?

            DR. KEARNS:  Absolutely, right.

            DR. FLOCKHART:  The absence of a phenotypic probe for the receptor.


            DR. LESKO:  We'll talk about genetic solutions tomorrow.

            DR. CAPPARELLI:  I would also emphasize at least a lot of the variability experience where there are these major changes, besides the newborn, is when you get into oral drugs.  So I think there are, at least from my experience ‑‑ again, the diet is different.  Controlling for when they take it relative to food, all those things that may be a little bit easier to do in adults is much more difficult in kids.  The formulations themselves ‑‑ while you can do bioavailability studies in adults and show similar formulations, it doesn't always extrapolate to kids.  So you have those sorts of things, I think, contributing as well to the variabilities.

            DR. VENITZ:  Any other questions or comments?         (No response.)

            DR. VENITZ:  Then let's take our afternoon break.  It's now 2:35.  So let's reconvene at 3:05, in 30 minutes.  Thank you.


            DR. VENITZ:  Let's reconvene our meeting, please.

            Our next and our last presentation for today on the pediatric topic is Dr. Gene Williams.  He's a pharmacometrics reviewer currently on detail at the Office of Clinical Pharmacology and Biopharmaceutics, and he's going to give us an update on the pediatric database.  Gene.

            DR. WILLIAMS:  Thank you, Jurgen.

            This is the title of my talk, kind of a long title.  The notion is to become better at predicting peds clearance and to take advantage of what we usually know at the time that we see peds submissions or proposals for peds studies, that is, child age, a lot of information in adults, and the knowledge of in vitro metabolism.

            I'm going to ask four questions of the committee.  It makes sense to show them first so you know where I'm headed.

            The first one is sort of an overall scope and method.  That is, is the general approach that I'm going to suggest in the presentation rational and logical?  And the approach proceeds from a very empiric method to a more mechanistic method.

            Secondly, is there anything special that you think that I should be aware of, some difficulties that you think I'm likely to encounter, and if you can identify such, how can I avoid them?

            The third question is of particular interest.  Are there data sources you could recommend?  One committee member has already been referred to as "in the business."  We'll get there I guess.

            The fourth question I have is, do you have any suggestions regarding the form of the non-physiologic-based PK mechanistic models?  That will become clear as I proceed, I hope.

            What brought on this project and what exactly are we talking about trying to accomplish?  What we'd like to be able to do is construct a model that allows us to predict pediatric systemic drug clearance from, as I said, adult PK and in vitro microsomal metabolism data.  That would be a short-term goal.  Obviously, we have a longer vision.  It seems like if you could construct such a model, it would aid us internally.  It would also be of potential interest to industry scientists, and finally perhaps even health professionals in the community could make use of such a model.

            It's probably appropriate to begin talking about the data that we have because that largely drives how we'll be able to model.

            The most fundamental data unit we're talking about using here is clearance, whether it would be from sparse or dense data, and age for each individual in the data set.  A number of demographic data also we would use; that is, the weight and height for each individual, renal function for each individual, and gender and race for each individual.  Finally, as I've alluded to earlier, we would also want to make use of what we know about in vitro metabolism data for each drug.

            I should probably add here that, as many of you I believe have appreciated, FDA is well positioned to do this sort of work because the data that we have often is very specific.  We not only see data summaries, but we also see individual data, which is a limitation that if you use literature data, you face, but we often don't face.  We usually get fairly raw data where we do have all these values.

            Our data set.  I've taken this statistic from a website that I've included here.  I believe it's publicly available.  I don't think it's just on our intranet.  In mid-March, about a month ago, we had 72 active moieties that had received pediatric exclusivity.  As Larry said earlier, most if not all of those would have pediatric data available.

            As I've gotten ahead of myself a little bit, the data that we see is usually raw.  It's actual measurements of individuals, not summaries across individuals.  And for the models that we want to explore, that's of a lot of utility.

            Further, our data is usually reviewable to an extent that literature data is not.  The analytical methods, dropout, salient features of data accumulation and choices made in data analysis are often presented to us.  So we can do a good job of assuring data quality.

            However, there are some limitations of the data we see.  First, studies are often not powered to compare PK across age groups.  People are submitting data to us for regulatory purposes, not always to discern carefully small age effects.  That's in distinction to studies sometimes performed by academicians where they're specifically trying to see age effects or, I should say, reasonably small age effects.

            The ages with the greatest difference from adults, often the very young, are often most poorly represented in the data sets that we see.  The data sets we see are motivated by the desire to treat, not necessarily the desire to see an age effect.

            Finally, most of the drugs that we see are not probe substrates.  People, again, are not asking mechanistic questions.  They're trying to get a drug approved.  So the ability to tease out effects may be more difficult since we don't have good markers for each individual effect that may be present.  As a result of this, it may be necessary to use some function of in vitro metabolism such as the Km, that is, for an enzyme, as a covariate when we do our analysis.

            I'm now going to carefully consider a data set that I took from the literature.  I did this for a number of reasons.  As Larry said earlier, organizing our data set is a considerable effort, and since this data set was sitting out there, I thought I'd use it not because we want to analyze it, but because it makes a good platform for discussing the methods that we intend to use.

            This is taken from Ginsberg, et al.  There are somewhere 21 and 27 drugs represented here.  The y axis is children's clearance relative to adult.  You'll see at the bottom of the slide I've described the units that are used.  These data are standardized for weight ‑‑ they looked at kilograms ‑‑ and age.  Age here is not a continuous variable.  Rather, it was grouped categorically, a decision the authors made.  They took these data from the literature.  It's not their own data, and I guess the data lent themselves or, for some reason, they organized for categories in this way.

            I've shown a line at 1.  That would be where the child is exactly the same as adult.  You can see at ages 12 to 18, that's accomplished.

            I don't want to give much attention to this slide, but the question is likely going to arise as to what sorts of drugs they were.  This is also taken from their database.  This database is available on line for anyone who wanted to explore it.  As I said, I don't want to discuss this, but the drugs represented a number of classes.

            Before you attempt to model these sorts of data, it's necessarily to normalize clearance.  The reason why is you want to consider each drug on its own and not have your analysis complicated when you compare drugs whose adult clearances differ widely.  So the method we chose to normalize clearance, similar to the method that Ginsberg used, is to divide each individual pediatric clearance by the mean adult clearance.

            Again, this is Ginsberg's data.  The y axis is clearance ratio versus age.  However, unlike the plot that I showed you from their paper, this data has had the element of weight removed.  So this is no longer adjusted according to the representative body weight of the data.

            The line shown here is a simple least squares fit, no weighting has been performed.  This is unlike what we intend to do when we analyze our database.  We'll probably use NONMEN extended least squares.

            As you can see, or perhaps as I need tell you, I have fit the effect of weight on clearance in this plot.  So the equation is shown beneath the line and it's a simple exponential relationship.  The maximum ratio I allowed to happen was 1; that is, where the ratio of child to adult would be 1.  So essentially this is one parameter.  I fixed the maximum ratio.

            As you can see ‑‑ I was somewhat surprised ‑‑ it provides a reasonably good fit.  This is a little at early ages, and this is sort of consistent with what we generally expect to happen, that during development and maturation, things may be a little different.

            DR. SHEINER:  Excuse me.  Just to clarify.  I guess I'm not sure what it is.  You haven't fit the data on the y axis to the data on the x axis.  Your equation there is in weight which is ‑‑

            DR. WILLIAMS:  Correct, yes.

            DR. SHEINER:   So tell me again what I'm looking at.

            DR. WILLIAMS:  Indeed.  I have not fit age here.  I fit weight.  So what I did is, although I'm representing it age because that's the thrust of our interest, before I went there, I wanted to isolate the effect of age as opposed to the effect of weight.  So first, I fit the effect of weight, and in the next slide I will then add in the effect of age.  I should have clarified.  Thank you.

            DR. SHEINER:  So the brown line that I'm looking at there is the equation that you wrote in the lower right-hand corner.

            DR. WILLIAMS:  Correct.

            DR. SHEINER:  And the way you know where to plot it on the age axis is what?

            DR. WILLIAMS:  By converting each age to a weight based upon standard CDC pediatric tables.

            DR. SHEINER:  Okay, but the fit was actually to the blue points where you knew what those weights were.

            DR. WILLIAMS:  I did not know what those weights were.  I had to go by the age.  So if we back up a little bit, these are the ages I had, but I have summary data.  I don't have individual data.  So what I did is for each bar I took the mean age.  Then I went to the CDC tables to get the weight ‑‑

            DR. SHEINER:  Transformed it to a weight.

            DR. WILLIAMS:  Exactly.

            DR. SHEINER:  So if we go back to that picture, we're really looking at a transformation ‑‑ a fit of the blue points on the y direction to a to a transformation, defined by these tables, of the data on the x axis.

            DR. WILLIAMS:  Indeed.  It would have been more straightforward to plot weight on the x axis, but the reason why I didn't do that is twofold.  One is for continuity with the next example where I'm going to fit weight and age.

            DR. SHEINER:  Where you're going to have both pieces of data.

            DR. WILLIAMS:  Exactly.

            DR. SHEINER:  Okay.

            DR. WILLIAMS:  Is that clear to everyone?

            DR. JUSKO:  From the previous graph, that should start at .5, at the age near 0.

            DR. WILLIAMS:  We have birth, which is ‑‑

            DR. JUSKO:  The ratio at birth on that graph is around .5.

            DR. WILLIAMS:  Correct.  But the y axis here is different because these are weight-adjusted.  The y axis here is child clearance divided by kilograms, quantity divided by adult clearance divided by kilograms.

            DR. SHEINER:  You're going to regret ever having shown that picture.


            DR. WILLIAMS:  So what I've done is I've taken out the kilograms so I could fit weight.

            DR. RELLING:  What is your goal?


            DR. RELLING:  Why would you do that?

            DR. WILLIAMS:  The reason why I chose to do it this way is because you want to independently describe weight effects and age effects.  You expect there to be weight effects, and you also perhaps expect there to be age effects.  But you want to be able to independently address are there age effects that are not simply a consequence of weight.

            DR. RELLING:  Okay.  Let's see what you have.

            DR. WILLIAMS:  I won't suffer from this difficulty when I have the FDA data because it doesn't initially present itself as normalized to kilograms.  Is this making sense a little bit more now?  Okay.

            So in spite of the fact that the x axis is not saying so, I have fit the relationship between this ratio and weight here.

            The next model I looked at is a combination of a weight effect and an age effect.  The weight effect is what you saw on the previous slide.  Here I've added in the age effect.  The effect of the two summed together, each of which is a simple exponential relationship, is shown with the green line.  The weight effect, which is what I described previously, is shown with the line in the middle, sort of the pink dashed line, and finally, an age effect which is what's new on this slide.

            Now, you'll notice that I fit 6 points with 4 parameters.  My point here is not to show that I can draw pretty lines.  The reason why I'm presenting this to you is because it shows the sort of strategy you might take when we have a larger database and how we might think about developing the models on our own data set.

            Did this confuse everyone further?  Can I aid anyone?

            DR. SHEINER:  Yes.  You don't have any independent information in this particular data set.

            DR. WILLIAMS:  Correct.

            DR. SHEINER:  You have weight, which is a deterministic function of age, and then you have age, which is a deterministic function of age.

            DR. WILLIAMS:  Indeed.

            VOICE:  (Inaudible.)

            DR. SHEINER:  No.  He got it from a table.  So it's a deterministic function of age.  So if I were to write your equation, it really is Rmax 1 minus E to the minus Kf of age plus Rmax 1 minus E to the minus K of age.  So all you've done is done a shape thing.  By restricting it to exponentials, you get more information out of two exponentials than one.  It's just like when we have time.  And you said the right thing at the end which is this is just an illustration.


            DR. WILLIAMS:  I would agree.  What we do have going here, though, is that the shapes ‑‑ I haven't looked at this specifically, but the shapes ‑‑ well, actually I have to an extent.  The shapes are different.  If you plot ‑‑

            DR. SHEINER:  No.  I'm saying if I used a spline or some flexible function of age, I could only get one term in age because it would be as many parameters as I needed, but because you've broken it up into two exponentials, you can get two terms in age because they don't have the same shape because the function of age that weight represents is another shape change.  So it's just like saying I have a polynomial in age.  It's not a polynomial.  It's a flexible function in age.

            DR. WILLIAMS:  I would agree.

            DR. SHEINER:  But it doesn't prove that you fractionated an age effect away from a weight effect.

            DR. WILLIAMS:  I would agree on that.

            DR. SHEINER:  Okay.

            DR. WILLIAMS:  Interestingly, I was somewhat surprised it turned out as consistent as it did because one thing I did as a check ‑‑ like I said, it wasn't essential to my purpose because I'm just trying to show you the kind of strategy I would employ.  But one thing I did do is I went back and switched the order in which I added the two, and interestingly I got the same relationship.  I don't know if that's surprising, meaningful, or what, but it did happen.

            DR. SADEE:  Can I ask you about this?  Using weight as a scaling may not be all that appropriate.  So rather than saying there's an age effect, is there any information on body surface area which would just do away with this ‑‑

            DR. WILLIAMS:  No.  The answer is no.  In this database everything was normalized according to body weight, and other than the numbers as presented, which were always per kilogram, I had no raw data.

            DR. SADEE:  Well, could you translate this into body surface area which would provide you with a different scale and it may actually do away with the need to invoke age?  Because body surface area changes with respect to weight and age, so it may account for both.

            DR. WILLIAMS:  Perhaps.  When I actually do this on our own data set, the path that I intended to follow was, first, to describe the effect of weight or mass or BMI, ideal body weight, BSA.  I would investigate a number of weight metrics.  Then whichever one fit best I would use and then proceed to looking at what an age effect, in the absence of a weight effect, would be to see if I can describe one, and if so, what it is.  That's my intent here.

            The reason why I did this in this way is to try and not have the effect of age confounded by weight and the only weight metric I had, which I couldn't pull out in any other way, was kilograms.

            The path, in addition to what I've just indicated for the models that we would consider ‑‑ and this is speculative.  It will be driven by data, but this is sort of how I see it going and thought it would be useful to share.

            First, we would consider adding exponentials if the data supported it.  You've seen a two-exponential fit, and as Dr. Sheiner said, it's effectively both descriptions of weight.  But the idea is if we should discover that at early ages there appears to be a unique phase of the curves ‑‑ it doesn't occur at later ages ‑‑ and is not as simple as a single exponential fit, then we would consider adding additional exponentials, a structural model, if you would, of more than one term.

            One thing also I haven't shown here is an offset for the age effect.  That is, you might not expect that the age effect would be 0 at birth.  Whatever processes are occurring, maturation may begin ‑‑ almost certainly did begin ‑‑ in utero.  So you might want to add a term that would give an offset for the age effect.  Again, this is speculative, but not based upon a real data set that I'm currently evaluating.

            Finally, we would begin to look at more physiologic covariates.  Can you enter covariates such as the percent excreted unchanged, the Km for a given enzyme, the Km ratios across the enzymes for which metabolism of the drug is responsible?

            Finally, the approach I've shown is very empirical.  Whether or not it will be successful is a question, and if it were unsuccessful, the next step would be to consider models which are more mechanistic.

            The first mechanistic model perhaps to be considered would be one that is less than a full-blown PD/PK model but which does incorporate what I'm calling process constants, such as GFR, such as Km and the percent non-renally eliminated. 

            Finally, if even such mechanistic models were not successful, you might have to go into more of a full-blown physiologically based pharmacokinetic model.  The difficulty with that is clear.  Usually it's difficult to obtain data to support such a physiology-based model.

            Would anyone like to tell me if they think this is reasonable?

            DR. VENITZ:  Thank you, Gene.

            DR. DERENDORF:  I have a question for clarification.  Are these all lumped together, hepatically cleared drugs, renally cleared drugs, high extraction, low extraction, no difference?

            DR. WILLIAMS:  Yes.  The initial cut would be the raw data is age versus clearance independent of the physiologic mechanism by which drug is excreted or eliminated.

            DR. DERENDORF:  Well, then I would not agree with the first question that it's logical because I would expect there to be major differences depending on the mechanism of clearance.  Obviously, for a renally cleared drug, the enzymes don't matter, and for high extraction drugs, the intrinsic clearance would matter, and so on.  So I think to break it down into several subgroups would make a lot of sense.

            DR. WILLIAMS:  You might expect that as the physiology differs, so will the relationship between clearance and age.  But I guess it's dependent upon a few elements.  One is to what extent is each one of those different elements not a function of allometry.  In other words, if the enzyme maturation and, say, GFR are both linear functions of weight, then you might expect that this approach would work.

            Now, the clear difficulty is at early ages, there's a whole literature, which many of the members of this committee have helped develop, that speaks to the fact that that is sometimes or perhaps oftentimes not the case.  So the question is perhaps, how much data do I have that can address that?  Specifically, do I have a lot of data at early ages or how much does it vary?

            Would you agree that ‑‑

            DR. DERENDORF:  I'm not an expert, but if I recall correctly, glomerular filtration rate in a 2-year-old is almost like an adult.  Right?  There's no difference.  Whereas, clearly the number shows that for a 2-year-old the clearance per body weight is almost twice as high.  So there are differences depending on the route of elimination, clearly, because you wouldn't expect a difference for a renally cleared drug.

            DR. LEE:  Can I add to that?  What we're planning to do is look at different drug classes.  For example, we will look at a drug class which is purely renally cleared and then try to see what's the relationship between clearance and age.  Then we will look at a bunch of drugs, for example, 3A4, purely 3A4, and then see how it's going to change in our clearance versus age.  So that will probably address your question.  And we know from the literature that the maturation of different enzymes are very different.  I mean, 2D6 may be fast and 3A4 may be slower.  This is what we planned.  We want to look at different drug classes and build a model perhaps about one drug class at a time, and then finally we have an individual model.  Then we will look at a drug that has a combined pathway, maybe a drug with 20 percent 3A4, 40 percent 2D6, and see if the model can actually predict the age effect.

            DR. SADEE:  Do you consider changes between males and females and the various sexual developments and so on?  If you talk about maturation of enzymes, which sounds a little fuzzy of a term, but males and females are probably very different, but that may also depend on the age.  I don't know.

            DR. KEARNS:  Actually with respect to drug metabolism, they're not very different at all.  There are a few examples of substrates for P450s that during adolescence differ a bit and it probably has to do with the things that make for differences in linear growth more than sexual maturation.  But for the most part, it's pretty boring, boys and girls, before puberty.

            DR. SHEINER:  If you go back to that picture, the one with the graph that we had the problem with, if I just connected the dots, I'd have a function defined by line segments of that y axis value, the relative clearance, versus age.  And everybody would be completely clear that I had totally explained all the data with my function, and consequently there's no way to partition out the effect of age from the effect of something else in your case that is a function of age, which is weight.  So the fact that you could partition out, means you made some kind of very powerful assumption that allowed you to take this function of the x axis and see it as a separate contributor to this curve which is practically connecting the dots.  So that's the point I'm making. 

            It's nice that you started with this one because here we all know there's only one variable age.  We don't know anything about weight, not the real weight.  So you've got perfect what's called multi-colinearity or you've got a perfect problem that the one substitutes for the other.  The value from your table can be translated back exactly into the ages and the other way around.

            So this is what's called, you know, an ill-posed inverse problem or an unidentifiable model.  It depends on what field you're in.  And the only way you can get something out of those ‑‑ and I'm going to get to the key question here in a minute.

            But the only way you can get something out of those, if you're trying to learn about these different effects, is you have to make some very powerful assumptions, some assumption that allows you to separate.  Here, the assumption you made is the exponentiality.

            So I could say, if I didn't know anything about this, if you have a solid basis in physics for that exponential equation, something at the level of theory that's as powerful as physics, that says each of these has to influence and the only way it can influence the spread of these points is through a rising exponential, then I'll believe that what you sort out of the two effects is right because that's the key piece of information that you've added that you told me essentially everybody knows this is true.  But if you say, well, I just ‑‑ the exponential because it kind of went up and then it went over, you know, then there's no reason to believe you've got it sorted out right now.

            I don't want to criticize this picture because you were clear this was an example, but you're going to have a similar problem with the real data.  That is to say, there's going to be very high correlation between age and weight.  So you got an almost ill-posed problem.  You got an almost unidentifiable model.

            So if your goal is to sort out the independent effects of things we can measure like size ‑‑ even that's a surrogate for something else ‑‑ and what we will all agree is an unknown age-related something called maturation or something like that that we don't measure when we measure renal function, that we don't measure when we measure these other things, if your idea is to sort that out and then look at the shape of that thing, we're going to have a lot of trouble believing it even when you're not as bad a situation as this because of that high colinearity.

            So I told you I'd get to the question.  So my question is, what's the question?  What do you really want to learn?  If you want a predictive equation, you could do anything.  I'm not being facetious.  If you want a predictive equation and you're not going to interfere with the system, you're not going to deliberately change people's ages or weights or whatever, you can just let it come as it falls, and you find some general predictive equation for clearance of all drugs as a function of a few easily measurable things, that would be completely valid as long as you don't go in and mess up the system, even though it doesn't give you an understanding necessarily of what the causes are.

            But on the other hand, if you say no, I want a predictive equation for a whole new drug, then you're interfering because you're not sampling from the same world.

            So I guess my question is, what's your question?

            DR. WILLIAMS:  First of all, I would say if it does turn out as you're saying and you do have this high colinearity and this difficulty, it seems to me that would be good news because it means that you have the ability to describe the relationship between the ratio and the age as only a function of weight.  If the data is well described by small or perhaps a single parameter, that would be fantastic.

            But to get to your later question, perhaps you can educate me some more.  It's not so clear to me if you had the situation that you're defining where you did have this high amount of colinearity and you then parsed out your drugs as a function of all of the things that you can look at, metabolic route, percent renally eliminated unchanged, et cetera, and you found that you could not identify covariates which interfered with that relationship, then wouldn't it be legitimate to extend it to the new drug?

            DR. SHEINER:  That would certainly be an empirical basis on which you would guess that a new one would look just like the other ones because you had a whole bunch that all looked the same and you hadn't chosen them for that purpose.  I agree it wouldn't be a mechanistic basis.  That's why I said sort of a whole new drug class.  Somebody might say, sure, you know, you've been right 85 times out of 90 so far, but you've never looked at one like this before.  So that's the problem.  That's what I'm saying. 

            I agree with you, it's good news if they all look the same because it gives you more faith that the new one, even though we know it's a little bit different, will also look the same.  But that's just counting how many times have I been right out of how many times I've tried.

            So is that your goal?

            DR. WILLIAMS:  Yes, that would be the goal.

            I'm new to this area, but from my read of the literature, I think it's unlikely that we would see that at young ages.  Now, the question is the quality of my data.  Do I have a lot of data at young ages?  Because I would expect that that's where they will separate from a simple function of weight.

            DR. JUSKO:  Don't you at each age have data from many children where there's a distribution of weights at each age?

            DR. WILLIAMS:  Yes.

            DR. JUSKO:  So if you do have that, then you have the ability to discern the separate effects of age and weight.  So I don't understand Lew's ‑‑

            DR. WILLIAMS:  Yes.  Our data set will be an improvement of that.  But if the colinearity is ‑‑ if they're very highly correlated ‑‑

            DR. JUSKO:  Certainly they're highly correlated, but there are factors in addition to age that control weight that you would be able to discern through this kind of analysis.

            What do you select as the upper limit in age or weight to reach the maximum?  Because for many physiological functions, the graph will look like this for a certain age range, but reach a maximum at about 18, and then as we all know, we steadily deteriorate until we reach some lower level.


            DR. JUSKO:  So a more insightful empirical function may be more of a U-shaped type of ‑‑

            DR. WILLIAMS:  I didn't grapple with that here obviously.  What I did is I just fixed it to 1.  But yes, it's a question.  I guess what I would do is I would try and look at as great a wealth of adult data as I can to see if there is an age relationship, and if so, where's the maximum I suppose.  But you're right.  That's something that will have to be worked through.

            DR. SHEINER:  Bill, I think you're right except I'm not sure what you get out ‑‑ so if I've got different ages and I've got a lot of weights at different ages, then it's quite true that if I want to say there is some sort of sum of these two effects that's operating, then if it were only weight, then I should be able to go across age and find it at the same weight.  Everybody had the same clearance.  And if I didn't, then I'd have to explain that by age.  Is that sort of where you're ‑‑ yes.

            DR. JUSKO:  Age is likely to be the strongest determinant, but then we have to bring in genetics and diet and the rest as additional determinants of weight.

            DR. SHEINER:  Yes.  I think Bill is completely right, that having independent variability in age and weight will help.  It's just how much of that do you need to feel comfortable about what you get out.  And strange things like Simpson's paradox can happen where, as you move from age to age, the regression within each age versus weight could be actually opposite the direction that it was when you do the whole thing.  These sorts of things happen.  So it's just a matter of what the data contain.

            I guess one of the things that bothers me is that with kids it's going to be a very, very strong relationship; whereas, with adults we expect a small age effect ‑‑ the deterioration that Bill is talking about ‑‑ and a large weight effect. 

            But again, for predictive purposes, it doesn't make any difference.  If you're not going to do anything different the next time, then your data has got all the information.  I suppose what you're saying is an interesting result here would be if there was a relatively simple equation that predicted most of what you see across lots of drugs.

            DR. WILLIAMS:  Yes.  I anticipated that I would get the reaction that I believe Hartmut is expressing, which is this is sort of naivete to expect that it would go that way and that it's probably very important to incorporate physiologic covariates.  But one of the things that drove me to think about it in this way is it made sense to operate on parsimony and start simple and see what the data set would support.

            DR. DERENDORF:  But you want to use anything that you already know.  So if you know that you have gentamicin, you can use glomerular filtration rate as a pretty good estimate.  So if you know what the physiological value is for that, you should use that and not ignore it.

            DR. KEARNS:  Just two comments, one to speak in support of doing this in the context of getting you in the ball park because I think it clearly has the ability to do that.  You're right in that for very young infants and perhaps even up to 6 or 8 months of age, it's not an issue of weight for much of the maturation.  Probably post-conceptional age falls out as a way to best predict these things because you do come to the field with a little bit of activity depending upon when you come to the field.  That's clear.

            But from a practical standpoint, I was sharing at the break our experiences in doing a study with a drug that was 100 percent renally cleared.  This drug was studied under that list of 72 who've been given exclusivity, and they got their 6 months of extended marketing exclusivity and had a big party and everyone was happy.

            But in going back in time and looking at those studies, we were able to simulate the results of the trial before we enrolled one patient, as you might believe you could do.  We made an argument that we thought was reasonably passionate, but perhaps not sufficiently so, that the trial that was done needed only to contain children in the first 3 months of life, that everything else could be predicted.  We were sent away believing that, indeed, some revelation would occur.  We could use the knowledge that we had to simplify and improve and streamline the process only to find out that we were wrong, that ultimately we were expected to fill in all the pieces of the puzzle of the barnyard despite knowing that it indeed contained animals of identity that was known.  At the end of it all, there was a lot of time, effort, and money spent for no good reason.  No good reason.

            So for all the pimples on an approach like this mathematically, this has the ability to improve the design of pediatric studies if, within the halls of this wonderful agency, it would just be used to do so.

            DR. WILLIAMS:  I guess, Larry, that helps justify my detail.


            DR. LESKO:  I was just going to comment on what is the question.  Actually Greg's comments are discouraging to what I was going to say.


            DR. LESKO:  Nevertheless, when Lewis asked the question about what is the question, it would seem what we want to know is not necessarily the empirical relationship we're talking about, but rather the more mechanistic one where you can look at a drug and it would be a whole new drug, but look at it not as a whole new drug in a therapeutic class per se but a whole new drug with certain attributes of processes of elimination, cytochrome enzymes, what we know about Bmax and Km's of those enzymes, and based on that information and based upon the analysis of a pediatric database, know where there are breakpoints in the age groups and perhaps do a limited study that might bracket age groups, and then you can fill in the blanks in between based on some model to say I know this from these relationships between routes of elimination and age.  I might do some limited studies, but then cover age groups I haven't actually studied in terms of extrapolating that information.  It's kind of what you were saying in terms of knowing something in the first 3 months and then using that to predict the rest of the puzzle, but it just struck me that if there's a difficulty in doing that with a drug renally excreted, the difficulty becomes magnitudes more for a drug that is out the enzyme system.

            But nevertheless, that's our noble mission here to try to look at the database.  Maybe we just need more examples of this using data we already have as opposed to something new.  I don't know, but I think where we want to go eventually is to take attributes of a drug and be able to make better predictions and maybe even excuse pediatric studies from being conducted if we're confident enough that we can predict clearance in those age groups.

            DR. KEARNS:  And the other thing ‑‑ and what I'm about to say is not mine.  I really owe this to Steve Spielberg who began preaching this some time ago ‑‑ is that if you can define a breakpoint, if they really exist, then it's possible to design your trial in such a way to enrich it so that you can get the most information out of the least numbers of subjects studied.  That's okay because in the process, we don't compromise the end game result, and we also don't put children in trials just for the sake of confirmatory purpose.

            Parents always want to know, especially for a nontherapeutic trial, they say, explain to me again why this is important.  As an investigator, you really have to be able to tell them that.  If they're convinced, you have a child on your study and you have good data.  If they're not convinced, maybe you shouldn't be either.

            DR. KARLSSON:  Maybe in addition to what you presented here, your analysis, if you looked at it, could also settle the debate that was before the break regarding what about variability in elimination capacity with age.  Does it actually decrease?  Does it increase?  Is that dependent on the elimination pathway, et cetera?

            DR. WILLIAMS:  Yes.

            DR. FLOCKHART:  I guess I'm back to what's the question.  It seems to me, Larry, that you gave a different answer to the question. 

            Greg, I think if the playing field is so big and has very significant error within it, I've got to ask the question, what's the point of asking where it is on the playing field.  I can't tell by looking at the playing field whether the ball is in the goal or is at the centerpoint.  If it's a very vague thing from doing this kind of activity before, I'm not sure knowing it's on the playing field is a valuable exercise.

            On the other hand, I guess we're all biased as scientists towards believing that a more mechanistic, physiologically based approach would work better.  But I have to say that that real hypothesis even in adults has not been really hard core tested.  We don't really know that.

            So this becomes, therefore, a testable hypothesis.  Each time you add a new drug, you're testing the idea, and if it turns out even between the ages of 1 and 16 that things fit, that would be a tremendously valuable thing to know that we kind of got gratis, we got free as we went along. 

            The error is the key, though, because it's very variable.

            DR. KARLSSON:  Are these 72 studies intravenous studies, or are you going to mix IV with oral studies?  If so, how do you handle bioavailability issues?

            DR. WILLIAMS:  Until I look at the actual data, what drugs are available and pick them, I won't know that answer.  In other words, how do I choose among the 72?  Perhaps the easiest way to start would be to look at IV drugs, but on the other hand, perhaps you would have a willingness to accept drugs whose metabolic route is well defined and is thought to be largely one process.  So the 72 will be a mix, but which ones I choose to look at first and how I develop it is something that we have to consider.

            DR. CAPPARELLI:  I would just like to echo what Mats was referring to, that the oral component is going to be huge, especially in the younger age groups, and everything that's been said before about some of the formulation issues.  So I think that it's going to take careful selection.  I'm excited to see this direction, and I lean on the mechanistic fence of things, starting off with what we know and building on it.  I think renally eliminated drugs make a lot of sense from the standpoint of what we know about renal function, how we can measure it or how we can, at least, estimate it in different pediatric populations and relate it.

            But you start getting into drugs where there are bioavailability issues in adults, it's going to go all over the map, and you're going to have the additional confounding issue if you've got active transport or gut metabolism.  And those may not parallel what's going on in the liver.

            So, again, trying to simplify it and at least starting it at points that I think you'll have buy-in and belief in the model I think is very important.

            DR. SHEINER:  How variable will be the way in which the clearance was determined in the individual children across these studies?

            DR. WILLIAMS:  I guess fundamentally you can separate into sparse and dense, and we're likely to see both of those.  But beyond that, how studies are conducted, sampling times, populations, numbers, probably a wide range.

            DR. DERENDORF:  Just to clarify, because I was under the assumption from the beginning it was only IV data.  Now you're saying that there was some oral data.  So they were the ratios of the oral clearances between kids and adults that you showed in that very first table?

            DR. WILLIAMS:  My recollection of ‑‑ oh, this is perhaps unfortuitous.  This is slide 4, not data IV.


            DR. WILLIAMS:  So the answer to the question is yes.  Certainly some of these or perhaps all of them are.  I did not separate this out into oral versus IV.

            Now, when we actually perform this on the FDA database, of course, we will have the luxury of choosing the order in which we consider the drugs.  Obviously, there's an advantage, especially given Dr. Capparelli's comments, of beginning with the simplest case which would be IV drugs.

            DR. DERENDORF:  The probability of getting anything useful out of it then is very low in my opinion.  I think you have too many things that are lumped together here.

            I'm amazed at the ratios, that they come out to be so close to 1 in that figure there.  If oral bioavailability is included there, it's almost hard to believe.

            DR. LESKO:  But in our own data set, we can control for that.  We can select drugs, as we said, taking care of those differences.  We would combine drugs in different ways that take those similarities into account, whether they're IV or oral, and not mix them.  He's working with a published data set, but I think your question and issue would be resolved if we picked from the data set appropriately that we have within the FDA.  Isn't that right?  Am I misunderstanding?

            DR. DERENDORF:  If you stick to high bioavailability drugs where there is not a big difference and where we don't expect a big difference, but if you have high extraction drugs in the group, you would really get numbers all over the place.

            DR. LESKO:  But I think one of the plans would be to look at different ways of categorizing the drugs that we're looking at to see if that makes a difference or not.

            DR. DERENDORF:  Well, the question changes completely.  Initially it was a question of how do metabolism and clearance develop, and now we have included bioavailability.  We have formulation issues.  We have transporter issues.  We have intestinal metabolism, I mean, a whole bunch of things that happen all lumped in one number.  And I think the chance of filtering out anything that teaches us something is very, very slim.  We'll get some kind of an average curve and we can fit a line through it, but what does it mean?

            DR. KEARNS:  But, Hartmut, if I told you that the data set that they had had over 300 patients intensively studied with midazolam, half of them on oral, half on IV, from ages of 6 months to 16 years, all of a sudden it becomes a little more interesting.  And that's in their data set.  So there is some gold in there to be mined.  But your point is well taken in that it's not something to be done in a reckless way not paying attention to all the assumptions and limitations.

            DR. DERENDORF:  Again, I think where I'm coming from is to utilize anything that we already know, and the physiological information about blood flow, for example, is something that doesn't vary that much.  That should be included in the data analysis.  It should focus on intrinsic clearance as the number to correlate with.  I think then it makes much more sense.

            Then you have a chance that you can identify maturation rates for the various enzymatic pathways that you then can use to extrapolate for new compounds.  Once you know that for a new compound which is the breakdown of different pathways, you already know how the rate of maturation occurs and you can make good predictions without any study.

            DR. WILLIAMS:  If you're right ‑‑ and I initially tended to think that way too, but like I said, this is new to me ‑‑ then we will get there because what will happen is the simple models will fail.

            DR. LESKO:  Gene, do you know the size of these studies, just speaking about the small to large of the pediatric studies in the database?  What's the typical n in these studies in terms of getting an estimate of precision of the pharmacokinetic measurements?

            DR. WILLIAMS:  I really don't.

            DR. LESKO:  Greg, what do you do?  Or, Ed, what's the typical size of a PK study in a pediatric population in your experience in terms of a single age group or all the groups?

            DR. CAPPARELLI:  When you say all the age groups, that incorporates a little bit of gray as you get further on down and looking at degree of maturation at birth.  Some of those studies get to be very large.

            Typically most of the stuff that one sees may not be optimal, but rarely do you see anything less than 50.  Most of it is in the 100 to 200 range if it's incorporated into the safety trial.

            But the driving force isn't often the optimal PK component.  It's really the other aspects of the study. So again, the precision issue really comes near the cut-points which I think was brought up earlier.  There is often a lack of information where the action is, unfortunately.

            So having a large data set that has three patients under the age of 2, and you get this spread that's here, here, and here, and then trying to make some sense that no, nothing is going on down there or there's something very dramatic going on down there really is based on the belief of who's looking at the graph rather than any real aspect of the data.

            DR. KEARNS:  Larry, for the phase II things done under most of the written requests that I've seen, it's about 24 subjects to 36.

            Some of it's ‑‑ I need to pick my words right because we're on tape, but it's interesting some of the ways the designs are done.  For instance, if we had a drug whose metabolism we knew or believed changed greatly in the first 3 months of life, you might see a written request that asks the sponsor to include 24 infants from the age of birth to a year and that the infants should be equally distributed across the age spectrum, so there should be at least 3 infants or 4 infants in the first month of life.  And oh, by the way, you can study babies all the way down to 800 grams.  Now, the chance of coming out of that at the end of the day with a revelation of therapeutic utility is slim to none.

            But unless my experience is somewhat not representative, this is happening every day under the context of negotiating a written request for drugs studied under the Best Pharmaceuticals for Children's Act, which is why I believe that until we put some of this bit of science and ingenuity into the action plan, we're really not serving the intent of the people who put that act together or, even worse really, the children as we try to make a fact out of fancy many times.

            DR. VENITZ:  Any other comments, questions, or further discussion?

            DR. WILLIAMS:  Are there data sources anyone here can recommend?

            DR. VENITZ:  Dr. Sheiner.

            DR. SHEINER:  At the risk of stupefying Mary, are you going to go back to the original data, the measurements of concentrations versus time, or were you thinking to use those clearances?

            DR. WILLIAMS:  The notion is that we would just go back to the clearances.

            DR. SHEINER:  Okay.  And those clearances will be some, you know, from 15 samples after a single dose and a nice area under the curve.

            DR. WILLIAMS:  Right.  Some of them would probably come from population post hoc and some would come from dense.

            DR. SHEINER:  Well, you want to think about how you mix those because those posterior Bayes' estimates are funny creatures.  They're centrally biased.  So they do odd things to regressions when you use them either as the explanatory variable or as the variable to be explained.  So you need to think about that.

            DR. WILLIAMS:  Yes.  It seems to me it gets pretty complicated if you don't do that, but would you like to propose an alternative?

            DR. SHEINER:  Well, if you don't do that, it doesn't get complicated.  That complication goes away.  If you use the original data, the complication I was just talking about goes away because you're not summarizing the data with a strange estimate.  But it does mean that you have lot longer run times and a lot more modeling to set up and all those nasty things.

            I don't know.  I would almost be tempted to say, since you're going to use them in subsequent regression and you have lots of data, that you should use unbiased estimates of each individual's clearance which you would get essentially by taking the prior variant system infinity or fitting the trapezoidal rule with three points. I mean, that's bizarre to talk about.  But I don't know.  I have to think about it.  It's not obvious.

            DR. WILLIAMS:  Perhaps we can discuss this further and I can come to the committee as a whole or perhaps even yourself showing you the actual characteristics of the data set.

            DR. VENITZ:  With that, I think we're ready to conclude.

            DR. WILLIAMS:  Actually if I could make a ‑‑

            DR. VENITZ:  You conclude.


            DR. WILLIAMS:  First, a number of the committee members are very active in this area.  If you would like to share your data, we would certainly welcome it.  As I said, some limitations of our data is we often don't see very young ages and we often don't see probe substrates.  So if you'd like to contribute, we sure would welcome that.

            Finally, should the very empirical models not be successful, the form that you would use not as far as a full physiologic-based model, but incorporating some of these physiologic covariates into the description of age versus clearance is not entirely clear to us.  We've worked on it a little bit, but we don't have any firm conclusions. If anyone would like to contribute here or perhaps even off line how they would see the form of those equations running, we'd be grateful.

            DR. SHEINER:  Just a quick question about that.  Usually PBPK to me means the various compartments connected in various ways and blood flow from the gut going to the liver and things like that.  Is there a large collection of physiologic models of clearance?

            DR. WILLIAMS:  I'm not sure I understand.

            DR. SHEINER:  Well, I mean GFR and renal clearance.  They seem to be linked, and usually people do it linearly.  I haven't seen too many what I'd call physiologic models of clearance.  That would be a thing where you'd have some model of the uptake mechanism and then the transport and the metabolism.  That might be *McHale-Smitton or something like that.

            DR. WILLIAMS:  Well, the ones I'm most familiar with sort of group all those sorts of things into a global parameter, intrinsic clearance. 

            DR. SHEINER:  But you're asking for a model for the intrinsic clearance.  Right?  Or for the Q times clearance over Q plus clearance, or something like that.  Because you're not asking for a model of the drug level.

            DR. WILLIAMS:  Right.

            DR. SHEINER:  That's what we usually think of when we say PBPK, models of how do concentrations relate to physiologic processes.  But you already decided that you are looking at a physiologic process called clearance, and I'm not aware of an awful lot of physiologic models, except I guess everybody who thinks about it could come up with things they think would be more, let's say, reciprocally related and things that would be more directly related.  But other than that, I think ‑‑ most of those models are empirical, and even in the middle of a population analysis, there's some little equation that says that clearance is a linear function with an intercept often, which doesn't make any sense, of weight, age, and some measure of renal function or of hepatic function.

            DR. VENITZ:  Sounds like a topic for another meeting.

            DR. WILLIAMS:  Thanks, everyone.

            DR. VENITZ:  Thank you, Gene, and thank you all for hanging in for today's agenda.

            We are adjourning the meeting.  We are reconvening tomorrow at 8:30.

            (Whereupon, at 4:10 p.m., the subcommittee was recessed, to reconvene at 8:30 a.m., Wednesday, April 23, 2003.)