Clinical Validation of Prognostic Biomarkers of Risk and Predictive Biomarkers of Drug Efficacy or Safety

1
Clinical Validation of Prognostic Biomarkers of
Risk and Predictive Biomarkers of Drug Efficacy
or Safety

• Gene Pennello, Ph.D.
• Team Leader, Diagnostics Devices Branch
• Division of Biostatistics
• Office of Surveillance and Biometrics
• Center for Devices and Radiological Health, FDA

SAMSI Risk Perception Policy Practice
Workshop October 3, 2007
2
Outline

• FDA and Device Regulation
• Types of Biomarkers
• Validation of Diagnostics
• Predictive and Prognostic Biomarkers
• Definitions, Endpoints
• Study Designs for Predictive Biomarkers
• Prospective Designs efficiency comparison
• Prospective-Retrospective Designs
• Summary

3
FDA
CDERDrugs
CDRH,Devices
CBER,Biologics
CVM,Veterinary
CFSAN,Food
NCTR
4
What are Medical Devices?
An item for treating or diagnosing a health
condition whose intended use is not achieved
primarily by chemical or biological action within
the body (Section 201(h) of the Federal Food Drug
Cosmetic (FDC) Act). Definition by
exclusion Simply put, a medical device is any
medical item for use in humans that is not a drug
nor a biological product.
5
Example of Medical Devices

• Relatively Simple Devices tongue depressors
  thermometers latex gloves simple surgical
  instruments
• Ophthalmic devices intraocular lenses PRK
  lasers,
• Radiological devices MRI machines CT
  scannersdigital mammographycomputer aided
  detection

Cardiovascular Devices pacemakers
defibrillators heart valves coronary stents
artificial hearts Monitoring Devices
glucometers bone densitometers Diagnostic
Devices diagnostic test kits for
HIVprostate-specific antigen (PSA) testhuman
papillomavirus (HPV) test
6
Example of Medical Devices
Dental, Ear, Nose, andThroat Devices hearing
aidsbronchoscopy system General, Surgical, and
Restorative Devices breast implants artificial
hips spinal fixation devices artificial skin 
Emerging technologies multiplex genetic tests
(e.g., for multiple mutations or
microbes) Genomic and proteomic Dx
tests Nanotechnological devices Microspheres for
molecular treatment of cancer Robotics Theranostic
s (predictive biomarkers of response or adverse
reaction to therapy). Artificial pancreas
7
Example of Medical Devices
Due to the wide variety in technology,
complexity, and intended use, medical devices
can present novel statistical design and analysis
challenges.
8
Device Regulation

• Decision to approve a PMA application must rely
  upon valid scientific evidence to determine
  whether there is reasonable assurance that the
  device is safe and effective.
• Valid scientific evidence is evidence from well
  controlled studies, partially controlled studies
  and objective trials without matched controls,
  well documented case histories conducted by
  qualified experts that there is a reasonable
  assurance of safety and effectiveness . . .
• U.S. Code of Federal Regulations, Title 21 (Food
  and Drugs), U.S. Government Printing Office,
  Washington DC, 2001, Part 860.7 Web address
  http//www.access.gpo.gov/nara/cfr/waisidx_01/21cf
  r860_01.html (Accessed February, 2002)

9
Device Regulation

• Least Burdensome Provisions of FDA Modernization
  Act (1997)
• Secretary shall only request information that is
  necessary to making substantial equivalence
  determinations.
• Secretary shall consider, , the least
  burdensome appropriate means of evaluating device
  effectiveness that would have a reasonable
  likelihood of resulting in approval.
• U.S. Code of Federal Regulations, Title 21 (Food
  and Drugs), U.S. Government Printing Office,
  Washington DC, 2001, Part 513(i)(1)(D) and
  513(a)(3)(D)(ii). Web address http//www.access.g
  po.gov/nara/cfr/waisidx_01/21cfr860_01.html

10
FDA Least Burdensome Guidance

• FDA Guidance The Least Burdensome Provisions of
  the FDA Modernization Act of 1997 Concept and
  Principles (2002)
• Modern statistical methods may also play an
  important role in achieving a least burdensome
  path to market. For example, through the use of
  Baysian sic analyses, studies can be combined
  in order to help reduce the sample size needed
  for the experimental and/or control device.

11
Examples of Less Burdensome

• Non-U.S. data
• Surrogate endpoints (e.g., acute follow-up)
• Interim analysis, Adaptive design
• Bayesian methods (e.g., to reduce sample size)
• Propensity Scores for historical controls
• Sensitivity analysis for missing data.Note,
  could trade clinical for statistical burden
• FDA Draft Guidance for the Use of Bayesian
  Statistics in Medical Device (released May 23,
  2006) www.fda.gov/cdrh/osb/guidance/1601.html

12
Least Burdensome Provision

• Least burdensome provision in FDAMA of 1997 is
  directed to both medical devices and diagnostics
  (including biomarkers).

13
Device Risk Classification

• Class I Devices for which general controls
  provide reasonable assurance of the safety and
  effectiveness.
• Class II General controls insufficient, Can
  establish special controls (performance
  standards CLIA, ISO, FDA guidance. May require
  clinical data on a 510(k).
• Class III General and special controls
  insufficient. Life-sustaining/supporting,
  substantial importance in preventing impairment
  of human health, potential unreasonable risk of
  illness or injury. Needs pre-market approval
  (PMA).

14
Post-Market Transformation

• Make postmarket data more widely available to
  Center staff and supplement search and reporting
  tools
• "Investigate the use of data and text mining
  techniques to identify the "needles in the
  haystack" by identifying patterns in the incoming
  data that equate to public health signals.
• Example is WebVDME Bayesian data-mining
• Design a pilot project to test the usefulness of
  quantitative decision-making methods for medical
  device regulation across the total product life
  cycle

http//www.fda.gov/cdrh/postmarket/mdpi-report-110
6.html
15
Types of Biomarkers

• Diagnostic
• Early detection (screening), enabling
  intervention at an earlier and potentially more
  curable stage than under usual clinical
  diagnostic conditions
• Monitoring of disease response during therapy,
  with potential for adjusting level of
  intervention (e.g. dose) on a dynamic and
  personal basis
• Risk assessment leading to preventive
  interventions for those at sufficient risk
• Prognosis, allowing for more aggressive therapy
  for patients with poorer prognosis
• Prediction of safety or efficacy (response) of a
  therapy, thereby providing guidance in choice of
  therapy

16
Types of Biomarkers

• Diagnostic
• Early Detection (screening)
• Monitoring
• Risk Assessment
• Prognostic
• Predictive of Safety or Efficacy
• The first three are considered together, where
  the focus is on identifying the disease or
  condition.

17
Types of Biomarkers

• Diagnostic
• Early Detection (screening)
• Monitoring
• Risk Assessment
• Prognostic
• Predictive of Safety or EfficacyThe last three
  are attempting to predict the future.

18
Analytical Validation

• How well are you measuring the measurand?
• Precision / Reproducibility
• Method Comparison
• LoB, LoD, LoQ
• Linearity
• Stability
• Clinical Laboratory Standards Institute (CLSI)
• (http//www.nccls.org/)

19
Clinical Validation (Qualification)

• Does the test have clinical utility?
• Does it have added value over standard tests
  (e.g, clinical covariates like age, tumor size,
  stage)?
• May or may not require a clinical study
• EX. Roche Amplichip

CDRH guidance document Statistical Guidance on
Reporting Results from Studies Evaluating
Diagnostic Tests issued in final form in March,
2007, concerns reporting agreement when there is
no perfect standard and also discrepancy
resolution. http//www.fda.gov/cdrh/osb/guidance/
1620.html
20
Roche AmpliChip CYP450 Test (CDRH de novo 510(k)
K042259)

• Genotypes two cytochrome P450 genes (29
  polymorphisms in CYP2D6 gene, 2 in CYP2C19) to
  provide the predictive phenotype of the metabolic
  rate for a class of therapeutics metabolized
  primarily by CYP2D6 or CYP2C19 gene products.
  The phenotypes are
• (1) Poor metabolizers (3) Extensive
  metabolizers
• (2) Intermediate metabolizers (4) Ultrarapid
  metabolizers
• Cytochrome P450s are a large multi-gene family of
  enzymes found in the liver, and are linked to the
  metabolism of approximately 70-80 of all drugs.
  Among them, the polymorphic CYP2D6 and CYP2C19
  genes are responsible for approximately 25 of
  all CYP450-mediated drug metabolism. A
  polymorphism in these enzymes can lead to an
  excessive or prolonged therapeutic effect or
  drug-related toxicity after a typical dose by
  failing to clear a drug from the blood or by
  changing the pattern of metabolism to produce
  toxic metabolites.

http//www.accessdata.fda.gov/scripts/cdrh/cfdocs/
cfPMN/pmn.cfm
21
Adding Value to Standard Clinical Predictors

• Head to Head Marker superior to clinical
  predictors at predicting outcome.
• Incremental Improvement Combination superior to
  clinical predictors alone.
• Marker Predictive within Clinical Strata e.g.,
  HR(, ) significant within age, tumor grade,
  tumor size groups.

22
Multivariate Index Assays

• An IVDMIA is a device that
• Combines the values of multiple variables using
  an interpretation function to yield a single,
  patient-specific result (e.g., a
  classification, score, index, etc.), that
  is intended for use in the diagnosis of disease
  or other conditions, or in the cure, mitigation,
  treatment or prevention of disease, and
• Provides a result whose derivation is
  non-transparent and cannot be independently
  derived or verified by the end user. MIA result
  could be a binary (dichotomous) (such as yes or
  no), categorical (such as disease type), ordinal
  (such as low, medium, high) or a continuous
  scale.
• Source FDA MIA Draft Guidance
• http//www.fda.gov/cdrh/oivd/guidance/1610.html

23
Typical Endpoints for Prognostic or Predictive
Biomarkers

• Time to Event
• Event by Time t

Treatment Median Survival Time
A 6 months
B 12 months
Hazard Ratio 0.5
Treatment R Not R Response Rate
A 30 30 0.50 (30/60)
B 10 50 0.13 (10/60)
24
Relative Risk vs. Diagnostic Accuracy
Event by Time t
RelativeRisk 3.0 (30/60)/(10/60)
Se 0.75 (30/40)
Sp 0.63 (50/80)
PPV 0.50 (30/60)
NPV 0.83 (50/60)
E No E
30 30 60
10 50 60
40 80 120
Marker

• Relative Risk looks good, but Dx accuracy not
  great ? limited clinical utility?

Example taken from Emir, Wieand, Su, Cha,
Analysis of repeated markers used to predict
progression of cancer Statist. Med., 17, 2563-78,
1998.
25
Hazard Ratio vs. Diagnostic Accuracy

• NCCTG Mayo Clinic Study. CA15-3 ratio as
  diagnostic for progression of breast cancer (as
  determined by physical exam).

Hazard Ratio 2.3 (p 0.0002)
Se 0.30 (0.17,0.43)
Sp 0.82 (0.74,0.89)
PPV 0.27 (0.21,0.33)
Example taken from Emir, Wieand, Su, Cha,
Analysis of repeated markers used to predict
progression of cancer Statist. Med., 17, 2563-78,
1998.
26
Diagnostic Performance
Sensitivity Specificity (TP rate) (TN
rate) FP rate fraction of fraction
of fraction of responders non-responders
non-responders who test who test who test
Test is useful if TP rate gt FP rate, i.e.,
sensitivity specificity gt 1. EX. Useless
test sensitivity 0.80, specificity 0.20
27
Diagnostic Performance
Positive Negative predictive predictive
value (PPV) value (NPV) 1 NPVfraction
of fraction of fraction of test s
who test s who test s whorespond dont
respond respondTest is useful if PPV NPV gt
1 EX. Useless test PPV 0.60, NPV 0.40
28
d
A ROC curve is a plot of sensitivity (true
positive rate) vs. 1-specificity (false positive
rate) over all possible cutoff points for the
test. The test is informative if the area under
the curve is greater than 0.5.
29
Prognostic Biomarker (Strong Defn)

• Prognostic factor. Informs about an outcome
  independent of specific treatment (ability of
  tumor to proliferate, invade, and/or spread).
• Prognostic biomarker is associated with
  likelihood of an outcome (e.g., survival,
  response, recurrence) such that magnitude of
  association is independent of treatment.
• On some scale, treatment and biomarker effects
  are additive, that is, do not interact.

30
HR(A,B)0.67
HR(A,B)0.67
31
(No Transcript)
32
Prognostic Biomarker (Weak Defn)

• Prognostic factor. Informs about an outcome
  independent of specific treatment (ability of
  tumor to proliferate, invade, and/or spread).
• Prognostic biomarker is associated with
  likelihood of an outcome (e.g., survival,
  response, recurrence) in a population that is
  untreated or on a standard (non-targeted)
  treatment.
• If population is clearly defined, than can use
  to choose more or less aggressive therapy, but
  not specific therapies, per se.

33
HR(A,B)0.67
HR(A,B)0.67
34
Prognostic Biomarker

• Her2-neu for node-negative women with breast
  cancer prognostic for recurrence
• Breast cancer prognostic test based on microarray
  gene expression of RNAs extracted from breast
  tumor tissue to assess a patients risk for
  distant metastasis for women less than 61 with
  Stage I or II disease with tumor size less than
  or equal 5.0 cm and who are lymph node negative.
• (Ref. Buyse et al. JNCI 98, 1183-1192)

35
Agendia Mammaprint Gene Signature for Time to
Distant Metastasis (N302)
5-year Low risk group 0.95 (0.91-0.99) High
risk group 0.78 (0.72-0.84) 10-year Low risk
group 0.90 (0.85-0.96) High risk group0.71
(0.65-0.78) Buyse et al JNCI (2006),
98,1183-1192
36
Proportion alive at 10 years

• Clinical Gene N Proportion
• Signature
• Low Risk Low Risk 52 0.88 (0.74 to 0.95) Sp
• Low Risk High Risk 28 0.69 (0.45 to 0.84) 1Se
• High Risk Low Risk 59 0.89 (0.77 to 0.95) Sp
• High Risk High Risk 163 0.69 (0.61 to 0.76) 1Se
• Buyse et al JNCI 2006

37
Predictive Biomarker

• Predictive factor. Implies relative sensitivity
  or resistance to specific treatments or agents.
• Predictive biomarker predicts differential effect
  of treatment on outcome.
• Treatment and biomarker interact.Predictive
  biomarker can be useful for selecting specific
  therapy.

38
HR(A,B)0.5
HR(A,B)1.0
39
Predictive Biomarker of Efficacy

• Marker HER2/neuTreatment Trastuzumab
  (Herceptin)
• Objective response rate
• HerceptinChemo ChemoFISH 95/176 (54)
  51/168 (30)FISH- 19/50 (38) 22/57
  (39)
• Arch. Pathol. Lab Med Jan 2007 (ASCO/CAP
  Guidelines)

40
Predictive Biomarkers for Safety

• Predict risk of an adverse event dependent on the
  biomarker
• Example
• UGT1A1, cleared by FDA, to predict the risk of
  neutropenia in patients taking irinotecan for
  colorectal cancer

41
Prospective Study Designs for Predictive Markers

• Untargeted Design (Reference)
• Validate Treatment, Marker Simultaneously
• Marker by Treatment Design
• Targeted Design (Marker Subset Only)
• Marker Strategy Design
• Historical Control

42
Untargeted Design (Reference)

• Test if drug works in entire population.
• Mixture of marker and drug effects.
• Can store samples if test is not ready.

43
Marker by Treatment (Interaction) Design

• A Randomized Block Design
• Can test for biomarker by treatment interaction
  (predictive biomarker)
• Test needs to be available before trial ensues.

44
Marker by Treatment Design Questions

• Test Drug Overall and within Marker Subset
• 0.04, 0.01 tests suggested to control Type I
  error rate at 0.05 (Simon), but subset could
  drive overall result.
• Frequentist multiplicity penalty may preclude
  subset testing as good business strategy.
• Statement about drug, not biomarker
• Test Marker Overall and within Drug Subset
• Statement about marker, not drug.
• Test for Treatment by Marker Interaction
• Simultaneously validates drug and marker.

45
Targeted Design

• Test if drug works in subset.
• Cannot test if marker discriminates. Only PPV
  available.

46
Efficiency of Designs
Relative Efficiency Relative Efficiency
Marker Prevalence Relative Efficacy Targeted Design Interaction Design
25 0 16x 8x
50 0 4x 2x
75 0 1.8x 0.9x

• Efficiency gain depends on marker prevalence,
  relative efficacy, and difference tested.

Marker to Marker Patients Simon
Maitournam, CCR 2004 Marker by Treatment
Design Test for Interaction approx. efficiency
enriching with half s, half s.
47
Efficiency of Designs
Relative Efficiency Relative Efficiency
Marker Prevalence Relative Efficacy Targeted Design Interaction Design
25 25 5.2x 1.5x
50 25 2.6x 0.7x
75 25 1.5x 0.4x

• Efficiency gain depends on marker prevalence,
  relative efficacy, and difference tested.

Marker to Marker Patients Simon
Maitournam, CCR 2004 Marker by Treatment
Design Test for Interaction approx. efficiency
enriching with half s, half s.
48
Efficiency of Designs
Relative Efficiency Relative Efficiency
Marker Prevalence Relative Efficacy Targeted Design Interaction Design
25 50 2.5x 0.3x
50 50 1.8x 0.2x
75 50 1.3x 0.1x

• Efficiency gain depends on marker prevalence,
  relative efficacy, and difference tested.

Marker to Marker Patients Simon
Maitournam, CCR 2004 Marker by Treatment
Design Test for Interaction approx. efficiency
when enriching with half s, half s.
49
Improving Efficiency of Interaction Design

• Enrich with Test Positives if Pr() is low
• Find scale such that marker and treatment effects
  are additive
• Adaptive Randomization
• Bayesian subset analysis
• If reader variability (e.g., IHC), then use
  multiple readers.
• Prior Information

50
Possibilities for Increasing Efficiency of
Interaction Design

• Enrich with Test Positives if Pr() is low
• Estimates of Sensitivity and Specificity are
  biased because they depend on Pr().
• Use inverse probability weighting (Horvitz,
  Thompson, 1952) or Bayes Theorem (Begg, Greenes,
  1983) to obtain unbiased estimates.

51
A Marker-Based Strategy

• Pro More ethical, perhaps. More patients given
  experimental drug. Test utility based on PPVE,
  NPVE.
• Con Cannot assess test-treatment interaction.

52
Marker-Based Strategy
Response
E Naïve E Unbd
Se a / (ac) a / (a2c)
Sp d / (db) 2d / (2db)
PPV a / (ab) same
NPV d / (cd) same
R Not R
E a b
P 0 0

Test
R Not R
E c d
P e f

Test
53
A Marker-Based Strategy
Response
E Naïve E Unbd
Se 20/43(0.47) 20/66 (0.30)
Sp 157/177(0.89) 314/334(0.94)
PPV 20/40(0.50) Same
NPV 157/180(0.88) Same
R Not R
E 20 20 40
P 0 0 0
20 20 40
Test
R Not R
E 23 157 180
P 24 156 180
46 314 360
Test
54
Possibilities for Increasing Efficiency of
Interaction Design

• Transformation
• Find a transformation (Box-Cox?) of outcome that
  makes treatment and effects additive.
• Can then pool marker effect estimates within
  treatments A and B.
• Can also pool drug effect estimates within marker
  and marker s.

55
Possibilities for Increasing Efficiency of
Interaction Design

• Adaptive Randomization
• Adapt randomization ratio to treatment A and B
  within biomarker subsets to maximize (a) power,
  or (b) fraction of patients on better treatment
• If response rate lt 0.5 for both treatments, then
  (a) and (b) are compatible, otherwise in tension.
• Pr() disturbed, so need to adjust Se, Sp

56
Possibilities for Increasing Efficiency of
Interaction Design

• Bayesian subset analysis (cf. Dixon, Simon)
• Subsets modeled as exchangeable via random
  effects.
• Subset estimate borrows strength from complement
  subset, increasing precision of estimate.
• However, interaction estimate more conservative
  relative to usual non-Bayesian analysis.

57
Bayesian Subset Analysis

• Power is enhanced to show drug works in marker
  subset (blue).
• Power is enhanced to show marker works
  (discriminates) in patients taking drug (red)

58
Possibilities for Increasing Efficiency of
Interaction Design

• Use Multiple Readers
• EGFR IHC test (Dako) and Cetuximab and
  Panitumumab (Amgen) for Colorectal Cancer. of
  cells stained and maximum staining intensity
  subject to reader variability
• Use multiple readers, account for random reader
  effects.
• Multiple Reader, Multiple Case Designs (MRMC) are
  used for digital mammography systems and computed
  aided detection (CAD) systems
• Analysis can be difficult.

59
Possibilities for Increasing Efficiency of
Interaction Design

• Prior Information (Bayesian analysis)
• Borrow strength from previous study regarded as
  exchangeable with current study.

60
Marker Based Strategy Design
Marker Level (-)
Treatment A
Marker Based Strategy
Marker Level ()
Treatment B
Register
Randomize
Test Marker
Non Marker Based Strategy
Treatment A
Sargent et al., JCO 2005
61
Marker Based Strategy Design
Marker Level (-)
Treatment A
Marker Based Strategy
Marker Level ()
Treatment B
Register
Randomize
Test Marker
Treatment A
Non Marker Based Strategy
Randomize
Treatment B
Sargent et al., JCO 2005
62
Marker Based Strategy Design

• Lacks power Differential effect comparison
  diluted because some patients in non-marker-based
  strategy arm get marker-based treatment (could
  eliminate these to increase power).
• Might be best suited if have gt 2 treatments or gt
  2 markers
• EX. Irinotecan regiment (dose, timing, frequency)
  determined by UGT1A1 genotype (6/6, 6/7, or 7/7)
  in colorectal cancer patients.

63
Marker Based Strategy

• If no gold standard, then can be only way to
  assess effectiveness of a test.
• EX. Detection tumor of origin in cancers of
  unknown primary.
• No gold standard IHC, imaging, may fail to
  identify TOO.
• Randomize patients to be managed with
• new test standard, or
• with standard alone
• Compare arms on survival

64
Targeted Design w. Historical Control

• Drug already on market, but has poor response
  rate.
• If response rate in marker study is
  significantly greater than historical rate, then
  marker discriminates.
• Limitations
• Lacks power because effect diluted.
• Need to calibrate historical rate to marker
  study (adjust for covariates).

65
Prospective-Retrospective Designs

• Prospectively apply marker to stored samples (in
  retrospect).
• Can test overall, w. subset, or for interaction.
• Missing samples could introduce bias.
• RCT samples. Randomization ensures case and
  control samples have similar characteristics.
• Case-control samples. Avoid selection bias by
  matching on sample processing date, processing
  sites, etc., and not excluding censored times.
• Reserve samples only for analytically validated
  markers that are biologically plausible.

66
The Challenge of Multiplicity

• Multiplicity of classifiers
• Microarrays and proteomics
• Many predictive models could be built with so
  many inputs.
• The challenge is to confirm any such model with
  an independent data set.
• A caveat the independent test data set cannot
  be continually reused. Great discipline is
  required in this regard.

67
Cross-Validation Pitfall
Simon, Radmacher, Dobbin, McShane (2003),
Pitfalls in the Use of DNA Microarray Data for
Diagnostic and Prognostic Classification, JNCI,
95 (1)
68
Summary Remarks

• How to assess a test or biomarker is well-known,
  but not as well-known in therapeutic circles.
• Need to assess whether the biomarker adds
  anything to what we already know.
• The number of possibly good biomarker candidates
  is enormous but great care is needed in
  restricting the search.

69
Summary Remarks

• Need to encourage least burdensome approaches to
  validating biomarkers without compromising level
  of evidence
• Essential to confirm marker in independent
  dataset
• Studies to demonstrate informativeness of a
  biomarker can be quite difficult to design,
  conduct and analyze.

70
Acknowledgements

• CDRH Division of Biostatistics (DBS)
• Greg Campbell, Division Director
• Diagnostic Devices Branch (DDB)
• Lakshmi Vishnuvajjala, Branch Chief
• Estelle Russek-Cohen, Team Leader
• Gene Pennello, Team Leader

Bipasa Biswas Kyungsook Kim, Harry Bushar Samir
Lababidi Arkendra De Kristen Meier Shanti
Gomatam Kyunghee Song Thomas Gwise Rong Tang
71
More References

• Sargent et al (2005). Clinical trial designs for
  predictive marker validation in cancer treatment
  trials. J Clin Oncol 232020-2027.
• Pennello Vishnuvajjala (2005). Statistical
  design and analysis issues with pharmacogenomic
  drug-diagnostic co-development, In American Stat.
  Assoc. 2005 Proc. of the Biopharm. Section, Joint
  Statistical Meetings, Minneapolis, MN, August,
  2005 American Stat. Assoc. Alexandria, VA.
• FDA Drug-Diagnostic Co-Development Concept Paper.
  April 2005.http//www.fda.gov/cder/genomics/pharm
  acoconceptfn.pdf

72
(No Transcript)