Bayesianhierarchicalmodelsforadaptiverandomizationinbiomarker-drivenstudies:
UmbrellaandplatformtrialsWilliamT.Barry,PhD
NancyandMorrisJohnLurieInvestigatorBiostatisticsandComputationalBiology
Dana-FarberCancerInstitute
Nov9th,2017
Motivationforbiomarker-driventrialsinoncology(inbrief)
• Molecularheterogeneityofcancerisnolongerahypothesis,butknown,measurable,andquantified.
Personalized/precisionmedicine: A fundamentalassumptionisthatusingthegeneticmakeupofthetumorandthegenotypeofthepatientwillenabletargetedtherapeuticstoimproveclinicaloutcomes.
• Increaseddevelopmentoftargetedtherapiesinoncology
• Componentsofmultiplexgenomicscreeningplatformsareconverging increasingoverlapping
Slide2
Motivationforbiomarker-driventrialsinoncology(inbrief)
• Manyinnovativeclinicaltrialsdesignsinoncology.Importanttodistinguishelements:
– Bayesianvsfrequentistanalysisplans
– Comparativevsnon-comparativehypotheses
– Single-stagevs.sequentialvs.continualassessment
– Adaptivevsfixedrandomization.
– Hypotheseswithinoracrossmarker-definedsubgroups
Slide3
Biomarker-drivendesignsIntegralbiomarkers- Testsinherentinthedesignfromtheonsetandmustbeperformedinrealtimefortheconductofthetrial(re:participantflow)
• Singlemarker/treatment– Enrichmentdesigns(e.g.B31/N9831)– Stratifieddesigns(TKIsandPI3KiinBr)
• Multiplemarkers/treatments– BasketandUmbrellatrial(BATTLE)– Platformtrials
• NCI-MATCH• I-SPY2
– Marker-strategydesigns(SHIVA)
Slide4
Herbst et al. Clin Cancer Res 2015;21:1514-1524
Slide5
Zhouetal.(2008)ClinicalTrials5:181-193– Method(butnocode)fullyspecified
Kimetal.(2011)CancerDiscovery1:44-53– Primaryresults
BATTLE:Biomarker-integratedApproachesofTargetedTherapyforLungCancerElimination(PI:Kim)
Slide6
BATTLEtrialdesign:• Hierarchicalmodel
• Bayesian(non-compara-tive)inference.
• Continualassessment
• Adaptiverandomization
BATTLE:Biomarker-integratedApproachesofTargetedTherapyforLungCancerElimination(PI:Kim)
Slide7
BATTLEtrialdesign:• Hierarchicalmodel
• Bayesian(non-compara-tive)inference.
• Continualassessment
• Adaptiverandomization
BATTLE:Biomarker-integratedApproachesofTargetedTherapyforLungCancerElimination(PI:Kim)
Slide8
BATTLEtrialdesign:• Hierarchicalmodel
• Bayesian(non-compara-tive)inference.
• Continualassessment
• Adaptiverandomization
BATTLE:Biomarker-integratedApproachesofTargetedTherapyforLungCancerElimination(PI:Kim)
Kim(2011):Weplannedtorandomlyassignatleasttheinitial80patientsequallytothe4treatments,toallowatleast1patientineachmarkergrouptocompletetreatment,thusprovidingsufficientdatatoestimatethepriorprobabilityof[diseasecontrol]
Slide9
Barryetal.JBS2015:TheuseofBayesianhierarchicalmodelsforadaptiverandomizationinbiomarker-drivenphaseIIstudies
Researchgoals:• EvaluatepropertiesofBATTLE
(PI:Kim),asoneofthefirstumbrellatrials
• Insilicosimulation(Rcodeasappendix)
• ContrastRARandcontinualassessmentversustraditionalSimontwo-stagedesigns
Slide10
Barryetal.JBS2015:TheuseofBayesianhierarchicalmodelsforadaptiverandomizationinbiomarker-drivenphaseIIstudies
Researchgoals:• EvaluatepropertiesofBATTLE
(PI:Kim),asoneofthefirstumbrellatrials
• Insilicosimulation(Rcodeasappendix)
• ContrastRARandcontinualassessmentversustraditionalSimontwo-stagedesigns
Assigned ineffective tx
Assigned effective tx
Slide11
Barryetal.JBS2015:TheuseofBayesianhierarchicalmodelsforadaptiverandomizationinbiomarker-drivenphaseIIstudies
Researchgoals:• EvaluatepropertiesofBATTLE
(PI:Kim),asoneofthefirstumbrellatrials
• Insilicosimulation(Rcodeasappendix)
• ContrastRARandcontinualassessmentversustraditionalSimontwo-stagedesigns
• Conclusions:• (Nearly)equalefficiency• LessvariabilityinE[N]
Slide12
BATTLE:Biomarker-integratedApproachesofTargetedTherapyforLungCancerElimination(PI:Kim)
LessonslearnedfromBATTLE:• Challengetomakereliableassumptions
aboutprevalenceofbiomarkers
Exp.
Obs.
n≥ 1 per group n≥ 4 per group
Slide13
NCI-MATCH:MolecularAnalysisforTherapyChoice
Schema of patient flow
StatisticalDesign:•1° Endpoint:• Obj resp (RECIST1.1)
• Null:5%• Target:25%
• Single-stagetest• Enroll35ptsperarm(N=31eval)• 5ormoreresp.• a =0.018• b =0.083
Protocolallowsforexpansioncohorts;notstatisticallydriven
StudyHistoryAug2015 Activatedwith 10initialdrugarmsandtargetN=3000Nov 2015 Suspendedenrollmentforplannedevaluation
795ptsregistered(739w/samplessubmitted)645ptscompletedscreening56ptswithamatchingmutation(8.7%)33ptseligibleandenrolled(5.1%)16ptsreceivedTx (2.5%)
Feb2016 Re-activatedwithaddendum#2Expandedeligibility tomyelomaIncreased toN=5000Increasedtototalof24treatmentarmsRevisedestimatewas23%ofptsmatch
Jun2017 Reached(revised) targetofN=6000pts19of26treatmentarmsstillseekingpatientsEnrollmenttosub-studiestocontinuethroughothermech’s
Slide14
NCI-MATCH:MolecularAnalysisforTherapyChoice
http://ecog-acrin.org/nci-match-eay131
Slide15
NCI-MATCH:MolecularAnalysisforTherapyChoice
Snapshotofstudystatus(Nov2016)
IncreasedTargetN:6000pts
24genealt’nsbeingtargeted
Arm/Target Expected #Patients
Arm/Target Expected#Patients
IPIK3CAmut 137 RBRAFnonV600 29WFGFR1/2/3 124 HBRAFV600 26PPTENloss 79 TSMO/PTCH1 18Z1ANRASmut 70 UNF2loss 17S1NF1mut 66 C1METamp 14Z1D dMMR 63 AEGFR mut 8NPTEN mut 62 GROS1transloc 8QERBB2amp 59 S2GNAQ/GNA11 3BERBB2mut 39 EEGFRT790M 1C2METex14sk 37 FALKtransloc 1Z1BCCND1amp 36 XDDR2mut 0YAKT1mut 32 VcKITmut 0
http://ecog-acrin.org/nci-match-eay131
Slide16
NCI-MATCH:MolecularAnalysisforTherapyChoice
OngoingworkbyRSapigao:• Insilicosimulationofthe
dynamicaspectofaddingarmstoNCI-MATCHovertimeandreplacingcompletedarms
• Explorethepropertiesoftwo-andthree-stagedesignsinthisframework
• Add(simulated)responsesandassessBayesianmethodsforcontinualassessment.
Num
ber o
f Arm
s to
reac
h an
alys
is
04
812
1620
24
Single (N = 31)
Minimax (N1 = 21) (N2 = 31)
Optimal (N1 = 16) (N2 = 42)
3-stage (N1 = 11) (N2 = 24) (N3 = 43)
Slide17
BATTLE:Biomarker-integratedApproachesofTargetedTherapyforLungCancerElimination(PI:Kim)
LessonslearnedfromBATTLE:• Challengetomakereliableassumptions
aboutprevalenceofbiomarkers
• Adaptingw/smallnjk
Barryetal.JBS2015:TheuseofBayesianhierarchicalmodelsforadaptiverandomizationinbiomarker-drivenphaseIIstudies
Slide18
BATTLE:Biomarker-integratedApproachesofTargetedTherapyforLungCancerElimination(PI:Kim)
LessonslearnedfromBATTLE:• Challengetomakereliableassumptions
aboutprevalenceofbiomarkers
• Adaptingw/smallnjk
Barryetal.JBS2015:TheuseofBayesianhierarchicalmodelsforadaptiverandomizationinbiomarker-drivenphaseIIstudies
Slide19
BATTLE:Biomarker-integratedApproachesofTargetedTherapyforLungCancerElimination(PI:Kim)
LessonslearnedfromBATTLE:• Challengetomakereliableassumptions
aboutprevalenceofbiomarkers
• Adaptingw/smallnjk• Inferencew/smallnjk
Slide20
I-SPY2:Umbrella/platform(andadaptive)
ImagescourtesyofDr Rugo
Studydesign• RandomizedphaseII• Comparetoconcurrentcontrolarm(T→AC)
• 1° endpoint:pathCR• Integralbiomarkers
• HER2• HR• Mammoprint
• Bayesiananalysisplan(nextslide)
• Intendedtoallowupto4experimentalarms.
LogisticmodelforpCR
Thresholdfor‘graduation’ofaregimenafter60pts.Evidence(bypCR)thatafutureN=300phaseIIIstudywouldbepositiveinanymarker-definedsubgroup:>85%PP
Thresholdforfutilityif<10%PPinallmarker-subgroupsafter20pts.
Note:functionoftwoparameters,pe andpc
ARisproportionaltotheposteriorprob.agiventxissuperior.Priors(appeartobe)fullyspecified;dependonI-SPY1
Slide21
I-SPY2:Neoadjuvant andPersonalizedAdaptiveNovelAgentstoTreatBreastCancer
Barkeretal.Clin Pharmacol Ther.2009;86:97– 100.
BATTLEtrialdesign:• Hierarchicalmodel
• Bayesian(comparative)inference.
• Continualassessment
• Adaptiverandomization
Rugo HSetal.NEngl JMed2016;375:23-34.
Slide22
I-SPY2:Neoadjuvant andPersonalizedAdaptiveNovelAgentstoTreatBreastCancer
StudyHistory NCT00409968
Mar2010 Activatedwith 3initialexperimentalarms:Figitumumab,Neratinib,Veliparib +Carboplatin
Dec 2013 Resultsonthefirstregimento‘graduate’(Veliparib +Carboplatin)werereportedatSABCS byRugo etal.
Rugo HSetal.NEngl JMed2016;375:23-34Apr2014 Resultsonthe2nd regimento‘graduate’(Neratinib) were
reportedatAACR
ParkJWetal.NEngl JMed2016;375:11-22Jun2015 Resultsfora3rd regimento‘graduate’,MK-2206[AKTi],were
reportedatASCO
Jun2017 Resultsfora4th regimento‘graduate’,Pembrolizumab,werereportedatASCO
Slide23
I-SPY2:Neoadjuvant andPersonalizedAdaptiveNovelAgentstoTreatBreastCancer
2010(TargetN =800)
2012 2014 2016(TargetN=1920)
Neratinib Ganitumab +Metformin
AMG386+Trastuzumab
PLX3397
Veliparib +Carboplatin
MK-2206+/-Trastuzumab
T-DM1andPertuzumab
Pembrolizumab
Figitumumab(droppedby2012)
Pertuzumab andTrastuzumab
Talazoparib +Irinotecan
+AMG386 Ganetespib Patritumab +/-Trastuzumab
+Conatumumab(droppedby2012)
• Nonegativearmshavebeenpublished(riskofreportingbias)• Asanongoingstudy,totalstudy-statushasneverbeen
publicallydisseminated(tomyknowledge)• Partialinformationcanbegleanedfromclinicaltrials.gov
Slide24
I-SPY2:Neoadjuvant andPersonalizedAdaptiveNovelAgentstoTreatBreastCancer
Rugo HSetal.NEngl JMed2016;375:23-34.
CONSORT:Veliparib/carboplatin
Slide25
I-SPY2:Neoadjuvant andPersonalizedAdaptiveNovelAgentstoTreatBreastCancer
Rugo HSetal.NEngl JMed2016;375:23-34.
Results:Veliparib/carboplatin
“…We do not report the raw data within biomarker subtypes or signatures; our analysis carries greater precision than would a raw-data estimate”
Veliparib +Carboplatin
Control(T→ AC)
Enrolled N=72 N=44TNsubset N=39 N=19pCR ?? ??No pCR ?? ??
Veliparib +Carboplatin
Control(T→ AC)
Enrolled N=72 N=44TNsubset N=39 N=19pCR 20 5No pCR 19 14
*imputedundersimplifiedassumptions
• MotivationandgeneralapproachofI-SPY2werepublishedwiththelaunchofthetrial(Barkeretal.2009)Insufficientdetailstoevaluatethespecificadaptivedesign.
• ConsistentwithICMJEpolicy,theprotocolwasprovidedassupplementalmaterialtotheNEJMarticles.
• Withmultipleappendices,thestatisticalmethodsappeartobespecifiedbutwillbeextremelychallengingtoreproduce.Priorsrequirepatient-leveldatafromI-SPY1.
• Softwarehasnotbeenmadepublic
• Thedecisiontoredactrawdatafrompublicationsisconcerning
• Unknownwhatthedisseminationplanswillbefornegativearms
Slide26
Commentsontransparency
Slide27
Ventz etal.Biometrics2017:Bayesianresponse-adaptivedesignsforbaskettrials
Researchgoals:• Developnovelmethodstobuild
offgenomicplatforms(Dana-Farber:Oncopanel)
• ApplyRARdesigns(e.g.I-SPY2)to‘basket’trial(NCI-MATCH)
• Constructhierarchicalmodelforadaptiveallocationandcontinualassessment
• Useinsilicosimulationtotuneandevaluateproperties
• ProvidedRpackage(s)formodelsandsimulation.http://bcb.dfci.harvard.edu/~steffen/software.html
Closingremarks
Slide28
• Theuseofmasterprotocols(whetherumbrella,basket,orplatformdesigns)willcontinuetogrowfortrialswithinandacrosstraditionaldiseasetypes.
• Choiceoftrialdesigndependsonmanyparameters:
– Distributionofclinicaloutcomes,andhypothesizedtreatmenteffects
– Markerprevalence,preliminaryevidenceabiomarkerispredictive/prognostic,feasibilityofreal-timeassessment, andoperationalresources.
• Adaptivedesignsgiveflexibility,butalways atsomecost;anditmaybehardtoascertain utility
– Response-adaptiverandomizationwillbecontroversialamongstatisticians.
– Adaptiveenrichmentdesignshavethepotentialtoachievegoalsofpopulation-findingwithtargetedtherapies.
• Adaptiveplatformtrialsareforcingustorevisitoldargumentsontransparencyandwaystofacilitatethereproducibleresearch
Acknowledgements
Slide29
DukeUniversityandUNCJoeIbrahimChuckPerouLisaCareyKellyMarcom
Dana-FarberCancerInstituteSteffenVentzLorenzoTrippaGiovanniParmigianiRosemarieSapigaoMeredithReganRichardGelber
NCTNbiostatistician:DonBerryMaryRedmanBobGrey