A PREDICTIVE MODELING OF TACROLIMUS … · v Expica've variables (SNPs) v Target variable...

1
References: C. Damon et al., Predic2ve modeling of Tacrolimus dose requirement based on high-throughput gene2c screening. Am J Transplant. 2016 Sep 6; N. Pallet et al., Kidney transplant recipients carrying the CYP3A4*22 allelic variant have reduced tacrolimus clearance and oMen reach supratherapeu2c tacrolimus concentra2ons. Am J Transplant (2015) Vol.15 (3):800-5; K. Torkkola, Feature Extrac2on by Non-Parametric Mutual Informa2on Maximiza2on. Journal of Machine Learning Research 3 (2003) 1415-1438; A-C.Haury et al. The Influence of Feature Selec2on Methods on Accuracy, Stability and Interpretability of Molecular Signatures. Plos One (2011) Vol.6 Issue 12 A PREDICTIVE MODELING OF TACROLIMUS PHARMACOKINETICS PARAMETERS BASED ON HIGH-THROUGHPUT GENETIC SCREENING APPROACH Nicolas Pallet 1,2,* , Margaux Luck 1,3 , Eric Thervet 1,2 , Cécilia Damon 3 1 Paris Descartes University, Paris, France; 2 Renal Division, Georges Pompidou European Hospital, Paris, France ; 3 Hypercube InsUtute, Paris, France. * Contact: [email protected] Tacrolimus (Tac) v Immunosuppressant for the preven0on of kidney transplant rejec0on v Narrow therapeu0c index v Drug metabolism inter-pa0ents variability v Associa0on between genome and Tac pharmacokine0cs: v High gene0c inter-pa0ents variability v CYP3A5 and CYP3A4 explain (only) a part of the varia0on v Hypotheses: involvement of a wider network of candidate genes Objec?ve and Data v Iden0fy a set of covariant germline polymorphisms predic0ve of the inter-pa0ents Tac pharmacokine0cs variability in kidney transplanta0on v Iden0fica0on of new poten0al candidate genes explaining a part of the Tac metabolism inter-pa0ent variability The Study Gene?c explanatory variables (input variables): DNA of 229 kidney transplant recipients genotyped using a chip with 16, 561 SNPs markers 16, 561 SNPs 229 subjects based on: v A variable selec0on scheme v A predic0ve model Mul?variate predic?ve model f(x) ≈ y Log(Tac C 0 /Dose) (229 subjects) Outcome (target variable): Tac blood concentra0on (Tac C 0 ) and dose monitored at each fellow-up 0me aSer transplanta0on (days 10, 14, 30, 60 and 90) d10 d14 d30 d60 d90 Supervised Analysis Ensemble variable selec4on scheme and mul4variate predic4ve modeling DATASET v Expica've variables (SNPs) v Target variable (log(C0/Dose) Train set (80 %) Test set (20 %) Dataset split preserving the distribu'on of the target variable EVALUATION v Performance: R 2 Score v Significance: Permuta4on test (1000 runs) ENSEMBLE VARIABLE SELECTION SCHEME 2 techniques: v Fisher’s test: p t p-value threshold v Mutual informa'on: p MI score threshold Op'miza'on (10 fold CV) v Intersec'on of the 2 subsets of variables based on the best couples of thresholds p * =(p t ,p MI ) * assessed with 1-std error rule on R 2 score PREDICTION v Op'mal dimension reduc'on v Explanatory predic've models M * with Par'al Least Squares: PLS1 at each follow-up 'me and PLS2 over all 'mes 10 folds cross-valida4on 10X Mul4variate predic4ve model M * M * (x) ≈ y Results and Discussion Conclusion and perspec2ves Performances and models v SNP networks explain 30 to 70 % of the inter-pa9ent variability of Tac metabolism, depending on the model generated and the 9me aCer transplanta9on v Genes interac9on networks related to oxidoreductase func9ons and mono-oxygenase ac9vity, including CYP3A4 and CYP3A5, have a major impact on Tac metabolism v The mul9drug transporter ABCC8 and the nucleoside carrier SLC28A3 appear to be involved in Tac metabolism v Further studies are needed to validate the clinical u9lity and impact of predic9ons tools, including pharmacogenomics, on transplant outcomes. PLS1 models PLS2 models Time aCer transplanta9on (days) 10 14 30 60 90 10, 14, 30, 60 and 90 days Performance (R 2 value) 0.30 0.27 0.41 0.7 0.62 0.28 Significance (p-value) 0.001 0.001 0.001 0.001 0.001 0.001 Model complexity (nb of SNPs) 5 19 12 44 33 7 PLS, par9al least squares; SNP, single-nucleo9de polymorphism. v All the performances are significant (p-value < 0.03; permuta9on test) v Best performances at day 60 and 90 (i.e. 3 months) with more complex models (Tac dosage more stable at later 9me) v Genes involved in the predic2ve models v Performances v Models include control genes (CYP3A4 and CYP3A5) + drug metabolism gene v Models genes are in molecular interac9on network related to monoxygenase func9ons v SLC28A3 (drug transporter) polymorphism as candidate gene v High weights in the model at day 60 v Higher ini9al Tac adjusted dose for homozygote v No linkage disequilibrium v Validated on an independent valida9on cohort of 189 kidney transplant recipients Distribu2on of the log(Tac C 0 / Dose) values according to 3 genotypes of SLC28A3 at each fellow up 2me and linkage disequilibrium Valida2on cohort Distribu2on of the log(Tac C 0 /Dose) values at 3, 12 and 24 month aNer transplanta2on. P-value calculated using Kruskal-Wallis one- way analysis of variance.

Transcript of A PREDICTIVE MODELING OF TACROLIMUS … · v Expica've variables (SNPs) v Target variable...

References:C.Damonetal.,Predic2vemodelingofTacrolimusdoserequirementbasedonhigh-throughputgene2cscreening.AmJTransplant.2016Sep6;N.Palletetal.,KidneytransplantrecipientscarryingtheCYP3A4*22allelicvarianthavereducedtacrolimusclearanceandoMenreachsupratherapeu2ctacrolimusconcentra2ons.AmJTransplant(2015)Vol.15(3):800-5;K.Torkkola,FeatureExtrac2onbyNon-ParametricMutualInforma2onMaximiza2on.JournalofMachineLearningResearch3(2003)1415-1438;A-C.Hauryetal.TheInfluenceofFeatureSelec2onMethodsonAccuracy,StabilityandInterpretabilityofMolecularSignatures.PlosOne(2011)Vol.6Issue12

APREDICTIVEMODELINGOFTACROLIMUSPHARMACOKINETICSPARAMETERSBASEDONHIGH-THROUGHPUTGENETICSCREENINGAPPROACH

NicolasPallet1,2,*,MargauxLuck1,3,EricThervet1,2,CéciliaDamon31ParisDescartesUniversity,Paris,France;2RenalDivision,GeorgesPompidouEuropeanHospital,Paris,France;3HypercubeInsUtute,Paris,France.*Contact:[email protected]

Tacrolimus(Tac)

v  Immunosuppressantforthepreven0onofkidneytransplantrejec0on

v  Narrowtherapeu0cindex

v  Drugmetabolisminter-pa0entsvariability

v  Associa0onbetweengenomeandTacpharmacokine0cs:

v  Highgene0cinter-pa0entsvariability

v  CYP3A5andCYP3A4explain(only)apartofthevaria0on

v  Hypotheses:involvementofawidernetworkofcandidategenes

Objec?veandData

v  Iden0fyasetofcovariantgermlinepolymorphismspredic0veoftheinter-pa0entsTacpharmacokine0csvariabilityinkidneytransplanta0on

v  Iden0fica0onofnewpoten0alcandidategenesexplainingapartoftheTacmetabolisminter-pa0entvariability

TheStudy

Gene?cexplanatoryvariables(inputvariables):

DNAof229kidneytransplantrecipientsgenotypedusingachip

with16,561SNPsmarkers

16,561SNPs

229subjects

basedon:v  Avariableselec0onscheme

v  Apredic0vemodel

Mul?variatepredic?vemodelf(x)≈y

Log(TacC0/Dose)(229subjects)

Outcome(targetvariable):

Tacbloodconcentra0on(TacC0)anddosemonitoredateachfellow-up

0meaSertransplanta0on(days10,14,30,60and90)

d10 d14 d30 d60 d90

SupervisedAnalysisEnsemblevariableselec4onschemeandmul4variatepredic4vemodeling

DATASET

v  Expica'vevariables(SNPs)v  Targetvariable(log(C0/Dose)

Trainset

(80%)

Testset

(20%)

Datasetsplitpreservingthe

distribu'onofthetargetvariable

EVALUATION

v  Performance:R2Score

v  Significance:Permuta4ontest(1000runs)

ENSEMBLEVARIABLESELECTIONSCHEME

2techniques:

v  Fisher’stest:ptp-valuethreshold

v Mutualinforma'on:pMIscorethreshold

Op'miza'on(10foldCV)

v  Intersec'onofthe2subsetsofvariablesbasedonthebestcouplesofthresholds

p*=(pt,p

MI)*assessedwith1-stderrorruleon

R2score

PREDICTION

v  Op'maldimensionreduc'on

v  Explanatorypredic'vemodels

M*withPar'alLeastSquares:

PLS1ateachfollow-up'meandPLS2overall'mes

10foldscross-valida4on

10X

Mul4variatepredic4vemodelM*

M*(x)≈y

ResultsandDiscussion

Conclusionandperspec2ves

Performancesandmodels

v  SNPnetworksexplain30to70%oftheinter-pa9entvariabilityofTacmetabolism,dependingonthemodelgeneratedandthe9meaCertransplanta9on

v  Genesinterac9onnetworksrelatedtooxidoreductasefunc9onsandmono-oxygenaseac9vity,includingCYP3A4andCYP3A5,haveamajorimpactonTacmetabolism

v  Themul9drugtransporterABCC8andthenucleosidecarrierSLC28A3appeartobeinvolvedinTacmetabolismv  Furtherstudiesareneededtovalidatetheclinicalu9lityandimpactofpredic9onstools,includingpharmacogenomics,ontransplant

outcomes.

PLS1models PLS2modelsTimeaCertransplanta9on(days) 10 14 30 60 90 10,14,30,60and90daysPerformance(R2value) 0.30 0.27 0.41 0.7 0.62 0.28Significance(p-value) 0.001 0.001 0.001 0.001 0.001 0.001Modelcomplexity(nbofSNPs) 5 19 12 44 33 7PLS,par9alleastsquares;SNP,single-nucleo9depolymorphism.

v  All the performances aresignificant (p-value < 0.03;permuta9ontest)

v  Best performances at day 60

and 90 (i.e. 3 months) withmore complex models (Tacdosage more stable at later9me)

v  Genesinvolvedinthepredic2vemodels

v  Performances

v Modelsincludecontrolgenes(CYP3A4andCYP3A5)+drugmetabolismgenev Modelsgenesareinmolecularinterac9onnetworkrelatedtomonoxygenasefunc9onsv  SLC28A3(drugtransporter)polymorphismascandidategene

v  Highweights inthemodelatday60

v  Higher ini9al Tac adjusteddoseforhomozygote

v  Nolinkagedisequilibriumv  Validatedonan independent

valida9on cohort of 189kidneytransplantrecipients

Distribu2onof

thelog(TacC0/

Dose)values

accordingto3

genotypesof

SLC28A3at

eachfellowup

2meand

linkage

disequilibrium

Valida2oncohort

Distribu2onofthelog(TacC0/Dose)valuesat3,12and24month

aNertransplanta2on.P-valuecalculatedusingKruskal-Wallisone-

wayanalysisofvariance.