CITELINESMARTSOLUTIONS

CITELINECLINICAL

WhitePaper

AIinPharma:

Benefits,Risks,andtheRoadAhead

September2024

AIinPharma:Benefits,Risks,andtheRoadAhead

Introduction

Useofartificialintelligence(AI)nowpermeateseveryindustry,andpharmaisnoexception.

Whethermachinelearning(ML)modelsthat

makepredictionsbasedonexistingdataor

generativeAI(GenAI)modelsthatcreatenew

databasedonthedatatheyweretrainedon,AIisbeingusedtostreamlineandaccelerateeachstepofthedrugdevelopmentprocessfrom

researchthroughapprovalandmarketing.

AccordingtoMcKinsey&Co.,generativeAI

alonecouldproduce$60billionto$110billionayearineconomicvalueacrossthepharmaindustryvaluechain.And$13billionto

$25billionofthatannualvaluealonewouldbeforclinicaldevelopment

.1

AIisabletohandlebothstructuredand

unstructureddata,includingmultimodaldatasuchastabular,text,images,andvideos.At

itsmostbasiclevel,AIcanautomatemundanetaskssuchasstructureddocumentandimageanalyses,enablingexpertstospendmore

timeontasksthatrequiretheirattentionandproficiency.

Onadeeperlevel,AIcanunveilinsightsfrom

historicaldatatoinformoperationsand

providealensintothefuturethroughpredictiveanalytics,supplementingtraditionaldescriptiveanddiagnosticanalyticsthatsolelyprovide

analyticalinformationanchoredonhistorical

patterns.Itcanalsoenableandaccelerate

expertisethroughprescriptiveanalytics,

advisingexpertsonthenextbestactiontotaketomaximizeaddedvalue

.2

Figure1.Value-difficultytrade-offfromtraditionaldescriptiveanalyticstoprescriptiveanalyticsviahuman-AIcollaborationforsustainablecompetitiveedge

Source:

Jaspersoft

2September2024Copyright??2024PharmaIntelligenceUKLimited,aCitelinecompany(Unauthorizedphotocopyingprohibited)

AIinPharma:Benefits,Risks,andtheRoadAhead

HowAIspecificallybenefitspharma

AIcanacceleratedrugdiscoveryand

developmentbysupportingtheanalysisofvastanddifferingdatasets,includingcomprehensivedrugdatabases,biochemicaldata,clinical

trialdata,andelectronichealthrecords(EHR).

AIanalysisismuchfasterandcheaperthan

traditionalmethodsatidentifyingpotentialdrugcandidates,reducingthetimerequiredfordrugdiscoveryandmaximizingthequalityofthe

novelcompound.

Forexample,AI-drivendrugdiscoveryplatformshavesignificantlyreducedtimetoidentifydrugcandidates.Whatusedtotakefourtofiveyearscannowtakeaslittleaseightmonths

.3

AIcanalsoempowerdrugrepurposing.Itcanidentifydrugcompoundsalreadyapproved

forotherindicationsandhelptopredicttheir

probabilityofsuccesswhenrepurposedto

treatdifferentdiseases.ThereareAI-enabled

systemsthathelpprioritizethemostpromisingcandidatesandestimatethesafetyprofile

andefficacyofexistingdrugsforother,similardiseases

.4,5

AIenablesboththedevelopmentof

personalizedtreatmentsandtitrationof

treatmentpathways.Itcanalsohelpadviseonmedicationswitchingandtailoringdosageto

anindividual’sspecificneeds.Itdoessoinpart

byanalyzingmultimodalpatientdatatopredicthowanindividualwouldrespondtoatreatment.

AIcansupportclinicaltrialplanningand

optimizeclinicaltrialdesignthroughtailored

protocoldesignsandinvestigatorandsite

selection.Furthermore,AIcanhelpmonitorandrescuestudies.Thisensureson-timepatient

recruitmentandtrialdeliverysodrugscanreachthemarketandthepatientsthatneedthemontimeandonbudget.

Whenitcomestoensuringclinicaltrial

diversity,AIsystemscanbeappliedtohelp

mitigatehumanbiasesinclinicaltrialdesign

andoptimizethetrade-offbetweenincreasingdiversityacrossvariouscharacteristics(e.g.,

race,ethnicity,demographics)anddelivering

thetrialontime.Thesetoolscanensureclinicaltrialdiversityrequirementsaremetandeven

exceeded.

AIcanautomateandstreamlineprocesses

suchasdrugmanufacturing,supplychain

management,clinicaltrialplanningand

execution,andpharmacovigilance.Thiscan

leadtoincreasedoperationalefficienciesfor

pharmaceuticalsponsorsandthecontract

researchorganizations(CROs)thatruntrialsontheirbehalf.

3September2024Copyright??2024PharmaIntelligenceUKLimited,aCitelinecompany(Unauthorizedphotocopyingprohibited)

AIinPharma:Benefits,Risks,andtheRoadAhead

Figure2.ApplicationsofAIindrugdevelopment

ADME:Drugabsorption,distribution,metabolism,andexcretionSource:

/pmc/articles/PMC10385763/

4September2024Copyright??2024PharmaIntelligenceUKLimited,aCitelinecompany(Unauthorizedphotocopyingprohibited)

AIinPharma:Benefits,Risks,andtheRoadAhead

RisksofemployingAItools

ApplyingAIinpharmainvolvesthepotential

useofsensitivepatientdata,includingpersonalidentifiableinformation(PII),whichraises

concernsaboutdataprivacyandsecurity.Thisrequiresleveragingdataasdirectedundertheapplicableregulationsandstandards,suchas

HIPAA

intheUSand

GDPR

intheEuropean

Union.

Withthisinmind,itisbestpracticetoleveragetheminimumamountofdatarequiredfora

businessapplication.Forinstance,whenutilizingreal-worlddata(RWD),useageinsteadofdateofbirthwherepossible,orpostalcodearea

insteadoffulladdress.

Itisalsocrucialtoapplytherequiredencryptiontosuchpatient-leveldataatrestandintransit,andtomodelsconsumingthem(e.g.,encryptedmodelendpoints).

Inlightoftechnologicaladvancessuchas

GenAI,integratingAIinpharmarequires

revisitingexistingregulationsandstandardstoensureitsethicalapplication.Guidingprinciplessuchasthe

GoodMachineLearningPractice

(GMLP),ajointeffortoftheUSFoodandDrugAdministration(FDA),HealthCanada,and

theUK’sMedicinesandHealthcareproductsRegulatoryAgency(MHRA),canhelpinthisregard.

Whileaddressingethicalandregulatory

concernsiscrucial,redefiningsuchregulationsandstandards,aswellasunderstandinghowtofullycomplywiththem,mayslowtheadoptionofAIinpharma.

Fromasoftwaredevelopmentperspective,goodpractice(GxP)compliancecanbeguaranteed

byensuringdataandmodelreproducibility

throughversioningwithtoolslike

DataVersion

Control

andappropriatelyloggedexperiments

andartifacts,aswellasdocumentationsof

methodologicalandevaluationstepsfollowedforbothtechnicalandnon-technicalaudiences.

OutcomesfromAI-drivensystemsrelyheavilyonthequalityofthetrainingdataused.Somefactorsaffectingthisqualityincludepresenceandextentofoutliersandmissingvalues,lackoforlimitedrepresentativeness,andnoisyorincorrectdata.Theterm“noisydata”refers

todatathatcontainirrelevantorerroneous

datapoints.

6

Asthesayinggoes,“garbagein,garbageout”—trainingwithdatathatare

suboptimalinqualitywillyieldsuboptimalresults.

WaystoimprovedataqualityforAImodeldevelopmentinclude:

?Imputationofmissingvalues:fillingin

missingdatapointstoachievecompletedatasetsviaeitherstatisticalorML-drivenapproaches

7

?Dataaugmentation:creatingnewdatapointsfromexistingonestoincreasetheamountofdata

8

?Datastandardization:ensuringthedatafollowaconsistentformatandstructure

ThereisalsoariskofAIperpetratinghuman

biases.Biasesinthetrainingdataresultfrom

humanprocesses;theyneedtobeanalyzedandmitigatedsothatAIsystemsdonotcontinue

toperpetuatethesebiasesintheirpredictions.

Somemethodsforaddressingthesebiases

includediversedatacollection;adversarial

training,whichinvolvestraininganeuralnetworktoevaluateAI-generatedcontentforbias

9

;anddataaugmentationtoenhanceparticipationofunderrepresentedpopulations.

Thehighcomplexity,andthereforelimited

transparency,ofcertainAI-drivensystems

likelargelanguagemodels(LLMs)withbillions

5September2024Copyright??2024PharmaIntelligenceUKLimited,aCitelinecompany(Unauthorizedphotocopyingprohibited)

AIinPharma:Benefits,Risks,andtheRoadAhead

ofparameterscanhindertrustandslowtheir

adoption.Thisisespeciallytrueinhighly

regulatedindustriessuchaslifesciences,wheretheoutcomesfromtheseAIsystemscontributetothedesignanddeliveryoftreatmentsto

patients.

Forexample,itisimportanttoreviewand

assesstheevidenceusedbyLLMstogenerateoutputs.Groundingandtailoringtheseoutputswithproprietarydataviaretrieval-augmentedgeneration(RAG)andfinetuningcanleadto

moreexplainableandaccurateoutputswhenusingLLMs.

WherepharmaisheadingwithAI

ThenextphaseofAIadoptionisalreadyunderwayinpharma.Companiesarehardatwork

devisingthenextgenerationofAI-enabled

technologiestosupportthedrugdevelopmentprocessfromdiscoverytolaunchandbeyond.

Forinstance,GenAIplatformssuchasNVIDIABioNeMoareacceleratingdrugdiscoveryby

optimizingmoleculardesigns

.10

IBMemployedGenAIineffortstorepurposedrugsfor

insomniaandParkinson’sdiseasetotreatthedementiathatoftenaccompaniesParkinson’s(PDD).

11

EdgeAIishelpingsupportmoredecentralizedtrialswithremotepatientmonitoringthrough

theuseofwearabledeviceswithsensorsto

monitorvitalsignsinrealtime.Thesedata

areanalyzedlocallyandcanalertdoctors

orpatientsandtheircaregiversofany

pathophysiologicalactivities,whichcanhelp

withtimelyinterventionandenhancequalityoflife

.12

Multi-modalGenAIisbeingusedtoanalyze

RWDtoidentifysubjectsmoreholisticallyandforprecisionmedicine.Forexample,using

GenAItosimultaneouslyconsiderdatasuch

asimaging,EHR,andmultiomics(atypeof

biologicalanalysis)helpsdeveloppersonalizedtreatmentsbyunderstandingpatients’

conditionsmoreholisticallythroughtheentirepatientjourney.

13

AIcanalsobeemployedtodevelop“digital

twins”forfurtherpersonalizedmedicine.In

oneinstance,adigitaltwincanbeusedto

simulateapatient’sglucose-insulindynamics,contributingtopersonalizedinsulindelivery

patterns.Thiscanhelppatientswithtype

IIdiabetesmanagetheirconditionmore

effectively.

14

Andinthecardiovascular

arena,digitaltwinsempoweredbymulti-

modaldatafromimaging,wearables,andelectrocardiogramscanhelppredicthow

theheartreactstodifferenttreatmentsanddosages

.15

QuantumMLcanpredictalternative

treatmentpathwaysandtheiroutcomes.

Quantumsupportvectormachineshavebeen

investigatedtodetectschizophreniafrom

electroencephalography(EEG)signals

16

,and

quantumneuralnetworkshavebeenexploredtopredicttreatmentoutcomesindepressionand

anxietyandimprovethediagnosisofpatientswithParkinson’s

.17

QuantumAIhasalsobeenemployedin

enhancingencryptionofsubject-leveldata

suchasRWDandEHR.Notably,IBMdevelopedquantum-safecryptographicsolutionsthat

areparticularlyrelevantforhealthcare,wheresensitivepatientdatarequirerobustprotectionagainstquantumattacks.

18

6September2024Copyright??2024PharmaIntelligenceUKLimited,aCitelinecompany(Unauthorizedphotocopyingprohibited)

AIinPharma:Benefits,Risks,andtheRoadAhead

CitelineSmartSolutions

CitelineisincorporatingAI,advancedML,andLLMsinitsSmartSolutionssuiteofproducts,whichhelpstudysponsorsreducecostly

protocolamendments,increasepredictabilityinclinicaltrialplanning,andaccelerateclinicaldevelopment

.19

ProtocolSmartDesigncombinesindustry-leadingdatafromCiteline’sTrialtroveand

Sitetrovesolutionswithreal-worldand

proprietaryperformancedataassetstobuild

anddelivermorereliableclinicaltrials.And

InvestigatorSmartSelectleveragesAI-enabledtechnologybuiltonSitetroveandTrialtrovedatatodeliveralistofhigh-performinginvestigatorswiththeexperienceandcapacitytodelivera

clinicaltrialontimeandonbudget.

7September2024Copyright??2024PharmaIntelligenceUKLimited,aCitelinecompany(Unauthorizedphotocopyingprohibited)

AIinPharma:Benefits,Risks,andtheRoadAhead

References

1McKinsey&Co.(2024)GenerativeAIinthepharmaceuticalindustry:Movingfromhypetoreality.Availablefrom

/industries/life-sciences/our-insights/generative-ai-in-the-pharmaceutical-industry-moving

-from-hype-to-reality#/(AccessedAug.21,2024)

2Jaspersoft(2024)WhatisPrescriptiveAnalytics?Availablefrom

/articles/what-is-prescriptive-

analytics

(AccessedAug.21,2024)

3SavageN(2021)TappingintodrugdiscoverypotentialofAI.BiopharmaDealmakers.Availablefrom

https://www.nature.

com/articles/d43747-021-00045-7

(AccessedAug.21,2024)

4Cyclica(2017)CyclicaLaunchesLigandExpressTM,aDisruptiveCloud-BasedPlatformtoRevolutionizeDrugDiscovery.

Availablefrom

/press-releases/cyclica-launches-ligand-express-a-disruptive-cloud-based-platform-

to-revolutionize-drug-discovery/

(AccessedAug.21,2024)

5KamyaP,OzerovIV,PunFW,etal(2024)PandaOmics:AnAI-drivedPlatformforTherapeuticTargetandBiomarkerDiscovery.JournalofChemicalInformationandModeling,64(10).Availablefrom

/doi/10.1021/acs.jcim.3c01619

(AccessedAug.21,2024)

6OttenNV(2023)DataQualityInMachineLearning–Explained,Issues,HowToFixThem&PythonTools.SpotIntelligence.Availablefrom

/2023/04/07/data-quality-machine-learning/

(AccessedAug.22,2024)

7VarsheniS(2024)Yourguidetomissingvaluesimputation.TrainInData.Availablefrom

/

your-guide-to-missing-values-imputation/

(AccessedAug.22,2024)

8AwanAA(2022)ACompleteGuidetoDataAugumentation.Datacamp.Availablefrom

/tutorial/

complete-guide-data-augmentation

(AccessedAug.22,2024)

9GuvvalaS(2023)BiasMitigationinGenerativeAI.AnalyticsVidhya.Availablefrom

/

blog/2023/09/bias-mitigation-in-generative-ai/

(AccessedAug.22,2024)

10PowellK(2024)NVIDIAGenerativeAIisOpeningtheNextEraofDrugDiscoveryandDesign.NVIDIA.Availablefrom

https://

/blog/drug-discovery-bionemo-generative-ai/

(AccessedAug.22,2024)

11Rosen-ZviM(2021)FindingNewUsesforDrugswithGenerativeAI.IBM.Availablefrom

/blog/

generative-ai-new-drugs

(AccessedAug.22,2024)

12NationalEdgeAIHub.Availablefrom

https://edgeaihub.co.uk/edgeai-health/

(AccessedAug.22,2024)

13PoonH(2024)MultimodalGenerativeAI:theNextFrontierinPrecisionHealth.Microsoft.Availablefrom

https://www.

/en-us/research/quarterly-brief/mar-2024-brief/articles/multimodal-generative-ai-the-next-frontier-in-

precision-health/

(AccessedAug.22,2024)

14Mosquera-LopezC,JacobsPG(2024)Digitaltwinsandartificialintelligenceinmetabolicdiseaseresearch.Trendsin

Endocrinology&Metabolism,35(6),549–557.Availablefrom

/trends/endocrinology-metabolism/

abstract/S1043-2760(24)00113-9

(AccessedAug.22,2024)

15Corral-AceroJ,MargaraF,MarciniakM,etal(2020)The‘DigitalTwin’toenablethevisionofprecision

cardiology.EuropeanHearthJournal,41(48),4556–4564.Availablefrom

/eurheartj/

article/41/48/4556/5775673?login=false

(AccessedAug.22,2024)

16AksoyG,CattanG,ChakrabortyS,etal(2024)QuantumMachine-BasedDecisionSupportSystemfortheDetectionofSchizophreniafromEEGRecords.JournalofMedicalSystems,48(29).Availablefrom

/

article/10.1007/s10916-024-02048-0

(AccessedAug.22,2024)

17StefanoGB(2024)QuantumComputingandtheFutureofNeurodegenerationandMentalHealthResearch.BrainSciences,14(1),93.Availablefrom

/20