2025年对制药行业有何影响？抓住你的帽子…

White Paper AI in Pharma:Benefits, Risks, andthe Road Ahead by Luca Parisi AI in Pharma: B enefits, Risks, and the Road Ahead Introduction Use of artificial intelligence (AI) now permeatesevery industry, and pharma is no exception.Whether machine learning (ML) models thatmake predictions based on existing data orgenerative AI (GenAI) models that create newdata based on the data they were trained on, AIis being used to streamline and accelerate eachstep of the drug development process fromresearch through approval and marketing. unstructured data, including multimodal datasuch as tabular, text, images, and videos. Atits most basic level, AI can automate mundanetasks such as structured document and imageanalyses, enabling experts to spend moretime on tasks that require their attentionand proficiency. On a deeper level, AI can unveil insights fromhistorical data to inform operations andprovide a lens into the future through predictiveanalytics, supplementing traditional descriptiveand diagnostic analytics that solely provideanalytical information anchored on historicalpatterns. It can also enable and accelerateexpertise through prescriptive analytics,advising experts on the next best action to taketo maximize added value.2 According to McKinsey & Co., generative AIalone could produce $60 billion to $110 billiona year in economic value across the pharmaindustry value chain. And $13 billion to$25 billion of that annual value alone wouldbe for clinical development.1 AI is able to handle both structured and AI in Pharma: B enefits, Risks, and the Road Ahead How AI specifically benefits pharma AI can accelerate drug discovery anddevelopment by supporting the analysis of vastand differing datasets, including comprehensivedrug databases, biochemical data, clinicaltrial data, and electronic health records (EHR).AI analysis is much faster and cheaper thantraditional methods at identifying potential drugcandidates, reducing the time required for drugdiscovery and maximizing the quality of thenovel compound. by analyzing multimodal patient data to predicthow an individual would respond to a treatment. AI can support clinical trial planning andoptimize clinical trial design through tailoredprotocol designs and investigator and siteselection. Furthermore, AI can help monitor andrescue studies. This ensures on-time patientrecruitment and trial delivery so drugs can reachthe market and the patients that need them ontime and on budget. For example, AI-driven drug discovery platformshave significantly reduced time to identify drugcandidates. What used to take four to five yearscan now take as little as eight months.3 When it comes to ensuring clinical trialdiversity, AI systems can be applied to helpmitigate human biases in clinical trial designand optimize the trade-off between increasingdiversity across various characteristics (e.g.,race, ethnicity, demographics) and deliveringthe trial on time. These tools can ensure clinicaltrial diversity requirements are met and evenexceeded. AI can also empower drug repurposing. It canidentify drug compounds already approvedfor other indications and help to predict theirprobability of success when repurposed totreat different diseases. There are AI-enabledsystems that help prioritize the most promisingcandidates and estimate the safety profileand efficacy of existing drugs for other, similardiseases.4,5 AI can automate and streamline processessuch as drug manufacturing, supply chainmanagement, clinical trial planning andexecution, and pharmacovigilance. This canlead to increased operational efficiencies forpharmaceutical sponsors and the contractresearch organizations (CROs) that run trials ontheir behalf. AI enables both the development ofpersonalized treatments and titration oftreatment pathways. It can also help advise onmedication switching and tailoring dosage toan individual’s specific needs. It does so in part AI in Pharma: B enefits, Risks, and the Road Ahead AI in Pharma: B enefits, Risks, and the Road Ahead Risks of employing AI tools Applying AI in pharma involves the potentialuse of sensitive patient data, including personalidentifiable information (PII), which raisesconcerns about data privacy and security. Thisrequires leveraging data as directed under theapplicable regulations and standards, such asHIPAAin the US andGDPRin the EuropeanUnion. and artifacts, as well as documentations ofmethodological and evaluation steps followedfor both technical and non-technical audiences. Outcomes from AI-driven systems rely heavilyon the quality of the training data used. Somefactors affecting this quality include presenceand extent of outliers and missing values, lackof or limited representativeness, and noisy orincorrect data. The term “noisy data” refersto data that contain irrelevant or erroneousdata points.6As the saying goes, “garbage in,garbage out” — training with data that aresuboptimal in quality will yield suboptimalresults. With this in mind, it is