药物研发中的人工智能：追踪AI在生物制药中日益增长的影响

EMERGING TECH RESEARCHAI in Drug Development PitchBook Data, Inc. Nizar TarhuniExecutive Vice President ofResearch and Market IntelligencePaul CondraGlobal Head of PrivateMarkets ResearchJames UlanDirector of EmergingTechnology Research Tracking AI’s growing influence in biopharma PitchBook is a Morningstar company providing the most comprehensive, most Institutional Research Group Ben RiccioAssociate Research Analystben.riccio@pitchbook.com Key takeaways •AI has proliferated across the biopharma ecosystem:Startups are deploying AIat every step of the drug development process: developing their own pipelinesof AI-discovered assets, inking massive partnership contracts for discovery Data Harrison WaldockData Analyst pbinstitutionalresearch@pitchbook.com PublishingDesigned byDrew Sanders •Biology shifts to foundation models:The introduction of foundation models in biology has already transformed protein-structure prediction and bioinformatics.Frontier applications focus on “virtual cells”: comprehensive simulations of Published on November 12, 2025 Contents •VC activity picks up:Over the past 12 months, VC investment in AI drugdevelopment has totaled $3.2 billion across 135 deals. AI-enabled pharma tools &services constitute 62.5% of the total deal count, while AI-native drug developers •AI-native biotechs have a nearly 100% valuation premium:In 2024, AI-nativebiotech companies raised money at a median valuation of $78 million, nearlytwice the broader biopharma median. The valuation gap highlights both the •AI pharma tools & services see dealmaking flurry:AI pharma tools & serviceshave seen a significant uptick in VC activity over the past four quarters, withdeal counts up 51.8% on a TTM basis. Startups in the space offer a faster path to models showing evidence of scaling behavior, continued progress hinges onovercoming data siloes and quality limitations. A competitive advantage in new Overview While AI is reshaping nearly every industry, its influence may prove mosttransformative in biopharma, where new technologies promise to address the highfailure rates, lengthy development timelines, and escalating costs that characterize The current iteration of AI in drug development is driven by the application offoundation models in biology. These models, trained on vast biological datasets,have built upon the capabilities of earlier machine learning applications in chemistrythat were developed over a decade ago and commercialized by startups suchas Atomwise and Insilico Medicine. These smaller models used task-specificarchitectures and were trained on labeled datasets, making them valuable tools AI technologies have emerged across the biopharma ecosystem: Startups aredeveloping their own pipelines of AI-discovered assets, inking massive partnershipcontracts for discovery services, offering AI-enabled platforms and services, and deploying AI in clinical trials. While approaches vary, all these startups are lookingto address the structural issues within drug development. Bringing a drug to marketcan take over 10 years and $1 billion, and with a clinical success rate of around 10%,failure is common.1These challenges have contributed to one of the most drasticfunding downturns for the sector in over a decade, as a general risk-off sentimentamong VC investors has pushed capital away from early-stage biopharma. A Technologies and recent developments Protein-structure prediction and de novo design Protein structure defines biological function, and accurate prediction of three-dimensional conformations enables researchers to better understand diseasestates and develop more targeted drugs. AI has revolutionized protein-structureprediction, overcoming a major experimental bottleneck in biology. DeepMind’s Subsequent iterations of protein models have expanded their functionality. Meta’sESMFold achieved faster inference, enabling less computationally intensive analysisof larger datasets, while AlphaFold 3 added modeling of protein-ligand and protein- Models have also moved beyond structure prediction to de novo design. RFdiffusion,developed by the Baker Lab, applies diffusion-based generation—like thoseused in image and video models—to create entirely new proteins with desired structures and functions. Results have included AI-generated binding proteins,complex enzymes, and, most recently, ion channels.2, 3Other breakthroughs include Genomic models AI has also become increasingly valuable in genomic analysis. Early tools such asDeepVariant and DeepConsensus improved the accuracy of variant detection andthe quality of sequencing data; however, these models lacked the scale to generate Importantly, these models appear to follow similar scaling laws to those observedin language and vision models, where performance improves with dataset andmodel size. More recent models have dramatically expanded their scale. Evo 2, developed by the Arc Institute and NVIDIA, was trained on 9.3 trillion base pairs