人工智能如何改写药物开发经济学

Please refer to important disclosures on pages 42 and 43. Analyst certification is on page 42.William Blair or an affiliate does and seeks to do business with companies covered in its research reports. As aresult, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of thisreport. This report is not intended to provide personal investment advice. The opinions and recommendations here- Portfolio Manager Summary.............................................................................................3Executive Summary...........................................................................................................3Ecosystem Overview..........................................................................................................9Public Company Coverage...............................................................................................11How AI Will Impact the “Life of a Drug”..........................................................................12Why AI Is Relevant: Drug Discovery Is Expensive and Biology Is Programmable.........12Biological Foundation Models – In the Dry Lab, Data Is the Limiting Reagent.............14How AI Is Impacting Antibody Discovery – Smarter Funnels and Faster Cycles...........20Target Selection (Steps 1-2)............................................................................................22Step 1 – Target Identification........................................................................................22Step 2 – Target Validation.............................................................................................23Step 2.5 – Antigen and Assay Preparation (Screening Preparation)............................24 Here is the short version: AI is not replacing the physical work of drug discovery, but it is reshapingit. It is true that AI is increasingly being leveraged as a “dry lab” (i.e., computational) tool for hy-pothesis generation, but those hypotheses will always need to be physically tested in the “wet lab”(i.e., the real world) and the overwhelming majority of drug discovery spend relates to the work Separately, the most balanced assessment of AI in drug discovery today is that it is good at generat-ing ideas. Not necessarily good ideas, but lots of them—quickly. Like all AI, the quality of outputsrelies on the quality of inputs, and biology today lacks the requisite inputs (standardized, fit-for-purpose, experimental data) to generate desired outputs (novel drug ideas). Hence, for AI to reach On thestocks: the life science tools and services group has traded off categorically in recent weeks,partly on fears that AI will dramatically reduce wet lab utilization in drug discovery. However, on balance we view AI as a net tailwind for most companies we cover given where their products sit inthe drug discovery workflow. Thus, the setup is asymmetric—alpha is available both if AI becomesa ubiquitous tool in drug discovery (underappreciated flow-through to physical lab work) and if itdoes not (dispelling the wet lab obsolescence thesis). Investors should be adding exposure as end- In a 2021 report (Programming Life), we discussed how advances in AI were enabling a new mod-el for drug discovery—one that leverages the inherent programmability of biology to automateand accelerate the design–build–test–learn (DBTL) cycle. As is often the case with new technol-ogy waves, excitement created hype that was more on par with promises made by the field thanprogress made on the field, which in turn created a bubble-popping cycle that led to persistentskepticism about the potential impact of AI in biology ever since. More recently, however, that par-adigm has flipped on the heels of announcements by Anthropic about Claude for Life Sciences andbroader awareness of progress made with in silico drug discovery tools (e.g., Microsoft CEO SatyaNadella’s recent X post about Microsoft’s GigaTIME). The S&P Life Sciences Tools & Services Indexis down 19% year-to-date and 25% from January highs, in part related to fears that the recent in- These concerns contemplate, perhaps with equal levels of imprecision, the potential for AI to 1)reduce discovery work by identifying targets and designing drugs and 2) reduce analytical workload by powering predictive models pertaining to efficacy and safety, with a potential offset from3) more molecules entering the funnel due to the improving efficiency of drug discovery and de- Our work builds on conversations with 20 companies and key opinion leaders in the space, a deepdive on biological models, and analysis of the drug discovery process and how cost distribution, AIdisruption, and companies’ products map to that process. Based on this, we see the narrative beingreshaped in the coming years as the use-cases and limitations of AI in drug discovery are better powering the production and adaptation of biological models, shifting test volume to high-contentand c