AI-driven Drug Discovery Faces 2026 Test

Writing in ProMarket, the publication of the Stigler Center at the University of Chicago Booth School of Business, business ethics professor Michael Santoro argues that a decade of AI drug-discovery investment, including Xaira's $1 billion 2024 launch and Alphabet's Isomorphic Labs raising $600 million in 2025, faces its first real-world efficacy check in 2026 as a meaningful cohort of AI-discovered drugs reaches Phase 3 trials. Santoro cites a Boston Consulting Group analysis of roughly two dozen AI-discovered molecules: in Phase 1 safety trials they succeed 80-90% of the time versus a historical norm of about 50%, but in Phase 2 trials, the first real test of whether a drug helps patients, success falls back to the industry's ordinary rate of roughly 40%. BCG's overall estimate is that AI roughly doubles a drug's odds of reaching market, from about 5-10% to 9-18%, with that gain concentrated in the cheap early stages (a Phase 1 trial averages about $4 million, Phase 2 about $13 million, Phase 3 $20 million or more, against a total drug-development cost of roughly $2.6 billion). Santoro points to Insilico Medicine's rentosertib, which posted positive mid-stage results published in Nature Medicine in 2025, as an early bright spot, against Recursion's 2025 discontinuation of its lead AI-discovered program.

Santoro's argument is that AI has mainly solved the tractable part of drug discovery, molecule design, which is fundamentally a chemistry and physics problem with abundant data, while leaving the harder problem largely untouched: choosing which biological target actually drives a disease, which depends on understanding human biology that remains sparsely mapped. For ML teams and investors evaluating pharma AI, his framing suggests the more valuable, more contested opportunity is applying AI directly to target validation, biomarker discovery, and trial design (the Phase 2/3 problem) rather than to additional molecule-generation platforms, which he argues are now a crowded, competed-down market.

Santoro flags the results of the current Phase 3 cohort of AI-designed drugs, due by the end of 2026, as the first real evidence on whether AI-driven discovery translates into approved, effective medicines. He also notes a secondary policy wrinkle: because AI has made generating drug candidates cheaper, companies can pursue more Orphan Drug Act designations, potentially straining the Act's rare-disease exclusivity provisions and its cost-based eligibility test.

This is a single-author opinion and analysis piece, not a data release or company disclosure. Santoro's specific numeric claims are sourced to a cited BCG analysis and other publicly available data he links, but his conclusion, that AI has solved the wrong bottleneck, is his own economic argument and remains contested within the field.