# Sources for "「AI創薬は第2相で止まる」はどこまで言えるか"

Read October 2026. Derived numbers: numbers.py / results/summary.json.

| Source | Used for | Read |
|---|---|---|
| Jayatunga et al., Drug Discov Today 2024;29(6):104009, doi:10.1016/j.drudis.2024.104009 (PMID 38692505) | Phase I 80-90%, Phase II ~40% "albeit on a limited sample size" | Abstract only (Europe PMC) |
| Santoro, "Pharma's AI Boom Has Bet on the Wrong Bottleneck", ProMarket, 2026-07-02 | Argument; figures attributed to BCG analysis; rentosertib "brightest result"; no count of Phase 3 cohort | Full page |
| Epstein & Niazi, Petrie-Flom Center, 2026-08-04 | Argument that evidence, not discovery, is rate-limiting; "10 times faster" attributed to a company news page, not verified by authors | Full page |
| Alsumidaie, The Clinical Trial Vanguard, 2026-08-11 | 117 tracked assets / 8 cleared Phase II and 80-90% attributed to a LinkedIn post; ~40% unsourced | Full page (LinkedIn posts not read) |
| Trey, Substack, 2026-04-01 | Bottleneck = evidence generation; 6-10 years and 10-15% without sources | Partial (paywalled) |
| Haque et al., ASCO 2026 abstract, J Clin Oncol 44(16_suppl):11072; doi:10.1200/JCO.2026.44.16_suppl.11072 | 117 assets / 63 companies; 60 phase 1 completions; 8 phase 2 completions. Entry deadline 2025-07-01, progress cutoff 2025-12-01 | Original abstract text in publisher-deposited Crossref record, retrieved 2026-11-03. Direct publisher HTML/PDF unavailable |
| IQVIA Institute blog 2026-05-14 on Global R&D Trends 2026: AI-enabled Phase I 75%, Phase II 'on par' with non-AI, no number, small cohort | Phase II comparison | Blog only (iqvia.com/blogs/2026/05/...) |
| Own articles: ai-phase2-sample-size, rentosertib-ph2a | 4/10 example; rentosertib 60 mg missingness 44% | Own work |

Not read: LinkedIn posts, Reddit, X, Fast Company (403), the original BCG/Jayatunga tables.


## Original abstract verification (2026-11-03)

Haque et al., JCO2026;44(16_suppl):11072, DOI10.1200/JCO.2026.44.16_suppl.11072: the abstract text in the publisher-deposited Crossref record was retrieved and read. It identifies assets entering phase1–3 interventional trials through1 July2025, with progress assessed on1 December2025:117 assets /63 companies,60 phase1 completions (51.3%) and8 phase2 completions (6.8%). Evidence of AI contribution to discovery or design was required at asset level. The abstract does not identify completion with efficacy success or supply a resolved-phase2 denominator. No asset-level recount was performed. The initial audit had used secondary commentary; that limitation is superseded by this original-abstract check. Direct publisher HTML/PDF requests returned403.

Crossref original abstract record: https://api.crossref.org/works/10.1200/JCO.2026.44.16_suppl.11072
