Nat. Biotechnol.｜AI驱动的新方法正在重塑动物实验，但全面替代仍需时间

TL;DR - A Nature Biotechnology news feature (Sheridan, 2026) surveying how AI-driven "new approach methodologies" (NAMs) — multi-agent virtual scientists, organ digital twins, human organoids, organs-on-chips, and PBPK simulation — are reshaping preclinical drug testing as the FDA pushes to reduce animal use, while full replacement remains blocked by regulatory qualification.

• Why animal models fail: cross-species mismatch drives ~95% development failure rates; TGN1412 passed primate testing at up to 500× the human dose yet caused life-threatening cytokine release syndrome in six Phase I volunteers. Parallel Bio reports its lymph-node organoids reproduce a TGN1412 shock-like response that human and NHP blood do not.
• Agentic hypothesis generation: Google's Co-Scientist was validated across drug repurposing, target discovery, and antimicrobial resistance; FutureHouse's Robin surfaced ripasudil (an approved Rho-kinase inhibitor) for dry AMD and flagged ABCA1 as a target. Incyte is piloting the successor system Kosmos via Edison Scientific.
• Digital twins and human-tissue platforms: a Toronto lung program built twins from 1,000 ex-vivo-perfused donor lungs; Duke's Randles group simulated 4.5M heartbeats (~6 weeks), with a full year under review. Vivodyne's vascularized organ models use 200k–500k cells each across 20+ organ types.
• Regulatory bottleneck: mechanistic PBPK is already accepted (Certara Simcyp used in 120+ FDA-approved filings; Novartis avoided ≥10 clinical pharmacology studies for asciminib), but FDA's ISTAND program has ~20 applications and zero fully qualified NAMs, with some rejections — qualification is context-of-use specific and non-transferable.