Recursion

Headlines · 3

1. 5d
   Leash Bio releases Minos, an agent sandbox for verifiable scientific reasoning

   A Latent Space episode features Ian Quigley of Leash Bio, which uses DNA-encoded libraries (DELs) to collect high-throughput protein–small molecule interaction data. The show says that data can supply tens of thousands of "lab-in-the-loop" simulations for scientific RL by agents, and has matured into Minos, with Leash Bio releasing a full preprint, blog and open dataset the same day. The episode also says Hermes, trained only on Leash Bio's own data, keeps pace with Boltz on public benchmarks, and that in its Kaggle competition about 2,000 teams failed to beat chance on an unseen chemical library. These claims come from the podcast description; the preprint and dataset are not included in the supplied material.

   Latent Space · 🔥 0
2. 6d
   AI drug discovery hits a bottleneck as wet-lab work slows the pace

   An industry commentary argues that AI is already embedded in pharma R&D — literature review, molecule generation and clinical-trial management — but models trained only on public data cannot create a moat, so companies are shifting toward building their own “dry-wet” design-build-test-learn loops to accumulate proprietary data. It notes that Anthropic was reported to have set up a biological wet lab in the San Francisco Bay Area, that Eli Lilly's Lilly TuneLab signed GenScript, Twist Bioscience and Ginkgo Bioworks to supply wet-lab data, and that in the Novo Nordisk breach the leaked proprietary biological and chemical training data amounted to only about 407MB. Citing estimates compiled by Southwest Securities, the piece says AI can cut timelines by 70–90% in discovery, 50–80% in pre-clinical and 50–60% in clinical stages, while stressing that AI remains a hypothesis-generating assistant rather than a decision system, and that success still hinges on whether a target truly causes the disease.

   虎嗅 AI · 🔥 0
3. 6d
   Roche says AI informs 40% of pipeline decisions, targets 80%

   At its late-September Pharma Day investor event, Roche said 40% of pipeline decisions from Q4 2025 to Q2 2026 carried a traceable AI or computational contribution, rising to 44% in Q2 2026, split 41% critical, 48% supportive and 10% informational. Genentech gRED head Aviv Regev presented the target-evaluation agent TargetNexus and a "Lab-in-the-Loop" system aiming to be involved in 80% of research portfolio decisions by Q4 this year. Roche says that with nominally flat resources and lead-identification cycles halved, projects entering lead identification rose 155% and internally generated development candidates rose 50% between 2022 and 2025, and one biologics program reached a clinical candidate in 18 months, with first patient dosed this year. Caveats: Roche has not disclosed how it scores AI contributions, 40% times 41% implies only about 16% of decisions were critically AI-influenced, and the Phase III success rise from 65% in 2025 to above 80% cannot be separated from the stricter "The Bar" framework; no approved drug has yet come from the closed loop.

   MIT科技评论中文 · 🔥 0

Experience and discussion from the community

Share my Recursion experienceAsk about Recursion