Follow
Three papers published simultaneously in Nature today, May 19, 2026, represent the concrete demonstration of AI systems performing end-to-end scientific discovery, not just assisting it (Two papers were available as preprints before).
FutureHouse published Robin, the first multi-agent system to fully automate both hypothesis generation and experimental data analysis in biology. Given dry age-related macular degeneration as its only input, Robin reviewed 551 papers in 30 minutes (estimated at over 800 human hours), identified RPE phagocytosis enhancement as a therapeutic strategy, and autonomously discovered ripasudil, a glaucoma-approved ROCK inhibitor never previously proposed for dAMD, which produced a 1.89-fold increase in phagocytosis confirmed in primary human RPE stem cells. A follow-up RNA-seq experiment, also designed and interpreted by Robin, revealed 3-fold upregulation of ABCA1 as a potential novel mechanistic target. When OpenAI Deep Research was given the same task as a benchmark, it produced zero hits and did not suggest ROCK inhibition at all.
Full Text: https://lnkd.in/ggVYGb9Z
Google DeepMind/Research published Co-Scientist, a Gemini-based multi-agent system that generates, critiques, and evolves scientific hypotheses through a self-play tournament loop that improves with more compute time. Validated across three biomedical areas, Co-Scientist identified the IRE1 inhibitor KIRA6 as a novel AML candidate with an 18-fold selectivity window between primitive AML cells and normal lymphoblastoid controls, identified Vorinostat (FDA-approved for cancer) as an anti-fibrotic compound validated in human hepatic organoids, and independently recapitulated an unpublished experimental discovery about bacterial gene transfer in antimicrobial resistance in just two days of compute time.
Full Text: https://lnkd.in/gnV4QZb8
Google DeepMind/Research published ERA, a system that uses LLM-driven tree search to iteratively rewrite and optimize scientific software against a defined quality score. ERA generated 40 methods that outperformed the best published approach for scRNA-seq batch integration on the OpenProblems benchmark, produced 14 models that outperformed the CDC CovidHub ensemble for COVID-19 hospitalization forecasting across all 52 US jurisdictions in the 2024-25 season, and achieved expert-level performance in geospatial segmentation and neural activity forecasting. Methods exploration that would take weeks now takes hours.
Full Text: https://lnkd.in/gP5hU2_8
Read together, Robin closes the wet-lab discovery loop, Co-Scientist scales hypothesis quality through iterative compute, and ERA converts empirical methods papers into runnable, improvable software. All three systems are designed to operate alongside human scientists, not replace them, but their combined effect is a measurable compression of the discovery timeline.
#AIScientist #DrugDiscovery #BiotechInnovation
