28 September 2026
Two AI agents, each built from a different paper, found a genetic link neither study reported
Stanford Medicine's Paper2Agent turns a scientific paper, along with its code and data, into an interactive AI agent that can answer questions and rerun the original analysis. In a demonstration, researchers let an agent built from a genetic-variant-effect model talk to an agent built from an ADHD genetics study; together they flagged a variant near the gene MPHOSPH9 as linked to ADHD risk, a connection nobody had reported before.
Source: med.stanford.edu →
An AI model searched 1.7 million molecules and handed back ten working antibiotics
Stanford engineers trained an AI model to scour a library of 1.7 million synthetic polymers for candidates that kill bacteria the way natural peptides do, without giving microbes an easy path to resistance. Lab tests on E. coli confirmed all ten finalists worked, and one was especially effective against biofilms, the stubborn bacterial colonies that resist most conventional antibiotics.
Source: news.stanford.edu →
A tool that flags suicide risk in crisis texts, and shows its work
MIT McGovern Institute researchers built a lightweight, interpretable model that scans crisis-line conversations for language tied to 49 known risk factors, trained and tested on roughly 16,000 de-identified chats and validated in a peer-reviewed study. Unlike a typical language model it runs on an ordinary computer, keeps conversations private, and shows counselors exactly which words drove its estimate; the team is clear it still needs further validation before clinical use.
Source: news.mit.edu →