Harvard releases privacy-preserving agent tool for blood biomarker discovery, +4.18 AUC points median
Harvard · hf · 2026-10-07
Harvard researchers released a study and tool on Hugging Face: agentic discovery of blood biomarkers from distilled private health records. The core problem: evaluating candidate biomarkers requires private records that can't be shared with frontier LLM agents.
Approach: They distilled evidence from Clalit Health Services' panel of 5.4M+ patients into a releasable scoring tool — for each of 13 immune-mediated diseases, a graph attention network was trained inside the data boundary to predict the case-control AUC of candidate CBC expressions; only trained weights were released. The tool grounds an agent's propose-score-refine loop without exposing patient data.
Results: Agent-discovered expressions beat their literature-seeded starting points by a median of 4.18 AUC percentage points in external validation; reranking candidates from three frontier research tools improved on their first choices in most comparisons across three independent cohorts.
More from Safety
- Dreadnode releases ScopeBench, a benchmark for agent scope adherence in offensive security — dyn___ · 2026-10-07
- AI Safety x Quant mixer in London on Nov 15 opens applications — NeelNanda5 · 2026-10-07
- EU EIC Accelerator offers €2.5M grants at zero equity, but founders warn of red-tape costs — DesireeCachette · 2026-10-07
- Experts weigh AI-enabled cyber risk: scaled CNE within a year, cryptoanalytic breakthroughs as phase change — teortaxesTex · 2026-10-07
- OpenSSH drops sshd sandboxing on macOS as SDK 27 removes key API — jedisct1 · 2026-10-07
- vf3 fuzzer unveiled at OAIC claims to outpace Jackalope and libprotobuf-mutator — dyn___ · 2026-10-07