Agentic discovery of blood biomarker from distilled private health records

Agentic discovery of blood biomarker from distilled private health records. By Seffi Cohen and nine others, from Harvard Medical School, the Clalit Research Institute, and one other institution. Published on the third of October. Its code repository is worked out: about forty-five source files, with a readme and a licence.

Routine complete blood counts could yield new biomarkers. The obstacle is a mismatch between where the evidence sits and where the discovery tools are. The private records needed to evaluate a candidate cannot be shared with frontier language model agents, and those agents are the very tools that excel at discovery. This work releases a scoring tool that stands between the two: it carries the evidence in those records without carrying the records.

The tool covers thirteen immune-mediated diseases. For each one, a graph attention network predicts the case-control AUC of a candidate expression built from blood count values, in effect, how well that expression separates patients with the disease from controls. An agent can propose an expression, score it, and refine it, and that loop is grounded in real-world data without exposing any patient data.

The evidence comes from the Clalit Health Services panel of over five point four million patients. Each network was trained inside the data boundary, and only the trained weights were released.

In external validation, expressions discovered by an agent improved on their literature-seeded starting points by a median of four point one eight AUC percentage points. Across three independent cohorts, reranking the candidates of three frontier research tools improved on their first choices in most comparisons, with gains that varied by cohort.

It is built for anyone generating biomarker hypotheses with agents who cannot reach the underlying records. The scorer supports that work while keeping the data private.

The data never moves. What leaves the boundary is a model of what the data would say about a candidate, and the agent searches against that model instead of the records. Privacy and discovery stop competing for the same access.