§01 — Strata · the discovery agent

Find the responder subgroup hidden inside the indication.

One method, not two drugs. Strata reads a methylome against any outcome — drug response, survival, toxicity — and names the methylation-defined subgroup that responds, with the biology behind it. A general engine pointing toward methylation-guided clinical decision support — research-use today; drug response is just the first proof.

the loop methylome × outcome → rank methylation markers → name the responder subgroup → validate against known biology

Two of 286 drugs in the GDSC panel — illustrative instances. The same engine runs on any methylome × outcome pair.

§02 — Agent transcript

Watch the agent reason.

What the agent is doing

An AI agent (Claude) is given a goal and four tools — list drugs · scan the methylome for markers · test a subgroup · write the brief. It decides which tools to run, reads the results, recognizes the biology (e.g. that MTAP is the known CDK4/6 control), and writes the brief. The engine does the statistics; the agent does the judgment. The headline evidence is the continuous correlation (Spearman ρ / q); the median-split subgroup is just the operational cut. Below is a recorded run, played back step by step.

§03 — What the screen found

Figures, markers, and the brief.

Top response markers

Per-gene methylation Spearman-correlated against drug AUC (lower AUC = more sensitive). A negative rho means methylation marks sensitivity. FDR q < 0.1.

Gene	rho	q-value

Opportunity brief

Generated by the agent in a portfolio-ROI framing — the artifact a business-development team reads.

Illustrative prototype — research use only. Markers are derived from the GDSC cell-line screen; novel hits are screen-derived hypotheses, not validated companion diagnostics. Their credibility comes from the same unbiased screen recovering the known 9p21 / MTAP–CDKN2A → palbociclib positive control: MTAP is the screen's #1 marker (CDKN2A's co-deleted 9p21 neighbor), and CDKN2A itself is recovered deeper in the ranked list — see the canonical-biomarker check below the marker table. Not for clinical decision-making.

The evidence

§04 — Scale · where this goes

The kernel is GPU-shaped by construction.

Strata's core operation is one Spearman correlation per gene per drug — embarrassingly parallel over (gene × drug × cohort). Discovery is cheap and decentralizing; what gets expensive is re-screening the whole space as the data grows, and that cost lands on exactly the layer we own.

The bridge. Every licensed marker→response association is a formal, provenance-stamped claim. Strata is the working on-ramp to the Polymer claims universe — a federated graph where contributors generate associations on their own compute and a GPU-bound center keeps the universe cross-checked and coherent. Generation decentralizes; verification is the moat — and verification is RAPIDS-native by construction. You build the engines for autonomous science — we're the structured memory they write to.