METHYLGKB by Polymer Genomics

§01 — Strata · molecular subgroup discovery engine

Start from the molecular state, not the disease.

Ask the agent about any drug or cancer type — it calls the report API and answers through the three structural gates.

Every report below is served by a standardized API the agent calls against. Replay the agent's tool calls, then drive the same reports yourself: pick a drug or cancer type, and watch the funnel, leads, and replication evidence render live.

§02 — Ask the agent

Ask a question; watch it reason over the API.

Type a question about any drug or cancer type and the live agent answers it — listing the report tools, calling the one it needs, and narrating through the three structural gates. The canned runs below the box are example queries; click one to prefill.

How to read the report
ρ  (methylation ↔ AUC) negative = hypermethylation marks sensitivity (lower AUC = more response)  ·  positive = marks resistance
direction label the sensitive / resistant tag already does the sign math — read it, not the sign
annotation score positive = predicted sensitive (sign-flipped for readability — opposite of ρ)
evidence level L3 GDSC only  ·  L4 + 1 external screen  ·  L5 all 3 (GDSC + PRISM + CTRP). Every lead is already response-associated and functionally silenced.
ASK

Example runs · click a tab to load, then replay

QUERY

§03 — Live report explorer

Drive the report API yourself.

Pick a report type, then a drug or cancer type. The page fetches the standardized report and renders it — the same shape the agent receives.

§04 — Patient methylation annotation

Drop in a methylation profile; read predicted drug sensitivities.

Hand the engine a promoter-methylation profile and it scores every validated lead against the patient — aggregating weak single-gene signals into a per-drug sensitive / resistant call with the driver genes that pushed it. Research-use enrichment and triage, not a clinical recommendation.

Held-out validation
Olaparib ρ = −0.25 p = 0.001
Trametinib ρ = −0.37 p = 5e−7
predicted sensitivity vs measured AUC, on held-out cell lines (negative ρ = works)
Paste your own profile (JSON {gene: beta})

A real patient tumor or PBMC sample is leakage-free — the honest input. Betas in [0, 1].

§05 — Honesty rails

What this is, and what it isn't.

The caveats travel with every report. They define the claim precisely enough to be falsifiable.