METHYLGKB by Polymer Genomics

§01 — The wedge

Methylation predicts drug response. Nobody has curated it.

Three facts make pharmaco-epigenomics a real, defensible frontier: the signal is proven, the curated resource doesn't exist, and the existing authorities openly contradict each other. Here's the evidence — caveats included.

§02 — The signal is real

Methylation alone predicts cytotoxic drug response.

Across 987 GDSC cell lines and 8 drugs, DNA-methylation-only models reached AUROC up to 0.97 — and transferred to temozolomide response in low-grade glioma patients. The canonical, clinically deployed pharmaco-epigenomic biomarker — MGMT promoter methylation → temozolomide (the Hegi/Stupp glioblastoma lineage) — is the proof that this generalizes from bench to bedside.

Honest caveat. AUROC 0.97 is best-case, and the clean patient transfer was temozolomide / low-grade-glioma specifically. The defensible claim is "promising, single-domain proven" — not "solved." Generalization beyond TMZ/LGG is an open question, and the 5hmC-vs-5mC confound applies here as in the TET2 work.

§03 — The white space

Two worlds of drug-response data. Neither curates methylation.

PharmGKB + CPIC / DPWG / CPNDS / RNPGx Clinical germline variant → dosing guidelines, expert-curated, evidence-graded. Variant-centric — no methylation.
germline · curated
PharmacoDB 2.0 Integrates GDSC, CCLE, CTRP, gCSI cell-line screens against molecular features — mutation, CNV, expression only.
omits methylation
ROC Plotter (cell-line mode) Drug-response prediction built exclusively on transcriptomics.
no methylation
CellMinerCDB v1.2 The one resource that integrates methylation for biomarker discovery — but an analysis portal, not a curated knowledge base.
portal, not KB

MethylGKB owns the empty quadrant: methylation as a first-class, curated, queryable drug-response predictor. And unlike biophysical claims, these come pre-graded — clinical truth here exists and is scored — so the database is validatable by construction.

§04 — Contradiction as signal

When the authorities disagree, that's the data.

The FDA's Table of Pharmacogenetic Associations and CPIC's guidelines cover ~106 vs ~59 drugs, overlap on 39, and are fully concordant on dosing for only 5. Most knowledge bases would force a winner. MethylGKB keeps both as a first-class CONTRADICTS edge — a flag for a hidden conditioning variable nobody has named yet.

~106
Drugs on the FDA pharmacogenetic table.
FDA
39
Drugs the FDA and CPIC both address.
overlap
5
Of those 39 where they fully agree on dosing.
Hertz, AJHP 2022

§05 — Sources

Verified references

From a deep-research sweep — 23 of 25 claims survived adversarial verification. The two that failed (specific consortium gene-drug counts; an over-strong "no methylation features" claim) are excluded here.