xX-its-amit-Xx/cyp-challenge-derived

CYP Challenge — derived data and research notes

Open artifacts from an autonomous, agent-driven campaign on the OpenADMET CYP Inhibition Blind Challenge (CYP3A4 / CYP2C9 / CYP2D6 / CYP1A2; pIC50 regression scored by RAE, plus TDI classification).

Shared so others can reuse the curation work and — just as usefully — our negative results.

What is here

file	size	description
cyp_external_wide.parquet	336 KB	Per-InChIKey median pIC50 across the 4 isoforms, aggregated from public sources.
pubchem_cyp_panel.parquet	975 KB	Harmonised PubChem NCGC qHTS CYP panel (AIDs 410/883/884/885/891/899).
chembl_cyp.parquet	617 KB	ChEMBL IC50/Ki records for the 4 challenge isoforms, converted to pIC50.
eda_official.json	5 KB	Aggregate EDA statistics (sparsity, cross-isoform correlation, noise floor, coverage).
findings.jsonl	30 KB	The campaign knowledge base: one durable research finding per line.
research_swarm_ledger.md	8 KB	The decision + state log driving the autonomous research swarm.

What is deliberately NOT here

The official challenge data is not redistributed. It is published by OpenADMET under their own terms and the test labels are blinded. This repo contains only aggregate statistics derived from it (correlations, counts, noise estimates), never per-compound rows. Get the real data from openadmet/cyp-challenge-train-test.

Headline findings (as of 2026-08-23)

These are measured, and several contradict what we assumed going in:

1. Public CYP data does not cover this challenge. Our 14,055-compound aggregated public panel shares 0 of 748 InChIKeys with the blinded test set — zero even on the InChIKey skeleton, and only 2.9% with the training set. Nearest-neighbour read-across has no neighbours to use. If you are planning to lean on public potency data for this challenge, check your overlap first.
2. CYP2D6 is orthogonal to the other three isoforms. Pearson r on co-measured compounds: 3A4–2C9 0.71, 2C9–1A2 0.45, 3A4–1A2 0.42, but 3A4–2D6 0.04, 2D6–1A2 0.08, 2C9–2D6 −0.10. A single joint 4-task head is the wrong architecture.
3. The label matrix is 73% single-isoform (33% fill rate, mean 1.33 isoforms/compound). Masked losses are mandatory; dense-target formulations fabricate two-thirds of the labels.
4. The noise floor is shipped with the data. The per-measurement _std column implies an irreducible RAE floor of ~0.16–0.25 per isoform. A local RAE far below ~0.15 indicates leakage, not skill.
5. A standard descriptor baseline gets macro scaffold-CV RAE 0.837 — but CYP2D6 sits at 1.001, i.e. no better than predicting the training mean.

Reproducing

git clone https://github.com/xX-its-amit-Xx/OpenADMET-cyp-challenge
uv sync
python scripts/external/pull_cyp_external.py   # rebuild the public panel
python scripts/eda/eda_official.py             # rebuild the EDA (needs the official data)
python scripts/train/multitask_baseline.py --source official

Provenance & licensing

Derived tables are aggregated from PubChem (public domain) and ChEMBL (CC BY-SA 3.0 — attribute ChEMBL if you redistribute). Our own notes, code, and analysis are released CC BY 4.0. Potency values are harmonised to pIC50 = −log10[M]; see scripts/external/pull_cyp_external.py for the exact conversion.

Assembled with substantial automation. Findings are logged with an explicit confidence level; treat anything marked speculative as a lead, not a result.