yw — OpenADMET CYP direct-inhibition method report
Updated: 2026-11-03T16:48:24.354536+00:00
Submission
This report describes the final public challenge submission under the anonymous leaderboard alias yw. It uses no proprietary training data and no external assay data. No code is published.
Prediction file SHA-256: 80695a19012518a3fbfd5ec3086dd112701b307b27abacba41b9472ced05ed0f.
The file combines, per enzyme, prediction columns from earlier model runs (no joint refit and no new training). Each enzyme's column was chosen by comparing the public live-leaderboard ST-RAE of earlier submissions that contained it. The live leaderboard covers only part of the blinded series, so these choices are tuned to that subset.
Live leaderboard result
Overall MA-ST-RAE of this file: observed 0.7404.
| Enzyme | Source run | Live ST-RAE of the column |
|---|---|---|
| CYP1A2 | 20261031T032022-cyp1a2-nested-support-conditioned-family-gate |
0.772 |
| CYP2C9 | 20261029T175417-cyp1a2-cyp3a4-fixed-descriptor-graph-ensemble |
0.5217 |
| CYP2D6 | 20261017T004159-cyp2d6-assay-tagged-nonhit-calibration |
1.172 |
| CYP3A4 | 20261029T175417-cyp1a2-cyp3a4-fixed-descriptor-graph-ensemble |
0.4958 |
Data and method of each source run
Each description is quoted from that run's own record. Only the enzymes listed for a run are taken from it, so statements about its other enzymes do not apply to this file.
20261031T032022-cyp1a2-nested-support-conditioned-family-gate (used for CYP1A2)
Model revision: c232bdc94f87c336f3253b2e51c52bae87474d32.
Method: CYP1A2 p=L+clip(0.5+Z beta,0,1)*(J-L), fixed eight-column support gate. Deterministic interval-error LP plus .01 L1 beta penalty, training weights constrained [0,1]. Full real-support gate is sole candidate; no control selection. Three other endpoints exactly F.
Features: intercept; m_z; d_z; abs_d_z; similarity_z; scaffold; m_zd_z; similarity_zd_z
Data sources: Challenge direct confidence intervals; Archived L tree and J descriptor-graph nested predictions; separately tagged TDI-condition screening supervision only in base models
Validation: Controller canonical random/scaffold five-fold, testlike holdout and archived Butina cluster five-fold. Gate training uses only each population inner OOF archive and permitted fitting labels; blinded fit uses full inner OOF. All 85 inner fits and 17 refits audited before gate training; no global outer OOF input.
20261029T175417-cyp1a2-cyp3a4-fixed-descriptor-graph-ensemble (used for CYP2C9, CYP3A4)
Model revision: 911812692f61783085b34085e0fb498683c28dcc.
Method: Fixed (L + GJ)/2 for CYP1A2 and CYP3A4; C CYP2C9/CYP2D6 copied exactly. GJ is the archived equal three-seed epoch-100 graph. No training, calibration, learned weights, checkpoint selection, or control substitution.
Features: Chemprop 2.3.1 depth-3 256-dimensional BondMessagePassing, default atom/bond features, fixed half-mean/half-componentwise-max pooling; X_d concatenation to 268-input two-layer eight-output RegressionFFN, hidden256, dropout0.1, no batch normalization; ['MolWt', 'MolLogP', 'TPSA', 'MolMR', 'NumHDonors', 'NumHAcceptors', 'NumRotatableBonds', 'RingCount', 'NumAromaticRings', 'FractionCSP3', 'HeavyAtomCount', 'NumHeteroatoms']; Historical LightGBM: chiral Morgan counts, RDKit descriptors, nested predicted TDI-condition screening responses
Data sources: Challenge direct-inhibition intervals; Separate TDI-condition single-concentration screen tasks; Hash-pinned completed L, C, and J archives; no new assay data
Validation: Unchanged canonical random/scaffold five-fold and testlike holdout; C fixed Butina cluster diagnostic. Every graph fit/held index and fitting-only descriptor/target normalization verified; all 153 saved inference arrays exactly reproduce seed-average exports. Historical stacking control flow and saved cluster screen memberships audited. TDI/Emax labels were not used in these components. All four predictions finite for every held row. No held assay enters a fitted component.
20261017T004159-cyp2d6-assay-tagged-nonhit-calibration (used for CYP2D6)
Model revision: 06e1089342f4ec8297830b103935684e73668812.
Method: Interval-bound LightGBM champion with a CYP2D6 assay-tagged screen nonhit probability correction
Features: Morgan radius-2 count fingerprints with chirality, RDKit descriptors, a CYP2D6 screen nonhit classifier and separate hit/nonhit L1 response models
Data sources: OpenADMET direct-inhibition training data (champion); OpenADMET single-concentration log2 fold change under TDI condition; Local screen-multitask-features champion predictions
Validation: Controller molecule-grouped random, scaffold, test-like direct splits and five-fold screen splits; all screen data from held-out canonical molecules excluded
Compounds were separated before fitting any model component or generating out-of-fold assay features. Direct, TDI, and screening readouts were treated as distinct assay observations. CYP2D6 direct measurements are selected from screening hits, limiting what random-fold direct validation establishes.