Can a molecule’s structure tell us how strongly it will inhibit a drug-metabolizing enzyme? The OpenADMET CYP Challenge supplies structures for 750 compounds while withholding their experimental answers. CYP enzymes are involved in the metabolism of many medicines. Predicting their inhibition is one part of studying drug interactions; it cannot by itself establish whether a medicine is safe in people.
I prepared predictions of direct inhibition for four CYP enzymes, along with time-dependent inhibition (TDI) classifications for two of them. My method report, published on 27 September, describes the models and files used for the 15 September submission. Later experimental models are kept out of that account.
An unmeasured value is not a zero
The official data assembled for production fitting contain 6,146 molecules, but not every molecule has every measurement. If a molecule has a measured CYP1A2 target and no value for another enzyme, that blank does not mean it failed to inhibit the second enzyme. I masked missing targets rather than filling them with zero or treating them as negative labels.
At prediction time, the inputs are identifiers and SMILES, a text representation of molecular structure. RDKit converts each structure into patterns of substructures and physicochemical descriptors. These molecular fingerprints let a model compare structural features. Branches that use auxiliary assays predict those signals from structure; they do not receive experimental assay values for the test compounds.
Combining six predictions with measurement uncertainty in view
The direct-inhibition predictor combines six branches: two LightGBM models using different molecular representations, a similarity-kernel regressor, a multi-assay neural network, a model that passes through predicted single-concentration assay signals, and a model called dose8 that relates potency to concentration-response information. The public report gives the branch settings and the weights used for each enzyme.
The combination weights are nonnegative and sum to one, with an intercept fitted separately for each enzyme. I fitted them using the distance beyond the experimental uncertainty interval. When the measurement supplies a plausible range rather than a single exact answer, a prediction inside that range is not penalized as an error against its midpoint. Accounting for measurement intervals does not resolve every source of uncertainty in the model.
For TDI, I retained the earlier five-branch classifier. Its outputs were calibrated on training out-of-fold predictions, and the positive/negative cutoff was selected from those predictions. An experimental TDI challenger considered during the regression update was not used in the submitted file.
Testing on groups of related molecules
A random split can put close structural relatives in both training and evaluation sets. Development used 5,378 molecules, with three repeated four-fold cluster splits and a separate three-fold scaffold split as a stress test. Combination weights, calibration, and classification thresholds were fitted on inner predictions within each outer training partition, keeping the outer evaluation molecules out of those fits.
The recorded development means were approximately 0.6490 for regression MA-ST-RAE and 0.2412 for TDI macro MCC. The first measures error relative to experimental uncertainty intervals; the second summarizes correct and incorrect binary classifications. Lower MA-ST-RAE indicates less error, while higher MCC indicates better classification. These are not official test scores. I explored and selected model families on the same development data, so nested fitting of the combination does not make the whole model-selection process an independent confirmatory evaluation.
Production fitting used the 6,146-molecule cohort while retaining the combination policy selected during development. Some molecules excluded from particular development fits therefore appear in the final fitting data. Applying the production model to those internal partitions would not constitute an unseen-data test.
What is public, and what remains unconfirmed
The public artifact is the method report. The code, fitted models, frozen policies, predictions, and validation records remain in a locally preserved package; I have not published that entire package. Preparing the report included recovering the original regression and classification files, checking that each contained 750 rows, and comparing their hashes, identifiers, and molecular ordering. It did not involve rerunning all historical training experiments.
On 27 September, I submitted the unchanged prediction files again with the report link and received acknowledgements from both submission screens. At that check, backend processing and replacement of the report links on the leaderboard were not independently confirmed. I therefore do not treat the acknowledgements as a new score or a confirmed ranking.
The report contains the exact settings, weights, runtime, and file hashes. The official tutorial explains the challenge and its evaluation rules. AI tools assisted with code development, experiment execution, and report preparation.