Predicting CYP Inhibition and Structures: What We Learned in OpenADMET

In OpenADMET, we tested ways to combine molecular models, learn from different assays, and predict protein–ligand structures. The clearest gains came from models that made different errors, small contributions from public data, and more structure samples.

We built models to predict inhibition, tested features based on chemistry, and compared training objectives. For structures, we ranked predicted poses by confidence and checked their geometry. This article explains what we tried, including promising ideas that did not improve predictions. [1][2][3]

Three tasks, three kinds of prediction

Cytochrome P450 enzymes help metabolize many drugs. Inhibiting one can change how another drug is processed. OpenADMET asks us to predict parts of this biology from molecular structure. [31][61]

Direct inhibition asks how strongly a molecule inhibits CYP1A2, CYP2C9, CYP2D6, or CYP3A4. We predict pIC50: a higher value means stronger inhibition. Time-dependent inhibition (TDI) asks whether inhibition increases after preincubation, using binary labels for CYP2D6 and CYP3A4. Structure prediction asks how 20 ligands fit inside CYP3A4. Below, D6 and C3 abbreviate CYP2D6 and CYP3A4. [2][3]

How to read the results. For potency, lower ST-RAE is better. Predictions inside the assay’s credible interval receive zero error. For TDI, higher MCC is better and accounts for both classes. For structures, lower ligand RMSD means a pose is closer to a known reference. The challenge’s LDDT-PLI instead measures agreement in protein–ligand interactions. The figures below show local comparisons and public controls, not final challenge standings. [3][48][59]

Activity models: combine complementary information

We started with Morgan fingerprints, RDKit properties, molecular similarity, and LightGBM. Fingerprints describe local fragments. Physical properties capture features such as size and polarity. Similarity models use measurements from related molecules. Chemprop graph networks learn directly from atoms and bonds. Combining these views was more useful than expecting one representation to capture everything. [14][15][16][17][18]

Our experiments compared single-task and multitask models, auxiliary assay endpoints, and separate blends for each enzyme. One early analogue holdout improved from 0.668 to 0.626 macro ST-RAE with an equal graph–classical blend. A 24-task model was weaker on its own but helped when added to the ensemble in a small amount. We therefore tested different combinations for each enzyme.

We tested which combinations of models, public assay data and chemistry-based corrections helped on new compounds. These comparisons shaped our enzyme-specific ensembles. The ensembles also included multitask and D6-specific Chemprop models reproduced from Workbench’s published approach. [5][6]

More data helped only when the measurements were compatible

We used public data to pretrain models, add related prediction tasks, and make small corrections where the same structures had public measurements. Some additions helped A2 and D6. Combining all sources was less reliable because assays used different protocols and measured different outcomes. In one test, correcting for differences between data sources improved the training fit but worsened predictions on held-out molecules. [20][21][22][37]

Chemistry was useful as a modest specialist signal

For D6, we grouped molecules using basic-nitrogen features and fitted global and specialist LightGBM models. They used absolute-error loss, 300 trees, learning rate 0.03, 15 leaves, minimum leaf size 20, and L2 regularization of 1. A 5% rank contribution within those groups provided a small correction while preserving the base prediction distribution. These groups encode structural rules, not calculated pKa values. [14][16][61]

The small correction helped a submitted comparison. Stronger mechanism features did not transfer, despite better local fits. Charge summaries, cation–aromatic pharmacophores, and predicting D6 through another CYP also failed their controls. These are useful negative results: a chemically plausible explanation does not automatically provide a better predictive feature.

Time dependence: model the outcome or the assays?

We treated D6 and C3 as separate classification problems. One route predicted the binary TDI label directly from fingerprints, descriptors and similarity. Another predicted direct and preincubation potency first, then used their difference and other assay features to infer TDI. We also tested maximum-inhibition features, public kinetic data, two-stage models, and multitask graphs. [25][26][27][30][31]

Dedicated classifiers were generally more useful than simple paired-assay rules. Predicting the assay measurements first added errors. An approach that helped one enzyme did not always help the other. Our later multi-assay graph study completed 40 fits, but neither the graph alone nor its features improved the four-model control on scaffold and analogue holdouts.

Longer-range molecular features offered a smaller, less certain signal. A functional-Morgan and atom-pair ensemble improved mean local MCC by 0.01168, with a 95% interval from −0.00226 to +0.02573. Because that interval includes no improvement, the experiment did not establish a reliable gain. [9][57]

The selected TDI combination uses our D6 calls, developed from classifier and similarity signals, and Jeremy’s publicly shared C3 predictions. Our independent C3 models did not surpass that public component in the relevant comparisons. [7][8]

Structures: generate, rank, then check

Our first structure workflow combined conformer search, ETKDG, MMFF and AutoDock Vina. A public CYP3A4 control exposed an important distinction: the pool contained a 1.58 Å pose, but Vina selected one at 5.58 Å. The search found a good candidate, but the scoring method did not select it. [47][50][51]

Our later workflow used Boltz-2 with the 470-residue CYP3A4 sequence from 5TE8 chain A, heme, and ligand SMILES. We compared five samples/three recycles with 25 samples/ten recycles. The larger run used 200 sampling steps. Its selected control pose improved from 3.87 Å to 1.44 Å RMSD. Sampling and recycling changed together, so the experiment does not separate their individual effects. [43][44]

We inspected poses in confidence order and selected the first with ligand ipTM of at least 0.5 that passed our PoseBusters checks. If none qualified, we retained a valid pose from the earlier run. The selected set contains 19 new poses and one earlier fallback. All 20 passed local physical checks. Chemical validity is a necessary check, not proof that a pose matches the unknown experimental structure. [3][48]

Additional contact and metal-coordination guidance did not reliably improve the controls. Some restraints recovered stereochemistry for one ligand while leaving another poorly placed. The useful change was broader sampling with a clear selection and validation procedure, rather than adding every available constraint.

Test the hypothesis on unfamiliar chemistry

We used scaffold, analogue and chemical-series holdouts, repeated splits, and grouped bootstrap comparisons. The challenge focused on expanding chemical series. Testing on new series was therefore important. Good predictions for close training neighbours may not carry over to a new family of compounds. [4][52][55][56]

Interval-aware losses, activity-cliff objectives and pretrained embeddings were reasonable hypotheses, but they did not reliably improve the tested controls. The experiment explorer shows whether each result came from a local test or a challenge submission. A method that helped in one setting may not help in another.

We learned to test the full workflow on new chemistry, from molecular features to the final prediction. Small additions of different information were often the most useful.