- CYP inhibition prediction achieves the strongest benchmark scores
- Phase II enzyme prediction lags behind CYP models
- Metabolite structure prediction remains the hardest task
- Benchmark performance varies sharply by modeling task
- Where ML metabolism models are used
- Key datasets remain limited for metabolite work
- Machine learning drug metabolism prediction still needs a confidence check
Machine learning drug metabolism prediction has become a routine early filter in medicinal chemistry, scoring how a candidate compound will be processed by cytochrome P450 (CYP) enzymes and Phase II conjugation pathways before a single metabolite is measured experimentally. Metabolic liability remains one of the most common reasons a chemically promising compound gets dropped, and models trained on large curated datasets are now sophisticated enough to influence which analogs a chemistry team synthesizes next.
Key takeaways
- Cytochrome P450 (CYP) inhibition models built on tree-based ensembles report area under the curve (AUC) values as high as 0.92 to 0.95 for major isoforms.
- Phase II enzyme prediction, including uridine 5'-diphospho-glucuronosyltransferase (UGT) substrate identification, has less public training data than CYP models and performs less consistently.
- Multitask graph neural networks now predict substrate status across five major P450 enzymes simultaneously with reported average AUC near 90.8%.
- Metabolite structure prediction remains the hardest of the three modeling tasks, since it requires generating plausible new chemical structures rather than classifying a known one.
- Model performance drops meaningfully when a compound falls outside the structural space the model was trained on.
CYP inhibition prediction achieves the strongest benchmark scores
CYP inhibition prediction is the most mature of the metabolism modeling tasks, benefiting from decades of accumulated inhibition data across the major drug-metabolizing isoforms: CYP3A4, CYP2D6, CYP2C9, CYP2C19, and CYP1A2. Pharmacokinetic and metabolic liabilities were once a leading cause of costly late-stage attrition, a share of failures that fell sharply after earlier screening became standard practice, and machine learning drug metabolism prediction pushes that screening earlier still. It sits within the wider shift toward target identification through clinical translation powered by AI, alongside the broader push toward AI-powered ADMET prediction that now anchors much of early-stage candidate triage, and the same models that flag a compound's metabolic risk also feed directly into how discovery teams weigh in silico ADMET predictions against in vitro assays before committing synthesis resources to a chemical series.
A 2024 evaluation of multiple machine learning algorithms found that XGBoost and CatBoost models achieved an AUC of 0.92 for CYP3A4, CYP2D6, and CYP2C9 inhibition when paired with combined fingerprint and physicochemical descriptor features, outperforming the deep learning architectures tested in the same comparison.
CYP2B6 prediction, a less studied but clinically relevant isoform, has also reached strong reported performance. One study found that its best CYP2B6 inhibitor model produced AUC values of 0.95 during cross-validation and 0.75 on an external test set, a gap that illustrates how cross-validation scores can overstate real-world reliability compared with genuinely external validation.
Rule-based and machine learning hybrid models such as the CYPlebrity models have also shown competitive external performance, with Matthews correlation coefficients ranging from 0.62 for CYP2C19 to 0.70 for CYP2D6, reflecting that some isoforms remain harder to model than others regardless of architecture choice.
This isoform-to-isoform variation matters practically because medicinal chemistry teams rarely care about a single average CYP inhibition score. A compound that scores well on aggregate but carries a specific CYP2D6 liability, an isoform with a well-documented history of clinically significant drug-drug interactions, may still warrant early prioritization for confirmatory assays even when its overall predicted profile looks favorable. Models that report isoform-specific confidence, rather than a single pooled prediction, give chemists more actionable information for exactly this reason.
Phase II enzyme prediction lags behind CYP models
Phase II metabolism, which includes glucuronidation, sulfation, and glutathione conjugation, has received comparatively less machine learning attention than CYP-mediated Phase I metabolism. Public training data for UGT and sulfotransferase substrates is sparser and less standardized than the CYP inhibition datasets that have driven most recent modeling progress, which limits how confidently a model can generalize across chemical classes.
Where Phase II models do exist, they tend to focus narrowly on predicting whether a compound is a substrate for a specific UGT isoform rather than attempting the broader multitask prediction that has become standard for CYP inhibition. This narrower scope reflects a genuine data limitation rather than a modeling choice, and it means Phase II predictions should generally be treated as lower-confidence signals than the equivalent CYP outputs, particularly for structurally novel scaffolds.
Discovery teams working with compound classes known to rely heavily on Phase II clearance, such as many nonsteroidal anti-inflammatory scaffolds, often supplement any available Phase II model output with earlier and more extensive in vitro glucuronidation screening rather than relying on prediction alone.
The data gap behind this limitation is structural rather than incidental. CYP inhibition assays have been run at industrial scale for decades because regulatory expectations around drug-drug interaction risk made them a near-universal part of preclinical screening, generating exactly the large, standardized datasets that machine learning needs. Phase II conjugation assays have historically been run more selectively, often only for compounds already flagged as having significant non-CYP clearance, which means the accumulated public dataset is both smaller and more biased toward compounds already suspected of Phase II involvement.
Metabolite structure prediction remains the hardest task
Predicting the specific chemical structure of a metabolite is a fundamentally harder task than classifying inhibition or substrate status, because it requires a model to generate a plausible new molecule rather than assign a label to an existing one. Site-of-metabolism prediction, which identifies which part of a molecule is most likely to be modified, has proven more tractable and is now reasonably reliable for common CYP-mediated reactions such as aromatic hydroxylation and N-dealkylation.
Full metabolite structure elucidation, by contrast, still generally requires combining computational prediction with confirmatory mass spectrometry, since predicted structures function best as a hypothesis-generation step that narrows the analytical search space rather than a standalone answer. A multitask deep learning study using fingerprints and graph neural networks improved substrate prediction across five major P450 enzymes, reporting an average AUC of 90.8% on the test set; that study focused on classifying which enzyme is likely to metabolize a given compound rather than generating metabolite structures, and separate structure elucidation work remains a distinct, comparatively less mature modeling task.
Benchmark performance varies sharply by modeling task
The following table summarizes reported benchmark ranges across the three core metabolism prediction tasks, synthesized from multiple independent studies rather than a single dataset.
| Metabolism prediction task | Reported benchmark range | Primary limitation |
|---|---|---|
| CYP inhibition (major isoforms) | AUC of 0.85 to 0.95 in cross-validation; external test performance can fall meaningfully lower for individual isoforms | External validation performance often trails cross-validation |
| CYP substrate prediction (multitask) | Average AUC near 90.8% | Performance varies notably by isoform |
| Phase II substrate prediction | Limited published benchmarks | Sparse, isoform-specific training data |
| Metabolite structure prediction | No standardized accuracy metric | Requires generative rather than classification modeling |
A recurring pattern across these benchmarks is that cross-validation scores, generated by holding out random subsets of the same dataset, consistently report higher performance than external test sets drawn from genuinely different chemical series. Teams evaluating a metabolism model for internal adoption should weight external validation figures more heavily than cross-validation scores when deciding how much to trust a prediction on a new chemical series.
The gap between the CYP2B6 model's 0.95 cross-validation AUC and its 0.75 external test AUC is a useful illustration of exactly this dynamic, and it is not unique to that particular model or isoform. Any evaluation of a metabolism prediction tool should specifically ask how the external test set was constructed, since a test set drawn from the same chemical series as the training data will produce an optimistic score that does not reflect how the model performs on a discovery program's actual, unseen chemistry.
Where ML metabolism models are used
Discovery teams typically apply metabolism prediction models at several distinct points in a program:
- Early virtual screening, to deprioritize entire chemical series carrying predictable CYP liabilities before synthesis.
- Structure-activity relationship analysis, to identify which specific structural features within an active series drive metabolic instability.
- Prioritization of confirmatory in vitro metabolic stability and CYP inhibition assays, focusing limited assay capacity on the compounds most likely to have a genuine liability.
- Late-stage risk assessment for lead candidates, where predicted metabolite structures help anticipate potential drug-drug interaction or toxicity concerns ahead of formal DMPK studies.
This staged use reflects the broader pattern across ADMET prediction: models are most valuable as a triage and prioritization layer rather than as a final determination, particularly for compounds outside well-represented chemical space.
Key datasets remain limited for metabolite work
Most published machine learning drug metabolism prediction models draw on a relatively small number of large, curated public datasets, supplemented in industry settings by proprietary internal assay data accumulated over years of medicinal chemistry programs. Public CYP inhibition datasets aggregating tens of thousands of compounds now exist for the major isoforms, and they have become the standard benchmark against which new modeling approaches are compared.
Metabolite identification datasets remain comparatively small and less standardized, since generating confirmed metabolite structures requires resource-intensive mass spectrometry work that has not been curated into public repositories at the same scale as inhibition assay results. This data gap is a primary reason metabolite structure prediction lags behind CYP inhibition prediction in reported accuracy, and it is likely to remain the binding constraint on progress in that specific modeling task for the foreseeable future.
Industry-generated proprietary data plays an outsized role in closing this gap for large pharmaceutical organizations, since decades of internal DMPK studies across many chemical series give bigger discovery groups a training data advantage that smaller organizations and academic groups cannot replicate from public sources alone. That asymmetry is part of why some of the more capable metabolism prediction platforms remain concentrated among a handful of large pharmaceutical companies and specialized software vendors, rather than being broadly accessible as fully open tools, even as the underlying published methods are shared in the scientific literature.
Machine learning drug metabolism prediction still needs a confidence check
Machine learning drug metabolism prediction has matured furthest for CYP inhibition, where large public datasets and years of iterative modeling have produced consistently strong benchmark performance across the major isoforms. Phase II enzyme prediction and full metabolite structure elucidation remain considerably less mature, constrained less by algorithm choice than by the depth and standardization of available training data.
Chemistry teams get the most reliable use out of these models by matching confidence to data depth: treating CYP inhibition scores as strong triage signals, treating Phase II predictions as lower-confidence indicators warranting earlier experimental confirmation, and treating metabolite structure predictions as hypotheses to be confirmed analytically rather than answers in their own right.
This article was produced under Drug Discovery News' AI Editorial Guidelines.
