- What ADMET prediction involves in drug discovery today
- Machine learning models built for each ADMET parameter
- Benchmark datasets show where performance holds up
- Where AI ADMET predictions are reliable enough to act on
- The novel chemical space challenge limits generalization
- Integrating ADMET AI into the discovery workflow
- AI ADMET prediction is redefining early candidate triage
AI ADMET prediction has moved from a screening curiosity to a routine checkpoint in early drug discovery, guiding which candidates advance before a single in vitro assay runs. Absorption, distribution, metabolism, excretion, and toxicity (ADMET) failures remain a leading cause of candidate attrition, and machine learning models built on curated chemical and biological data are now sophisticated enough to influence chemistry decisions, provided teams understand exactly where those predictions can be trusted.
Key takeaways
- AI ADMET prediction models now cover most absorption, distribution, metabolism, excretion, and toxicity (ADMET) endpoints, with performance varying widely by parameter.
- Cytochrome P450 (CYP) inhibition models routinely reach area under the curve (AUC) values near 0.90 to 0.92 on held-out test sets.
- Cardiotoxicity models targeting the human ether-a-go-go-related gene (hERG) channel report AUC-ROC scores above 0.92 under scaffold-split validation.
- Benchmark leaderboards such as the Therapeutics Data Commons have revealed reproducibility problems in several top-ranked ADMET models.
- Novel chemical scaffolds outside a model's training distribution remain the primary limitation on prediction reliability.
What ADMET prediction involves in drug discovery today
ADMET prediction spans five interconnected domains, each governed by distinct biology and modeled with different machine learning architectures. Absorption models estimate oral bioavailability and gut permeability, distribution models estimate plasma protein binding and tissue partitioning, and metabolism models flag which cytochrome P450 (CYP) enzymes will process or be inhibited by a compound.
Pharmacokinetic and toxicity issues once drove roughly 40% of clinical-stage drug attrition, a figure that fell sharply once companies adopted systematic early ADME screening. AI ADMET prediction extends that same discipline earlier still, scoring virtual compound libraries before synthesis and letting chemists deprioritize structures carrying predictable liabilities. That shift sits within a broader move toward target identification through clinical translation powered by AI, and it has made machine learning drug metabolism prediction one of the more mature applications of computational chemistry in the field, since metabolic stability and enzyme inhibition are among the ADMET parameters with the deepest public training data.
The interconnection between these domains is a persistent modeling challenge. A compound with excellent passive permeability may still show poor oral bioavailability once first-pass metabolism is accounted for, and a molecule that clears metabolism screens cleanly can still fail on plasma protein binding if the free fraction available to reach its target turns out to be too low. Predicting each parameter in isolation therefore only partially answers the question a discovery team actually needs answered, which is whether the whole ADMET profile, taken together, supports a viable dosing regimen.
Excretion models predict renal and biliary clearance, while toxicity models cover a wide range of endpoints, from acute cytotoxicity to organ-specific liabilities such as hepatotoxicity and cardiotoxicity. Each domain historically required separate wet-lab assays, and AI ADMET prediction consolidates much of that early triage into a single computational pass over a virtual library.
The consolidation matters because ADMET failures have historically clustered late in development, after significant capital has already been committed to a candidate. Screening in silico earlier shifts that cost curve, and it is the core rationale for why weighing in silico ADMET predictions against in vitro assays has become standard practice across large pharmaceutical discovery organizations rather than a niche computational add-on.
Several factors determine whether a given AI ADMET prediction is worth acting on for a specific discovery program:
- The size and structural diversity of the public or proprietary dataset a model was trained on for that endpoint.
- Whether reported performance comes from a random split, which tends to overstate reliability, or a scaffold split, which better approximates novel chemistry.
- How closely the compound being scored resembles the chemical space the model has already seen.
- Whether the endpoint has a clear biological mechanism that a model can learn, as opposed to idiosyncratic toxicity driven by off-target biology.
These factors recur across every ADMET parameter discussed below, and they explain why the same modeling approach can perform very differently depending on which property it is asked to predict.
Machine learning models built for each ADMET parameter
Different ADMET endpoints call for different modeling strategies, and the field has largely converged on architecture choices matched to data availability and endpoint complexity. Gradient-boosted tree methods such as XGBoost and CatBoost, paired with molecular fingerprints and physicochemical descriptors, remain competitive for CYP inhibition prediction, with one 2024 evaluation reporting an AUC of 0.92 for CYP3A4, CYP2D6, and CYP2C9 inhibition using tree-based ensembles, outperforming the deep learning models tested in the same study.
Graph neural networks and directed message-passing networks have made larger gains on endpoints with more structurally diverse training sets, including hERG channel blockade prediction and blood-brain barrier permeability, where three-dimensional and bond-level information appears to add real predictive value beyond flat descriptors.
For toxicity endpoints with sparse or imbalanced clinical data, such as drug-induced liver injury (DILI), decision forest and other tree-ensemble approaches trained on curated regulatory datasets have shown more consistent generalization than single deep learning models, in part because they degrade more gracefully when test compounds diverge from the training distribution.
Absorption and distribution parameters have received comparatively less architectural innovation than metabolism and toxicity, largely because passive permeability and plasma protein binding correlate strongly with well-understood physicochemical descriptors such as lipophilicity and molecular weight. Simpler regression models often perform competitively on these endpoints, which is a useful reminder that more complex architecture does not automatically translate into better predictions once an endpoint's underlying biology is already well captured by classical descriptors.
Benchmark datasets show where performance holds up
Public benchmarks have become the primary mechanism for comparing ADMET models on equal footing, and the Therapeutics Data Commons ADMET Benchmark Group now aggregates 22 standardized datasets spanning absorption, distribution, metabolism, excretion, and toxicity endpoints. That standardization has been valuable, but it has also exposed weaknesses in how leaderboard results are produced and reported.
A 2026 critical assessment of the Therapeutics Data Commons ADMET leaderboards found that only a handful of top-ranked methods passed reproducibility checks across all 22 endpoints, while several others showed unavailable code, broken execution environments, or direct data leakage between training and test sets. That finding is a useful caution for anyone treating a single leaderboard rank as proof of real-world reliability.
Data leakage in particular deserves attention because it is easy to introduce unintentionally and hard to spot from a leaderboard score alone. It typically occurs when structurally near-identical compounds appear in both the training and test partitions of a dataset, which lets a model perform well by memorizing close analogs rather than genuinely learning the underlying structure-property relationship. A discovery team evaluating an external ADMET model for internal use should ask specifically how the developer split training and test data, not just what the headline accuracy figure was.
The following table synthesizes reported benchmark performance across several ADMET endpoints, drawn from multiple independent studies rather than a single leaderboard snapshot.
| ADMET endpoint | Reported benchmark range | What drives the variation |
|---|---|---|
| CYP inhibition (CYP3A4, CYP2D6, CYP2C9) | AUC of 0.85 to 0.92 | Tree ensembles versus deep learning, descriptor choice |
| hERG channel blockade | AUC-ROC of 0.89 to 0.96 | Random split versus scaffold split validation |
| Drug-induced liver injury | Two-class accuracy roughly 70% to 75% | Dataset curation and class imbalance handling |
| Blood-brain barrier permeability | Accuracy of 0.90 to 0.98 | Passive permeability alone versus active transport features |
Reported numbers should be read as ranges rather than fixed accuracy claims, since cross-validation and scaffold-split results routinely diverge, and the scaffold-split figure is the more honest proxy for how a model will behave on genuinely novel chemistry.
Comparing these ranges against traditional screening timelines is instructive. A full in vitro ADMET panel across metabolism, toxicity, and permeability endpoints typically takes days to weeks per compound and consumes physical material that has to be synthesized first. A machine learning model scores an entire virtual library, often tens of thousands of structures, in a matter of hours, and it does so before a single milligram of compound exists. That speed advantage is precisely why benchmark reliability matters so much: a fast, wrong answer applied across a large library can eliminate promising chemistry just as easily as it can filter out genuine liabilities.
Where AI ADMET predictions are reliable enough to act on
Some ADMET endpoints have matured to the point where a discovery team can reasonably deprioritize a compound based on the prediction alone, at least as a triage step ahead of confirmatory assays. CYP inhibition and hERG blockade fall into this category for well-represented chemical space, given that both endpoints benefit from large, well-curated public datasets and years of iterative model refinement.
Blood-brain barrier permeability prediction has reached similarly strong reported performance, with one deep neural network model, the DeePred-BBB model, reporting accuracy above 98% on its validation set, and a separate quantitative model, LogBB_Pred, reporting 0.97 accuracy on an independent central nervous system (CNS) drug screening library.
Endpoints with sparser or more heterogeneous training data, including idiosyncratic hepatotoxicity and rare genotoxic mechanisms, remain far less reliable, and predictions there function better as prioritization signals than as pass or fail gates. A closer look at machine learning toxicity models and what they can and cannot tell a safety team is worth a dedicated read, since the coverage gaps differ meaningfully by endpoint. The practical distinction, covered in more depth in a companion decision-framework piece, is between endpoints where a model output can substitute for early experimental screening and endpoints where it should only reorder which assays run first.
The novel chemical space challenge limits generalization
Every published ADMET benchmark score describes performance on compounds structurally similar to the training set, and that caveat matters more than any single accuracy figure. When a discovery program moves into a genuinely novel scaffold series, model confidence typically degrades, even for endpoints that otherwise post strong benchmark numbers.
This gap explains why scaffold-split validation, which deliberately separates structurally distinct compounds between training and test sets, produces lower reported performance than random-split validation on the same dataset. A hERG model trained with directed message-passing networks reported its best-performing configuration at an AUC-ROC of 0.956 under random split but 0.922 under scaffold split, a meaningful drop that reflects how much harder true extrapolation is than interpolation within familiar chemical space, and that scaffold-split figure sits toward the strong end of published hERG models, several of which report scaffold-split performance in the high-0.70s to low-0.80s.
Practically, this means ADMET AI works best as a triage tool within chemical series a team already understands, and it should be applied with additional caution, and often paired with early experimental confirmation, whenever a program pushes into new scaffolds, novel target classes, or unusual physicochemical property space.
Model developers have started addressing this gap through active learning strategies, where a model flags the compounds it is least confident about and routes those specifically for experimental testing, then retrains on the results. That approach directly targets the novel chemical space problem rather than treating it as an unavoidable blind spot, and early adopters report that it narrows the gap between predicted and observed performance faster than passively accumulating more general training data.
Integrating ADMET AI into the discovery workflow
Most large-scale discovery organizations now embed AI ADMET prediction directly into compound design software, scoring virtual libraries before synthesis rather than waiting for physical compounds to reach an assay queue. That integration typically follows a staged approach:
- Score full virtual libraries against multiple ADMET endpoints simultaneously, using predictions to rank rather than eliminate candidates outright.
- Route top-ranked compounds through targeted in vitro confirmation for the endpoints carrying the highest developmental risk.
- Feed confirmed assay results back into the model's training set to narrow the gap between predicted and observed performance for that chemical series.
- Reserve full experimental ADMET panels for candidates approaching lead optimization or nomination.
The same staged logic extends beyond single-compound ADMET screening into predicting drug-drug interactions with AI, since polypharmacy risk assessment draws on many of the same enzyme-inhibition datasets used for metabolism modeling. This staged workflow treats prediction and experiment as complementary rather than competing sources of evidence, which is consistent with how regulators have approached computational toxicology more broadly. The FDA's AI development guidance emphasizes model transparency and validated performance rather than treating AI outputs as a wholesale replacement for experimental safety data, a framing that mirrors how regulators already treat quantitative structure-activity relationship (QSAR) models under the ICH M7 guideline for mutagenic impurity assessment.
AI ADMET prediction is redefining early candidate triage
AI ADMET prediction has earned a permanent place in early drug discovery, not because it has eliminated the need for experimental confirmation, but because it lets teams allocate that experimental capacity more deliberately. The endpoints with the deepest public data, including CYP inhibition, hERG blockade, and blood-brain barrier permeability, now support prediction-led triage with reasonable confidence.
The endpoints without that data depth still require the caution that has always governed early ADMET screening, and the novel chemical space problem means no benchmark score should be treated as a guarantee once a program moves into unfamiliar structural territory. Teams that treat AI ADMET prediction as a prioritization tool, rather than a final verdict, get the most reliable use out of it.
That distinction, between prediction as a triage step and prediction as a substitute for experimental evidence, is likely to remain the defining tension in AI ADMET prediction even as individual model architectures continue to improve. Benchmark scores will keep climbing on standardized datasets, but the harder problem, generalizing to genuinely new chemistry, is a data availability problem as much as a modeling one, and it will not be solved by architecture alone.
This article was produced under Drug Discovery News' AI Editorial Guidelines.
