How ADMET-AI results compare to the bench

You have an ADMET-AI result. Here is what it supports.

• Every number here was graded against a curated label, not against an assay. The comparison is to a dataset in the Therapeutics Data Commons (TDC) that someone else assembled from someone else's measurements. The assay is still the origin of the label, but it reaches the model through inclusion criteria, activity thresholds, and whatever chemistry happened to be screened.
• The toxicity classifiers are usable as triage signals. On TDC's scaffold-split test sets: hERG AUROC 0.840, Ames 0.850, DILI 0.887. Each of those is one dataset's view of one endpoint, not the endpoint.
• Do not read the pharmacokinetic regression outputs as quantities. On the same benchmark, half-life reaches Spearman ρ = 0.239 and volume of distribution ρ = 0.389. On the authors' own random-split analysis, the R² for both is negative — worse than predicting the dataset mean.
• Nothing here replaces an assay. Confirm hERG with patch clamp, mutagenicity with the bacterial reverse mutation (Ames) test, stability with microsomal or hepatocyte incubation, permeability with Caco-2. ADMET-AI is not a model qualified under ICH M7 or any other guideline that admits computation in place of a test.
• It predicts no off-targets and no metabolites. For a selectivity read, see DeepPurpose — which has no published comparison against panel screening either.

Jump to agreement with wet lab or against other in silico methods. The general framing for every tool is in reading an in silico result against the bench.

How in silico is used currently

Broad ADMET prediction is used to order a list, not to answer a question. In a live program that means three things:

• Library triage. You have more compounds than you can buy, make, or assay, and you want the ones with obvious liabilities pushed down the list. This is the use the tool was built for: the source paper frames it around high-throughput docking campaigns and generative chemistry that produce more hits than any assay queue can absorb.
• Multi-parameter flagging at hit triage. A hERG or DILI probability near the top of its range is a reason to look harder at a compound, or to put the assay earlier in the plan — not a reason to call it toxic.
• Features for a downstream model. The most informative published extension of ADMET-AI does not use its endpoints as verdicts at all. It feeds all 41 predictions into a gradient-boosting classifier trained on clinical cardiotoxicity labels, and reports the composite model's performance rather than any single endpoint's.

What the published literature does not contain is ADMET-AI used as a decision instrument and then graded against a fresh measurement. Almost every paper citing it uses it as a profiling stage inside a computational pipeline — docking, dynamics, then an ADMET filter — and reports no assay for the predicted properties. That absence is the central fact on this page.

Against other in silico methods

The comparison most readers assume was made was not made. The ADMET-AI paper compares itself to SwissADME, ADMETLab 2.0, pkCSM, vNN-ADMET, ADMETboost and PreADMET on speed only — a 45% reduction in time against the next fastest server for 1000 molecules. Its accuracy comparison is against other machine-learning models on the TDC leaderboard. There is no published head-to-head of ADMET-AI against ADMETlab, pkCSM or SwissADME on a shared endpoint with a shared test set. The nearest thing is the cardiotoxicity letter above, and even there the contest is between ADMET-AI-derived features (0.72 ROC AUC) and SwissADME-derived features (0.67) feeding the same downstream classifier — not between the two tools on an endpoint either of them predicts.

"Highest average rank" is an average, and it hides the spread. As of the leaderboard snapshot the paper reports (4 October 2023), ADMET-AI ranked 1st on lipophilicity and aqueous solubility, 2nd on Ames and microsomal clearance — and 10th on volume of distribution and P-gp inhibition, 11th on blood–brain barrier, and 7th on half-life. On VDss the best leaderboard entry reported Spearman 0.627 against ADMET-AI's 0.389. A tool can hold the best average rank and still be a poor choice for the specific endpoint you care about.

That leaderboard is itself under independent challenge. A 2026 audit took the top three models on all 22 TDC ADMET leaderboards and tried to reproduce them from their public artifacts alone. Only three models — CaliciBoost, MapLight and MapLight + GNN — passed every stage; the rest failed on unavailable code, non-reproducible environments or methodological flaws, and direct or indirect data leakage was traced in three of them. The authors conclude that leaderboard position cannot be read as a measure of model quality. ADMET-AI is not among the models that audit examined, so this is not a finding about ADMET-AI — it is a finding about the benchmark every figure on this page comes from, and a reason to treat a leaderboard rank as weaker evidence than a per-endpoint number with its n.

Other models on the same datasets, which is not the same as a head-to-head. Two independent groups report models on TDC ADMET data with scaffold-based validation. ADMET-Vault covers 12 TDC datasets and reports classification AUROC between 0.87 and 0.99, regression MAE from 0.28 (Caco-2, log cm/s) to 7.68 (plasma protein binding, % bound), and Spearman between 0.44 and 0.66. A cross-endpoint stacked ensemble reports six TDC pharmacokinetic datasets at the same sizes used here (HIA n = 578, Caco-2 n = 906, VDss n = 1111, PPBR n = 1614, hepatocyte clearance n = 1020, half-life n = 665), reaching HIA AUROC 0.935, Caco-2 MAE 0.386, VDss ρ 0.71, PPBR MAE 9.3%, clearance ρ 0.34 and half-life ρ 0.49. Those are different protocols and different splits, so the numbers are not directly comparable to this platform's own figures in agreement with wet lab — but the direction is worth noting: on VDss and half-life, ADMET-AI's two weakest endpoints, both studies report higher rank correlation than ADMET-AI does.

The off-target axis has not been measured for anything on this platform. The other wet-lab comparison for this class of tool is predicted off-target activity against a secondary-pharmacology panel screen. That comparison exists in the literature only inside pharma, against proprietary panels: the published example trains on an in-house 50-assay off-target panel covering roughly 4000 compounds. ADMET-AI predicts no off-targets at all — its nearest outputs are the 12 Tox21 nuclear-receptor and stress-response assays, which are a toxicity screen rather than a selectivity panel. DeepPurpose, which provides the kinome-wide read on this platform, has no published comparison against panel screening.

Two alternatives are gone rather than beaten. ADMETlab 3.0 is deprecated here because the external service is unavailable, and DeepChem is retired — neither was removed on the basis of a comparative evaluation, so do not read their absence as a result.

Koleiev I, Stratiichuk R, Shevchuk N, Melnychenko M, Nyporko A, Todoryshyn D, Husak V, Starosyla S, Yesylevskyy S, Nafiiev A. An end-user audit of reproducibility, data leakage, and overfitting of the top-ranked ADMET prediction models in TDC leaderboards. Journal of Chemical Information and Modeling. 2026;66(14):8045-8058. doi: 10.1021/acs.jcim.6c00819

Vishwakarma R, Lokhande V, Punia P, Mishra A. ADMET-vault: an interactive framework for real-time ADMET prediction and molecular optimization. Journal of Computer-Aided Molecular Design. 2026;40(1):129. doi: 10.1007/s10822-026-00816-3

Agboola OE, Agboola SS. Artificial intelligence-driven cross-endpoint stacked ensemble learning for separate but meta-feature coupled prediction of six human pharmacokinetic endpoints with RDKit molecular representation and applicability domain control on real TDC benchmark data. Naunyn-Schmiedeberg's Archives of Pharmacology. 2026. doi: 10.1007/s00210-026-05785-w

Naga D, Muster W, Musvasva E, Ecker GF. Off-targetP ML: an open source machine learning framework for off-target panel safety assessment of small molecules. Journal of Cheminformatics. 2022;14(1):27. doi: 10.1186/s13321-022-00603-w

In silico vs. wet lab

There is no single standard wet-lab technique on this axis. Each predicted endpoint has its own assay, and they are not comparable to each other: hERG is automated or manual patch clamp, Ames is the bacterial reverse mutation test (OECD 471), metabolic stability is a liver microsome or hepatocyte incubation, permeability is a Caco-2 or PAMPA monolayer, plasma protein binding is equilibrium dialysis or ultracentrifugation. That is the first way this page differs from a structural tool's: a docking pose has one ground truth (the co-crystal) and one settled agreement threshold (RMSD ≤ 2 Å). A property predictor has neither.

And the grading is one step removed from the bench. ADMET-AI was never compared to an assay it ran against. It was trained and tested on TDC datasets, so an AUROC here measures how well the model separates molecules that a curator labelled active from molecules that curator labelled inactive. Twelve of the 31 classification endpoints are the Tox21 nuclear-receptor and stress-response assays; the much larger ToxCast collection (117 endpoints) was deliberately excluded as too hard to interpret, and two further hERG datasets were dropped as redundant with the one that was kept. Your hERG probability is therefore a prediction about that dataset's definition of a hERG blocker. An AUROC is a property of the model and of the label set.

Read the accuracy column by cell, never down the column. An assay result and an AUROC against assay-derived labels are not the same kind of number, and an AUROC and a Spearman correlation are not either. What the table supports is comparing time and cost at a stated accuracy — minutes for a million molecules against days per compound per assay, at an accuracy that is good enough to sort a list and not good enough to decide one compound.

Where computation does replace the assay, it is not this kind of model. The ICH M7(R1) guideline admits two complementary (Q)SAR models in place of the bacterial mutagenicity assay for drug impurities. On an external validation set of 388 proprietary drug impurities (72 active, 316 inactive), the FDA-evaluated composite models reached 66–82% sensitivity, 91–95% negative predictivity and 96% coverage — and the reason that substitution is permitted is a documented validation exercise against an external, proprietary set, under a guideline, with negative predictivity as the controlling statistic. ADMET-AI's Ames AUROC of 0.850 on a public dataset is not the same claim and carries no equivalent standing.

Landry C, Kim MT, Kruhlak NL, Cross KP, Saiakhov R, Chakravarti S, Stavitskaya L. Transitioning to composite bacterial mutagenicity models in ICH M7 (Q)SAR analyses. Regulatory Toxicology and Pharmacology. 2019;109:104488. doi: 10.1016/j.yrtph.2019.104488

Agreement with wet lab

All figures below are author-reported, from the ADMET-AI paper and its supplement. They are reported in two regimes that give materially different numbers, and knowing which one you are reading is the single most important thing on this page.

The leaderboard regime (scaffold split). For the 22 datasets in the TDC ADMET Benchmark Group, five models were trained on Murcko scaffold splits and all were evaluated on the single scaffold-split test set TDC provides. Scaffold splitting is the realistic setting: the model must generalise to chemical series it never saw.

Classification endpoints, AUROC where TDC specifies it, mean ± standard deviation across the five seeds:

CYP inhibition is scored by AUPRC rather than AUROC on this benchmark: CYP3A4 inhibition 0.876 ± 0.003 (n = 12 328) and CYP2D6 inhibition 0.673 ± 0.007 (n = 13 130).

Regression endpoints, with TDC's chosen metric:

Classification and regression translate differently, and the split is obvious once the table is laid out. The classifiers that work are the ones with a large, chemically diverse label set and a crisp label. The regression endpoints that work are physicochemical — solubility, lipophilicity, permeability — where structure largely determines the answer. Every endpoint that requires a whole organism to produce the number (volume of distribution, half-life, clearance) is weak, and half-life is the weakest thing the model reports.

The random-split regime, and why the headline numbers are higher. Across all 41 TDC datasets the paper also reports an 80/10/10 random split with five seeds, and this is where the often-quoted summary comes from: R² > 0.6 for 5 of 10 regression datasets, AUROC > 0.85 for 20 of 31 classification datasets. The authors state the caveat themselves — random splitting improves generalisation of the ensemble "but at the potential cost of overly optimistic test set performance due to scaffold sharing between the train and test sets." The size of that effect is visible at hERG: 0.921 ± 0.025 single-task on a random split against 0.840 ± 0.007 on the scaffold split. Both numbers are real and they measure different things. The scaffold number is the one that describes a new series.

Two further things the headline hides. The five regression datasets that clear R² > 0.6 are Caco-2 permeability, hydration free energy, lipophilicity, aqueous solubility and LD50 — all five are physicochemical or acute-toxicity endpoints, and not one pharmacokinetic endpoint is among them. And two of the ten are negative: on the multi-task model, VDss R² = −0.610 and half-life R² = −1.552, with the single-task models at −0.749 and −2.183. A negative R² means the prediction is worse than reporting the dataset's mean.

Against a clinical label, two steps removed. The one published extension that puts ADMET-AI's outputs against an outcome rather than an assay feeds all 41 predictions into an XGBoost classifier trained on 555 drugs from DICTrank (293 "most concern" for drug-induced cardiotoxicity, 262 "no concern"), evaluated by 10-fold cross-validation with scaffold-based splits. It reached 0.72 ROC AUC and 0.75 PR AUC. The label there is derived from FDA drug labelling, not from a cardiac assay, so this is further from the bench than the hERG number, not closer to it. It is also author-reported by the same group, with a declared Greenstone Biosciences interest.

Swanson K, Walther P, Leitz J, Mukherjee S, Wu JC, Shivnaraine RV, Zou J. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics. 2024;40(7):btae416. doi: 10.1093/bioinformatics/btae416

Mukherjee S, Swanson K, Walther P, Shivnaraine RV, Leitz J, Pang PD, Zou J, Wu JC. ADMET-AI enables interpretable predictions of drug-induced cardiotoxicity. Circulation. 2025;151(3):285-287. doi: 10.1161/CIRCULATIONAHA.124.070413

Huang K, Fu T, Gao W, Zhao Y, Roohani Y, Leskovec J, Coley CW, Xiao C, Sun J, Zitnik M. Therapeutics Data Commons: machine learning datasets and tasks for drug discovery and development. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks. 2021. arXiv: 2102.09548 · https://tdcommons.ai

What the figure doesn't cover

• Applicability domain is the dominant caveat, and no published study measures it for this model. A benchmark AUROC does not predict performance on a scaffold series absent from training, and the scaffold-split figures above are the closest available proxy — they measure generalisation to held-out scaffolds inside TDC's chemistry, not to chemistry TDC never contained. No external-validation or prospective-test study of ADMET-AI exists, so there is no figure to quote here. The gap between the random-split and scaffold-split hERG numbers (0.921 against 0.840) is the best indication of the direction, and it is a lower bound on the effect, not an estimate of it.
• The DrugBank percentile is context, not confidence. ADMET-AI contextualises each prediction against 2579 approved drugs, optionally filtered by ATC code. That tells you where your molecule sits among marketed drugs on that property. It does not tell you whether the model has seen anything resembling your molecule, and it is not an applicability-domain estimate.
• AUROC flatters the imbalanced endpoints. Several toxicity datasets have very few positives, and their precision-recall performance is much weaker than their AUROC suggests. On the multi-task model: NR-PPAR-gamma (2.9% positive) AUROC 0.884 but AUPRC 0.373; aromatase (5.2% positive) AUROC 0.874 but AUPRC 0.454; ClinTox (7.0% positive) AUROC 0.870 but AUPRC 0.425. If you intend to act on positive calls, AUPRC is your statistic and it is considerably lower.
• Small label sets behind some headline endpoints. DILI's AUROC of 0.887 rests on n = 475; the carcinogenicity dataset has n = 278; skin reaction n = 404. A confident-looking number on a few hundred molecules describes a few hundred molecules.
• Version and date. The figures here are for TDC v0.4.1, Chemprop v1.6.1 and RDKit 2023.3.3, with leaderboard ranks as of 4 October 2023. A rank claim from that snapshot does not describe the leaderboard today, and a figure quoted without its version is unusable.

• No prospective validation, for this model or generally in this class. Two reviews are useful for framing and for nothing else: one on AI-driven toxicogenomics, which notes how few such models have been prospectively validated, and one surveying AI in safety pharmacology. Neither supplies a concordance figure for any tool, and neither is a source for a number about ADMET-AI.
• It does not predict metabolic pathways, metabolite structures, or drug–drug interactions. It predicts CYP inhibition and substrate probabilities, which are inputs to a DDI assessment, not a DDI assessment.
• No renal clearance. The clearance endpoints are hepatocyte and microsomal; there is no renal clearance model.
• Half-life is reported in hours and clearance in assay units, which makes them look like measurements. They are the weakest outputs on the page. See outputs and interpretation.

Islam MM, Ali SSM, Meravanige G, Nagaraja S, Chohan MS, Anwer MK, Gurung A. Artificial intelligence-driven toxicogenomics for predictive environmental and pharmaceutical safety. Frontiers in Genetics. 2026;17:1858338. doi: 10.3389/fgene.2026.1858338

Miller L, Treleaven I, El Amrani AI, Rossman EI, Winters BR, Pugsley MK. The emerging use of artificial intelligence in safety pharmacology and toxicology. Journal of Pharmacological and Toxicological Methods. 2026;139:108424. doi: 10.1016/j.vascn.2026.108424

Go / no-go

Go — act on the result when:

• You are ordering a library and the question is which compounds to look at first. Relative rank over a large set is what the benchmark figures support.
• You are deciding when to schedule an assay, not whether to trust its result. A high hERG or Ames probability is a reason to pull that assay earlier.
• The endpoint is physicochemical — aqueous solubility, lipophilicity, Caco-2 permeability — where the model is strongest and the underlying property is largely structural.
• You are feeding the full 41-endpoint vector into a downstream model trained on your own labels, which is the pattern the cardiotoxicity work demonstrates.
• Your chemistry resembles the TDC training sets, and you treat the DrugBank percentile as a sanity check on where the molecule sits rather than as a confidence score.

No-go — do not act on the result when:

• You need a number rather than a ranking for a pharmacokinetic property. Half-life (ρ = 0.239) and volume of distribution (ρ = 0.389, negative R² on the authors' own random split) do not support a quantity.
• You are deciding to drop a compound on a toxicity probability alone, especially on an imbalanced endpoint where AUPRC is far below AUROC.
• Your series has no close relative in the training chemistry. Nothing published measures performance there, and the scaffold-split numbers are a lower bound on the degradation.
• The output is headed for a regulatory submission. ICH M7 admits specific, separately validated (Q)SAR models for one endpoint; this is not one of them.
• The question is selectivity or off-target liability. ADMET-AI has no off-target panel, and the platform's off-target route has no published comparison against panel screening.
• You intend to skip the assay. No figure on this page licenses that, for any endpoint.

Caveats on reading this page: every figure above is quoted from the cited source with its original metric, dataset and sample size. Figures from different studies are not directly comparable, because label sets, splits and metrics differ between them. Where a comparison does not exist in the literature, this page says so rather than substituting a figure from a different model or a different endpoint.