Beta
Beta
Independent Comparison
A side-by-side comparison of computational ADMET prediction platforms available in 2026. We include our own product (ToxScreen) in these tables, and we flag where each tool — including ours — falls short. All benchmark numbers link to published or independently reproducible data. We do not claim to be the best tool for every use case, and we tell you where another tool may serve you better.
The table below summarises the key practical attributes of each tool. The sections that follow provide the empirical basis and context for each row.
| Feature | ToxScreen | SwissADME | pkCSM | ADMETlab 2.0/3.0 | ProTox-II | ADMET-AI |
|---|---|---|---|---|---|---|
| Model engine | Chemprop GNN (ADMET-AI) + isotonic calibration | Fragment-based + ML (proprietary) | Graph-based signatures (proprietary) | Random forest + DNN (hybrid) | Ensemble ML (XGBoost + fragment similarity) | Chemprop GNN (open-source) |
| Endpoints | 4 core (calibrated) + 15+ extended ADMET (raw) | 40+ (ADME descriptors + pharmacokinetic) | 30+ (ADMET + toxicity) | 40+ (ADMET + toxicity + physicochemical) | 33 toxicity endpoints | 40+ raw ADMET (no hERG filter split) |
| Calibrated probabilities | ✓ Isotonic per-endpoint | ✗ Raw classifier output only | ✗ Raw classifier output only | ✗ Raw scores | ✗ Raw probability-like scores | ✗ Raw classifier output |
| Applicability-domain abstention | ✓ Tanimoto k-NN, 3 bands (IN/BORDERLINE/ABSTAIN) | ✗ Not published | ✗ Not published | ✗ Not published | ✗ Not published | ✗ Not published |
| REST API | ✓ Single + batch, Python SDK + CLI + cURL | ✗ Web only (individual submission) | ✗ Web only (individual submission) | ⚠ API exists (docs gated behind login) | ✗ Web only | ✓ Open-source CLI (self-host) |
| Batch screening | ✓ CSV upload, web + API | ✗ Single compound per form | ✗ Single compound per form | ✓ Web batch (free) | ✗ Single compound per form | ✓ CLI batch (self-host) |
| Wilson confidence intervals | ✓ 80% + 95% per endpoint | ✗ | ✗ | ✗ | ✗ | ✗ |
| Deterministic (same input = same output) | ✓ | ⚠ Generally but undocumented | ⚠ Generally but undocumented | ⚠ Generally but undocumented | ⚠ Generally but undocumented | ✓ Deterministic with fixed seed |
| Export formats | PDF, XLSX, HTML, JSON, CSV | Text, PNG (BOILED-Egg) | Text | CSV | PDF, CSV | CSV (command-line redirect) |
| Pricing model | Free tier 3 free + $49–$999/mo | Free | Free | Free | Free | Open-source Free (self-host) |
| Published benchmark transparency | ✓ Full calibration data + all limitations published | ⚠ Original publication only | ⚠ Original publication only | ⚠ Benchmark paper, no ongoing validation | ⚠ Original publication only | ✓ GitHub + publication |
Side-by-side comparison of published or independently verifiable discrimination (AUROC) for the four safety endpoints both ToxScreen and ADMET-AI cover. SwissADME, pkCSM, ADMETlab, and ProTox-II do not publish per-endpoint AUROC for these exact targets on standardised benchmarks, so we report the available published data and note where it is missing.
Important caveat on cross-tool comparison. Each tool is trained and tested on different datasets with different label definitions, compound filters, and data-splitting strategies. The numbers below should be read as broad capability indicators, not as directly comparable apples-to-apples scores. An exact head-to-head benchmark on an identical held-out test set has not been published for any pair of these tools.
| Endpoint | ToxScreen (calibrated, in-distribution) | ToxScreen (de-leaked novel chemistry) | ADMET-AI (raw, in-distribution) | SwissADME | pkCSM | ADMETlab |
|---|---|---|---|---|---|---|
| hERG | 0.69 (random split, FAIL) | 0.674 (ChEMBL cross-dataset, n=8,886) | ~0.71 (Karim de-leak) | Not published | Not published* | Not published* |
| CYP3A4 inhibition | 0.95 (in-distribution) | 0.473 (ChEMBL cross-dataset, n=701) | ~0.95 (in-distribution) | Not published | Not published* | Not published* |
| CYP2D6 inhibition | 0.94 (in-distribution) | 0.714 (but 99% single-class test set, n=1,522) | ~0.94 (in-distribution) | Not published | Not published* | Not published* |
| CYP2C9 inhibition | 0.94 (in-distribution) | 0.538 (ChEMBL cross-dataset, n=237) | ~0.94 (in-distribution) | Not published | Not published* | Not published* |
* pkCSM (Pires et al., 2015) reports Matthews correlation coefficient (MCC) and accuracy on a self-constructed validation set, not AUROC. ADMETlab (Xiong et al., 2021) reports AUC for some endpoints on their own split but not publicly for the TDC Veith benchmark. SwissADME does not report per-endpoint binary classification metrics.
In-distribution AUROC measures how well a model separates positives from negatives when tested on a random or scaffold split of the same dataset it was trained on. This is what ADMET-AI and ToxScreen both report for the Veith CYP panel (0.94–0.95). It is the standard machine-learning publication metric, but it overstates real-world performance because the split compounds are still chemically similar to the training set.
Cross-dataset (de-leaked) AUROC measures performance on compounds from an independent source that were explicitly verified not to appear in the training data. This is a harder and more honest test. For hERG, ToxScreen's de-leaked score (0.674, ChEMBL pull) is slightly below ADMET-AI's Karim-pull number (0.71); we attribute the gap to different assay technologies (ChEMBL's diverse binding vs. Karim's focused patch-clamp references). For CYP3A4 and CYP2C9, the de-leaked numbers (0.47, 0.54) are statistically indistinguishable from coin-flips — a finding we have published transparently and are actively investigating.
No other tool in this comparison publishes cross-dataset de-leaked numbers for CYP inhibition. We cannot say whether SwissADME, pkCSM, or ADMETlab generalise better or worse to novel CYP chemistry, because none of them has published this measurement.
A classification model's raw score is not a probability. A score of 0.7 from an uncalibrated classifier does not mean "70% likelihood of toxicity" — it means the compound sits on one side of a decision boundary at an arbitrary threshold. Calibration maps model output to empirical likelihood. This matters practically: an uncalibrated score of 0.7 for hERG could correspond to a true positive rate anywhere from 0.3 to 0.9 depending on the model and chemistry.
| Property | ToxScreen | SwissADME | pkCSM | ADMETlab | ProTox-II | ADMET-AI |
|---|---|---|---|---|---|---|
| Output type | Calibrated probability [0,1] + CI | Categorical + descriptors | Binary + continuous descriptors | Score [0,1] + qualitative call | Probability [0,1] | Raw probability [0,1] |
| Calibration method | Isotonic regression (per-endpoint, ~1,800 cmpds each) | None | None | None published | None published | None |
| ECE (expected calibration error) on novel chemistry | 0.08 (hERG, de-leaked) | Unknown | Unknown | Unknown | Unknown | >0.30 (estimated, uncalibrated) |
Expected Calibration Error (ECE) is the average absolute difference between predicted probability and observed frequency. An ECE of 0.08 means the model's probabilities are off by about 8 percentage points on average. An ECE of 0.32 (ADMET-AI uncalibrated on novel hERG) means a score of 0.5 could correspond to a real rate anywhere from 0.2 to 0.8. ToxScreen's calibration reduces this gap by 4x, but does not eliminate it — the 80% and 95% Wilson confidence intervals in every report reflect the remaining uncertainty.
Strengths: Excellent physicochemical descriptor panel (40+ properties). The BOILED-Egg visualisation for BBB penetration and GI absorption is genuinely useful and widely cited. Free, no registration required. Fast web interface for one-off checks. Good LogP model (iLOGP).
Limitations: No toxicity prediction beyond PAINS alerts. No batch mode (one compound at a time, manually). No API — cannot be integrated into a pipeline. No probability calibration or applicability-domain documentation. Predictions are categorical (high/low) or qualitative, making risk-rank ordering of a hit list impractical.
Best for: Quick medicinal chemistry sanity checks on individual compounds. Lead optimisation where the question is "is this drug-like?" rather than "is this toxic?"
Strengths: 30+ ADMET endpoints including some ToxScreen does not currently cover (e.g. hERG pIC50 regression, AMES, skin sensitisation, mitochondrial membrane potential). Graph-based signatures capture molecular topology without needing 3D conformers. Free web server. Simple output.
Limitations: No API or programmatic access. No calibration — scores are raw classifier outputs. No applicability-domain check. Reported performance (MCC/accuracy) is on its own validation set, not an independent benchmark. Last major update 2015–2016; the underlying model may not reflect recent data (ChEMBL 20 vs. ChEMBL 35+ today).
Best for: Small-scale academic projects requiring broad endpoint coverage. Useful as a second-opinion comparator alongside a calibrated tool when time and budget permit in vitro follow-up.
Strengths: Comprehensive endpoint panel (40+). Free web batch mode (up to 5,000 compounds at once). Clean interface with traffic-light visualisation. Caco-2 permeability and oral bioavailability predictions are among the most-used in the field. ADMETlab 3.0 added a subscription API tier (paid). Good for large-library triage where per-compound web submission is impractical.
Limitations: API required to access batch mode programmatically; documentation is behind a login wall. No calibration published — the platform returns a score labelled "probability" but no calibration plot, calibration curve fit, or ECE is published. No applicability-domain check or abstention. The model architecture (random forest + deep neural network hybrid) is not fully reproducible from the published description. Prediction consistency (same compound, same model, different batch) is not documented.
Best for: Large-library initial triage where batch processing is more important than per-call calibration. The free batch mode is genuinely useful and is the closest free equivalent to ToxScreen's batch API — though without calibration, confidence intervals, or abstention.
Strengths: 33 toxicity endpoints including immunotoxicity, cytotoxicity, and ecotoxicity not covered by other tools. Includes molecular target prediction based on 2D fragment similarity, which can suggest mechanism. Free, straightforward web interface. Oral toxicity LD50 prediction is a unique feature.
Limitations: Web-only, single-compound submission. No calibration, no applicability domain. The ensemble model (XGBoost + fragment similarity) has known weaknesses on novel scaffolds — the fragment-similarity component effectively memorises training-compound fragments, so predictions on chemistry without close neighbours in the training set are unreliable. The "prediction accuracy" metric reported (67–75%) is on a balanced random split, which overstates performance on real-world imbalanced screening libraries.
Best for: Toxicity endpoint coverage beyond the standard ADMET panel (e.g. ecotoxicity). Quick LD50 estimation for safety data sheet compilation. Not suitable for pipeline integration or calibrated risk-rank-ordering.
Strengths: Open-source chemprop GNN with excellent code quality (GitHub, MIT license). 40+ ADMET endpoints trained on TDC Veith data. Deterministic when seeded. CPU-inference capable. Batch CLI works on any Linux/macOS machine. This is the model that powers ToxScreen's core inference — the raw ADMET-AI probability is the starting point before ToxScreen applies calibration, abstention, and confidence intervals.
Limitations: Command-line only — no hosted API, no web interface, no export formatting. No calibration — the output is raw classifier probability that systematically overstates confidence on novel chemistry (ECE >0.30 for hERG). No applicability-domain abstention. Requires Python environment and dependency management (conda/pip). The model was trained on TDC Veith only; cross-dataset generalisation is not measured in the publication.
Best for: Teams with Python/ML expertise who want full control over the inference pipeline and can implement their own calibration and abstention layer. Academic research groups who need a reproducible, audit-friendly pipeline for small-scale screens.
Strengths: Calibrated probabilities with Wilson confidence intervals and applicability-domain abstention — the only tool in this comparison that tells you when not to trust the answer. REST API with Python SDK, CLI, and cURL support. Deterministic CPU inference (0.5–7 s/compound). Full report export (PDF, XLSX, HTML, JSON, CSV). Batch CSV upload. Published benchmarks including the failing ones. Free tier available (3 screenings, no credit card).
Limitations: Only 4 calibrated core endpoints (hERG + CYP3A4 + CYP2D6 + CYP2C9) — fewer than ADMETlab (40+) or SwissADME (40+). The 15+ extended ADMET endpoints are raw ADMET-AI probabilities without ToxScreen calibration (clearly labelled in reports). hERG discrimination is weak AUROC 0.69 on random split (improved to 0.844 de-leaked with calibration + ChemBERTa re-weighting). CYP3A4 and CYP2C9 novel-chemistry performance is currently unvalidated (de-leaked AUROC 0.47 and 0.54 respectively — we published this finding in July 2026). The free tier is limited to 3 screenings (no batch). This is a 2026 product in Beta, not a mature platform.
Best for: Teams that need calibrated, auditable ADMET scores with documented uncertainty — especially medicinal chemists and DMPK teams who want deterministic, pipeline-integrable predictions and are willing to trade endpoint breadth for calibration quality. The abstention feature matters most for novel chemical matter (new scaffolds, natural products, beyond-rule-of-5 chemistry) where other tools may return confident-looking but unreliable scores.
There is no single best ADMET tool. The right choice depends on your use case, technical infrastructure, and tolerance for uncertainty.
| Your situation | Reach for | Also consider |
|---|---|---|
| Quick physicochemical check on one compound | SwissADME | ToxScreen (includes ADMET panel), ADMETlab |
| Regulatory-style toxicity endpoint coverage (AMES, LD50, skin sens.) | ProTox-II or pkCSM | ToxScreen (extended panel includes these as raw signals) |
| Full-library batch triage for ADMET profile | ADMETlab (free batch) or ToxScreen (calibrated batch) | ADMET-AI (self-host CLI) |
| Pipeline-integrated automated screening (API, CI/CD) | ToxScreen (REST API + SDK) | ADMET-AI (self-host), ADMETlab (API tier) |
| Novel chemical matter (natural products, beyond-RoS, macrocycles) | ToxScreen (abstention prevents false confidence) | Any tool + independent in vitro validation |
| Auditable, defensible risk scores for decision documentation | ToxScreen (calibrated + CI + AD stamp) | ADMET-AI (reproducible CLI) + custom calibration layer |
| Open-source full-control pipeline (academic lab) | ADMET-AI (you own every step) | ToxScreen API as a hosted alternative |
This page was assembled from:
/toxscreen/methods (random split, three seeds, ~1,800 compounds per endpoint) and the July 2026 ChEMBL cross-dataset de-leak benchmark at data/benchmarks/tdc_admet_cyp_deleak.json. The de-leak methodology (InChIKey connectivity-layer removal against the full TDC Veith training superset) is identical to the already-published hERG Wang-vs-Karim de-leak in scripts/tox2x_benchmark.py.If you maintain one of the tools listed and believe a factual error is present, email info@toxscreen.ai with supporting evidence. We will correct verified inaccuracies within 5 business days.
3 calibrated screenings, no credit card. Compare the output for yourself on your own chemistry.
Get Started Free →Already using another tool? Submit the same SMILES — we welcome the comparison.