Methodology

Prediction pipeline

Every screening runs the same deterministic, CPU-based pipeline. There is no docking, no 3D pose, and no GPU — predictions come from a calibrated ML model, inline, in roughly 1–7 seconds per compound.

1. SMILES validation & canonicalization. Input SMILES is parsed and canonicalized with RDKit. Invalid structures are rejected with a row-level error message before any compute is consumed.
2. Physicochemical pre-flight (sanity layer). RDKit descriptors are computed locally: MW, LogP, TPSA, HBD/HBA, QED, rotatable bonds, PAINS substructure match, Lipinski violations. These appear in the report as context and as a drug-likeness sanity check — they do not by themselves determine the tox call.
3. ADMET-AI inference. For each endpoint, ToxScreen runs ADMET-AI (Swanson et al., 2024) — an open-source chemprop message-passing graph neural network for ADMET — over the molecular graph to produce a raw per-endpoint probability. Inference is deterministic: the same SMILES always yields the same output.
4. Isotonic probability calibration. Each raw model output is mapped through a per-endpoint isotonic regression to a calibrated probability, so a reported 0.7 means roughly a 70% empirical liability rate at that score — not an uncalibrated relative ranking. Calibration is fit per endpoint (see calibration data).
5. Applicability-domain (AD) check & abstention. ToxScreen computes the Tanimoto similarity of your compound to the model's training distribution. If the compound falls outside that domain, the prediction is flagged low-confidence or abstained — ToxScreen tells you when a molecule is too novel to trust rather than guessing. This is the core differentiator (see confidence).
6. Wilson confidence intervals. Every calibrated probability carries 80% and 95% Wilson confidence intervals derived from the finite calibration set size per endpoint (n ≈ 1,800 compounds). A hERG call of 0.60 might span [0.51, 0.68] at 95% — an honest representation of what the data can support. CI bands propagate to CSV, Excel, and JSON report formats.

Endpoint panel

The current ToxScreen panel covers four critical safety endpoints:

Extended ADMET Panel (informational)

Every report also includes an Extended ADMET Panel — 15+ additional ADMET-AI predictions surfaced as supplementary signals alongside the core CTI. These are raw ADMET-AI probabilities without additional ToxScreen calibration. They complement the core panel but do not feed into the CTI score.

Planned future endpoints (no committed timeline): CYP2B6, CYP2C8 substrate, BCRP, BSEP, PXR, AhR, aqueous solubility. We add calibrated endpoints only; the Extended Panel carries raw ADMET-AI outputs for research guidance.

Calibration data

We calibrate each endpoint on a held-out random split with three seeds and roughly 1,800 labelled compounds per endpoint, then report the discrimination (AUROC) the calibrated model actually achieves. We publish the numbers as-is — including the endpoint that does not pass our gate.

Validation gate: AUROC ≥ 0.70 to enter the public panel as a fully weighted endpoint. Endpoints below the gate (currently hERG on the random split) are surfaced with an explicit low-confidence caveat and AD abstention rather than being presented as a high-confidence call.

Reproducing this: the calibration artifact is a JSON report the engine emits on each calibration run (random split, three seeds, ~1,800 compounds per endpoint). The numbers above are taken directly from that artifact, not hand-curated.

Composite Tox Index (CTI)

The CTI is a single weighted score in [0, 1] that aggregates the calibrated per-endpoint liability probabilities plus a physicochemical penalty term. It operates entirely on the calibrated ML probabilities described above — there is no binding energy, docking score, or kcal/mol anywhere in this calculation.

CTI = Σᵢ (weight_i × calibrated_prob_i)
    weight_hERG    = 0.35   (7CN1)
    weight_CYP3A4  = 0.25   (6MA7)
    weight_CYP2D6  = 0.20   (4WNT)
    weight_CYP2C9  = 0.20   (1OG5)
    calibrated_prob_i = isotonic-calibrated ADMET-AI probability for endpoint i, in [0, 1]

Risk classification bands and action thresholds:

Each endpoint's calibrated probability also appears individually in the per-target table. When the hERG endpoint is abstained for being out-of-domain, the report says so rather than forcing a number.

Confidence & applicability-domain abstention

Inference is deterministic, so there is no replicate variance to report — the same SMILES always returns the same calibrated probability. Confidence instead comes from where your compound sits relative to the model's training distribution. ToxScreen computes a Tanimoto-based applicability-domain (AD) score: the similarity of your molecule to the chemistry the model was actually trained on. This is the headline reason to trust a ToxScreen number — the model knows when it is being asked about chemistry it has never seen.

The applicability domain on real chemistry — 26 reference compounds

• IN DOMAIN (Tanimoto > 0.6). The compound is well inside the training domain. The calibrated probability is reported as-is.
• BORDERLINE (Tanimoto 0.3–0.6). The compound is near the edge of the training domain. The prediction is reported but flagged; weight it accordingly and prefer confirmatory assays.
• OUT OF DOMAIN / ABSTAIN (Tanimoto ≤ 0.3). The compound falls outside the model's domain. ToxScreen flags it as out-of-domain and may abstain from a hard call rather than extrapolate. An abstention is a feature, not a failure: it is the model declining to guess on chemistry it cannot support.

The AD status and Tanimoto score appear in every ToxScreen report (PDF, XLSX, HTML) as the Applicability Domain section. The raw Tanimoto score is also returned in the API result as ad_score. This aligns with OECD QSAR Principle 3 and EMA 2024 NAMs guidance requiring AD documentation in regulatory submissions.

Wilson confidence intervals. In addition to the applicability-domain flag, every calibrated probability carries 80% and 95% Wilson confidence intervals derived from the size of the calibration set per endpoint (n ≈ 1,800 compounds; wider for hERG n=875/925, narrower for CYP n=727/1073 to 671/1419). These appear in the XLSX Sheet 2 (per-target with CI columns), the JSON result, and the flat CSV export as ci_80_low / ci_80_high / ci_95_low / ci_95_high.

Known limitations

This list is exhaustive to our knowledge. If you find a limitation we don't list, we want to know.

• Pre-screening tool only. ToxScreen does not replace patch-clamp hERG, microsomal CYP IC50, in-vivo pharmacokinetics, or any FDA / EMA / PMDA-mandated assay. It is intended for triage before wet-lab spend.
• hERG discrimination is the weak endpoint on a random split. On a random split hERG reaches only AUROC 0.69 (FAIL). The calibrated ToxScreen pipeline reaches 0.84 AUROC on de-leaked novel chemistry (ADMET-AI baseline: 0.71), but treat hERG calls as directional and lean on the applicability-domain flag rather than the point estimate.
• CYP novel-chemistry performance is unvalidated, and where tested, no better than chance. The published CYP3A4/2D6/2C9 AUROC 0.94–0.95 is in-distribution only. An independent ChEMBL cross-dataset test (2026-07-07) found CYP3A4 AUROC 0.47 and CYP2C9 AUROC 0.54 — both statistically indistinguishable from chance; CYP2D6 was inconclusive due to a near-single-class test set. Do not treat the CYP panel as more reliable than hERG on chemistry outside its training distribution — right now, the evidence points the other way. See the calibration section above for full numbers.
• Domain shift. A model trained on one slice of chemical space may not generalize to another. Predictions on chemistry far from the training distribution are flagged low-confidence or abstained — but novel scaffolds, metals, large/atypical molecules, and unusual functional groups should be treated with extra caution even when not abstained.
• Training-set overlap. Performance on compounds that closely resemble the training data can look better than performance on genuinely new chemistry. A high-confidence, in-domain call is most trustworthy precisely because it is near known data; novelty cuts both ways.
• Frequent hitters / aggregators. Some compound classes (notably certain flavonoids and PAINS substructures) score positive on most endpoints. We flag PAINS hits explicitly; an aggregation counter-screen is recommended for any HIGH/CRITICAL scaffold.
• Stereochemistry sensitivity. Enantiomers can behave very differently. We canonicalize SMILES with RDKit but do not enforce stereochemistry. Submit explicit stereo for chiral compounds.
• Parent compound only. ToxScreen scores the molecule you submit. CYP metabolites (often more or less toxic than parent) are not predicted.
• No species-specificity. The hERG and CYP models are human; we do not currently offer dog/rat/monkey/zebrafish equivalents.
• No regulatory acceptance. No regulator (FDA, EMA, PMDA, ICH) has formally accepted ToxScreen output as a substitute for any wet-lab assay. ToxScreen output should be treated as supplementary only.

Reproducibility

Every report ships a methodology stamp footer with the exact configuration in use at scoring time:

Calibrated ADMET-AI · deterministic CPU inference
Calibration md5: (8-char hash of the active calibration config)
Frozen: (timestamp at scoring time)

Inference is deterministic. Re-running the same SMILES against the same calibration config produces an identical calibrated probability (up to floating-point ε) — there is no stochastic sampling and no random seed to vary. Reproducibility is exact, not statistical.

References

1. Swanson K, et al. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics 2024. — ToxScreen's prediction engine.
2. Pires DEV, Blundell TL, Ascher DB. pkCSM: predicting small-molecule pharmacokinetic and toxicity properties using graph-based signatures. J. Med. Chem. 2015. (Methodological background; not used in the ToxScreen pipeline.)
3. Daina A, Michielin O, Zoete V. SwissADME: a free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small molecules. Sci. Rep. 2017. (Methodological background; not used in the ToxScreen pipeline.)
4. Vargas HM, et al. Time for a fully integrated nonclinical-clinical risk assessment to streamline QT prolongation liability determinations. Clin. Pharmacol. Ther. 2018.

Privacy & data handling

• Your structures. Submitted SMILES are stored in an access-controlled database, associated only with your account, never shared with third parties, and deleted on request.
• Deletion. Email info@toxscreen.ai with your account email and we delete all associated SMILES, reports, and compute records. Data is retained until you request deletion.
• Compute provider. Predictions run on our secured servers using CPU inference. No third-party AI provider sees your structures.
• Data residency. Currently US-only. Contact us before submission if you have EU GDPR or APAC data-residency requirements.
• No tracking. We use first-party Plausible analytics for page views; no third-party trackers, no advertising cookies, no fingerprinting.
• Account deletion. Self-service in dashboard → Account → Delete account. Wipes all jobs, reports, payment records (subject to legal retention requirements for invoicing).

Legal disclaimers

Last updated: 2026-06-21. Methodology evolves as we add targets, expand calibration, and incorporate user feedback. Material changes are versioned and logged internally; a change log is available on request at info@toxscreen.ai.