Submit a SMILES, interpret scores, understand the applicability domain
ToxTwin is a toxicological prediction model that analyses the chemical structure of a molecule to estimate its probability of activity across 14 regulatory biological endpoints. It uses a graph neural network (GNN) trained on high-throughput screening data from the ChEMBL and Tox21 databases. ToxTwin estimates preliminary toxicological risk from structure alone, signals whether the molecule falls within the model’s applicability domain, and generates a synthetic pharmacological interpretation. It does not replace mandatory biological tests (ICH S2, S7B), does not predict in vivo toxicity, and does not provide a legally binding regulatory opinion.
ToxTwin accepts molecules in SMILES notation — the standard textual representation of chemical structure. Canonical SMILES is recommended (ToxTwin applies RDKit canonicalisation upstream). SMILES sources include PubChem, ChEMBL, ChemDraw/MarvinSketch and RDKit. The demo interface is accessible on twingital-ventures.com with a freemium quota of 3 free analyses. Invalid SMILES, molecules exceeding 500 heavy atoms and mixtures (. character) are subject to specific handling documented in the guide.
The 7 nuclear receptor endpoints (NR-AR, NR-AR-LBD, NR-AhR, NR-Aromatase, NR-ER, NR-ER-LBD, NR-PPAR-γ) assess endocrine disruption. The 5 cellular stress response endpoints (SR-ARE, SR-ATAD5, SR-HSE, SR-MMP, SR-p53) signal genomic, mitochondrial or protein toxicity. The 2 pharmacotoxicology endpoints — hERG (potassium channel, ICH S7B) and Ames (Salmonella mutagenicity, ICH S2R1) — have V2.3 performance below the regulatory target and are indicative only.
Each endpoint returns a score between 0 and 1 representing the calibrated probability of activity. Interpretation thresholds are: low (<0.25), moderate (0.25–0.50), high (0.50–0.75), critical (>0.75). A complete profile is read at three levels: dominant signals (>0.25), structural coherence with functional groups, and applicability domain context. The Aspirin example (low profile across 14/14 endpoints, Tanimoto 0.857, in domain) illustrates the standard reading.
The applicability domain indicates whether the molecule is structurally similar to the training corpus. ToxTwin uses a composite score based on three complementary signals. Common causes of out-of-domain results include novel scaffolds, unusual functional groups, peptides, polymers and organometallic compounds. An out-of-domain score does not mean toxicity — it means insufficient data for a reliable prediction.
The textual interpretation generated by a local LLM provides a narrative synthesis of signals, correlation with functional groups, risk level assessment and regulatory context. It does not constitute a regulatory opinion, an in vivo prediction, or a substitute for the judgement of a qualified toxicologist. It is a starting point for reflection.
ToxTwin is appropriate for early triage, mechanistic hypotheses, experimental prioritisation and comparative monitoring. It is not appropriate as a substitute for regulatory tests, as a component of a marketing authorisation dossier without experimental confirmation, or for out-of-scope classes. Endpoints not covered in V2.3 (DILI, ClinTox, extended cardiotoxicity, phototoxicity, DART, metal complexes) are planned for V3.0.
The guide addresses the most common questions: diagnosing invalid SMILES, interpreting a high hERG score (warning signal justifying a patch-clamp test, not a stop decision), using out-of-domain scores (indicative, a strong signal warrants investigation), deterministic prediction reproducibility, stereochemistry handling, professional API access, and the policy of not storing submitted SMILES.
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.