World's first foundational model for toxicology
Review endpoint-specific hazards, exposure, uncertainty, and analogue evidence before deciding what to make, test, advance, or stop.
Review individual compounds and related chemical series while keeping the endpoint evidence behind each pattern visible.
MLTox Small Molecule Safety
Review 15 visible hazard outputs, direct and tiered potency estimates, 15 ADME/PK models, physicochemistry, read-across, approved-drug context, and OECD-format model and prediction records in one report. Exploratory nephrotoxicity remains clearly labeled.
Explore Small Molecule Safety
SeriesLens workflow
SeriesLens is part of MLTox Small Molecule Safety. Compare predicted liabilities across user-defined series, then inspect the compounds and local changes that drive the pattern.
Explore the SeriesLens workflow
A score is not enough. Read calibrated probability, conformal ambiguity, endpoint performance, and chemical applicability as separate evidence.
In domain, backed by a reliable endpoint model, with a decisive prediction set.
The 90% conformal set includes both labels, so the model cannot distinguish clean from hazard.
Outside the applicability domain, below the reliability gate, or withheld because the evidence is too weak.
Fifteen visible hazard outputs do not come from one universal model. Each endpoint is independently trained, selected, and evaluated against its own data, performance metrics, reliability evidence, and applicability evidence. Exploratory nephrotoxicity and the retained immunotoxicity champion keep their lower-confidence caveats visible.
Classification, calibration, and model selection are performed for the biological question and data available at each endpoint.
Measured endpoint performance stays beside the result. Lower-confidence outputs retain their metrics, caveats, and maturity labels for scientific interpretation.
A structure outside the modeled chemical neighborhood is flagged as low-confidence instead of receiving a polished extrapolation.
A 90% prediction set can contain both labels. That state is presented as unresolved evidence, not compressed into a false binary answer.
High-value predictions can be routed into appropriate assays. Returned outcomes remain connected to their source; eligible small-molecule endpoint models may use them in a controlled learning loop.
A separate language-model assistant is grounded in a report-derived fact sheet, instructed to label general background, and to say when candidate-specific evidence is absent. It does not compute the underlying predictions.
Held-out benchmark evidence
Each task was rebuilt from scratch on its official Therapeutics Data Commons train fold and evaluated against public leaderboard methods.
Explore methods and full comparisonsFirst on the board. Built on MapLight + GNN, not an MLTox architecture — see Evidence for the caveats.
Highest mean on the board — a statistical tie, not a win.
Second on the board — the leader’s edge is inside the noise.
Second on the board; MiniMol keeps a separable lead.
Mean ± standard deviation across five seeds. Each task uses its own scale; do not compare interval lengths across cards.
A bounded first step
Start with a fixed compound set and a decision your team already understands. Agree on success before MLTox reviews the blinded outcomes.
A hazard flag does not say at what dose an effect may occur or whether a compound has a workable ADME profile. The report keeps those questions separate.
Could the structure carry an intrinsic toxicity signal?
Calibrated endpoint probabilities, conformal sets, measured AUC, structural alerts, and interpretable drivers.
At roughly what concentration or dose might the signal matter?
LD50, hERG IC50, ER AC50 as direct estimates, plus a tiered systemic POD and a derived cardiac IVIVE margin, with broad ranges where precision is limited.
Can the compound reach, persist, and clear in a workable way?
Absorption, permeability, distribution, binding, clearance, half-life, and CYP substrate predictions, each with its own metric and domain gate.
Nearest analogues, approved-drug percentiles, physicochemical rules, model and domain evidence, PDF/JSON, and OECD-format QMRF/QPRF records stay attached to the result.
Interpretation boundary: potency and ADME outputs are structure-based estimates for prioritization. They do not establish a safe human dose, exposure profile, or regulatory conclusion.
MLTox Proteins & Biologics
Protein and biologics work uses different evidence from small-molecule safety. Review selectable HLA panels, humanness, junctional neoepitopes, sequence liabilities, intrinsic solubility, and conservative de-immunization candidates.
Explore Proteins & Biologics
A decision, not a demo
Define one blinded comparison, make the success criteria explicit, and decide whether the evidence improves your next compound review.
Plan a pilot