Submit a small molecule or a protein and get a calibrated safety report in seconds — QSAR toxicology across twelve endpoints, or learned immunogenicity and developability screening for biologics. The models fine-tune themselves as your clinical and lab results come back.
Research tool. Every output is an in-silico hypothesis on an unvalidated model — a prioritization aid, not a regulatory or clinical determination.
mltox auto-detects what you submit and routes it to the appropriate model — a protein is never scored by a small-molecule QSAR. Real-world outcomes feed back in.
Paste a SMILES, InChI, MOL/SDF, or a
FASTA/amino-acid sequence. No account, no setup.
Format is auto-detected. Small molecules are featurized (RDKit or a pure-Python fallback); biologics get a sequence-liability pipeline.
Per-endpoint gradient-boosted classifiers, or learned per-allele epitope models — each with calibrated confidence and an applicability-domain gate.
A structured report with rationale. Submit clinical/lab outcomes; weighted retraining folds them back into the models.
How the loop closes. Every report is a prediction. When a real result comes back — a clinical readout, an in-vivo study, a bench assay — you submit it as feedback. Each outcome is stored with a trust weight based on its source (a clinical trial counts for more than a literature note), then folded into a retrain over the seed data plus all accumulated feedback. That updated model scores your next molecule. The more real-world data you feed it, the more it is tuned to your chemistry.
Everything happens inside your boundary. mltox is self-hosted: the models, the feedback store, and retraining all run on infrastructure you control (a container or host you own). Scoring is fully local — RDKit + scikit-learn, and ESMFold on your own GPU — so a submitted molecule, sequence, or clinical outcome is never sent to a third-party service to be scored. Feedback lives in your own local database and is used only to retrain your own models; nothing is pooled across tenants.
Scope note: the fine-tune loop updates the small-molecule endpoint models. The biologics epitope models are trained separately on curated IEDB data and are not retrained on submitted outcomes. The only outbound fetch anywhere is the one-time download of the public ESMFold model weights — model data in, never your data out.
The small-molecule models don't apply to a 300-residue protein — so proteins get a real sequence-based analysis instead of a fake QSAR score.
SMILES · InChI · MOL/SDF → QSAR toxicology report
FASTA · sequence → immunogenicity + developability
Each is an independent binary hazard model, trained on real assays and evaluated by scaffold-split CV (0.58–0.95 AUC by endpoint). Severity weights how much each drives the overall risk band.
A suite of sequence-aware capabilities for antibody and protein engineers. Each is an in-silico hypothesis to prioritize wet-lab work, never a verified fix.
In fusions and multispecifics, flags T-cell epitopes created at the seam — present in neither parent domain — across auto-detected linkers.
Proposes conservative BLOSUM62 substitutions the model predicts remove an epitope — anchors first, Cys/Pro/glyco fixed, with a greedy best-pair fallback.
Detects VH/VL, annotates CDR1/2/3 vs framework, and tags a risk as a redesign target (CDR) or a humanization gap (framework).
Selectable epitope panels — HLA-DR, DR/DQ/DP, and a supplementary class-I CD8 panel — each per-allele CV-AUC gated (0.80–0.93).
A continuous per-residue intrinsic-solubility profile from hydrophobicity, charge and β-propensity — flags aggregation-prone regions.
Every endpoint is isotonic-calibrated and carries a 90%-coverage conformal set — a two-way {clean, hazard} means “can’t distinguish.”
Folds the sequence with ESMFold on GPU, then uses per-residue solvent exposure to down-weight buried liabilities and up-weight exposed ones. Runs as a queued job (queued → running → done, with your place in line) so long folds never time out — results are cached and viewable in 3D.
ESMFold’s memory grows with the square of length, so a full-size protein won’t fit on one pass. Long chains are cut at flexible linker-like regions — which tend to fall between domains — and each segment is folded on its own, then recombined into one full-length exposure profile. The split is a heuristic, and how the segments sit relative to each other is not predicted.
Every report explains why — the drivers behind the overall band and the developability call, in language a reviewer can act on.
A safety tool that overstates certainty is worse than none. mltox is built to tell you when it doesn’t know — and to keep its numbers honest about novel chemistry.
Try it on your molecule →AUCs are measured across Bemis–Murcko scaffold groups on real assays, so the numbers reflect novel chemistry — not analog leakage from the training set.
Endpoints below 0.70 CV-AUC (the hard, data-scarce in-vivo ones) are shown with their score but held out of the overall-risk escalation, so a weak model can’t inflate a verdict.
Structures unlike anything seen in training are flagged low-confidence instead of extrapolating a confident-looking number.
90%-coverage sets quantify ambiguity: a set containing both labels is an explicit “can’t distinguish,” not a silent coin flip.
Every biologics output is labeled an in-silico hypothesis on an unvalidated model — a capability to prioritize experiments, not a verified answer.
No signup. Paste a structure or a sequence and read the report.