Scientific decision workspace
Run small-molecule and biologics analyses through the web portal, REST API, command line, or Python. Retain versioned analysis reports and deploy through the hosted portal or in infrastructure you control.
Platform role
mltox is a focused safety and engineering workspace for Small Molecule Safety and Biologics Engineering. It should connect with registries, ELNs, LIMS, scientific data platforms, modeling environments, and reporting workflows rather than become a generic system of record.
MLTox is not one general-purpose AI model. It combines modality-specific trained models with established computational methods and explicit scientific boundaries.
Small molecules
Production models use gradient-boosted classifiers and regressors to estimate endpoint-specific hazards, potency, and ADME/PK. Classifiers use isotonic calibration and split-conformal sets; regressors use cross-validated intervals; a chemical applicability gate scopes both.
Biologics
Per-allele HLA-binding models sit alongside sequence-liability methods and optional ESMFold structure prediction. Outputs are design hypotheses, not clinical ADA predictions.
Computed evidence
Physicochemical calculations, structural alerts, labeled analogue outcomes, and sequence rules remain distinct evidence types. They should not all be called AI.
Report assistance
The assistant receives a report-derived fact sheet. It is instructed to explain completed results, label general background, and avoid computing the scientific predictions.
Implementation truth
Website language follows the observed product state and the evidence available for each capability. This keeps current product behavior distinct from scoped enterprise work and claims that still need approval.
| Capability | Current state | Website treatment |
|---|---|---|
| Web portal, REST API, CLI, Python library, and batch scoring | Shipped | Describe as current product interfaces. |
| Analysis history, versioned model bundles, and structured reports | Shipped | Describe as analysis traceability, not an approval workflow. |
| PDF, JSON, QPRF, and endpoint-level QMRF outputs | Shipped | Show as current evidence and reporting outputs. |
| Weighted small-molecule retraining and model registry | Shipped with limits | State that biologics models are excluded and governance is still limited. |
| Docker or Windows-service self-hosting and hosted portal | Shipped | Describe the observable deployment modes without adding compliance claims. |
| Per-user identity, roles, approvals, audit trails, and ELN/LIMS connectors | Roadmap | Position as scoped enterprise work until implemented and validated. |
PLT-02 | Scientific workspace
The current platform preserves analysis history and report context. Reviewer assignment, approval state, rationale capture, and controlled return to an ELN or LIMS are sensible enterprise additions, but they are not presented as shipped features.
AI/ML feedback and model versions
For small molecules, confirmed outcomes can be stored with source-based weights and used to retrain versioned endpoint models. Biologics models are trained separately and are not updated from submitted outcomes. The repository does not yet implement a formal challenger review, approval, or promotion workflow.
Store an eligible small-molecule hazard outcome with its endpoint, observed result, source, and source-derived weight.
Combine seed data with weighted feedback, rebuild endpoint models, and register a versioned model bundle.
Evaluate a candidate against protected validation evidence, calibration, subgroup behavior, and decision impact.
Add named review, approval, monitored release, and rollback controls before treating retraining as governed production behavior.
Clear boundaries matter more than broad claims. These are the distinctions buyers and reviewers should be able to verify.
Yes. MLTox uses endpoint- and modality-specific machine-learning models alongside QSAR descriptors, chemical rules, analogue evidence, sequence methods, and optional structure prediction. It is not one universal AI model.
No. Scientific outputs are computed before the assistant is invoked. The separate assistant receives a report-derived fact sheet, is instructed to label general background, and is instructed to say when candidate-specific evidence is absent.
No. Eligible small-molecule classifiers use isotonic calibration and split-conformal sets; potency and ADME regressors use cross-validated intervals. A chemical applicability gate scopes small-molecule estimates. Biologics outputs use separate methods and caveats.
MLTox does not learn continuously from inference. Small-molecule feedback retraining is manual by default and can be explicitly configured after a feedback threshold; each retrain creates a versioned bundle. Biologics models are separate and are not updated from submitted outcomes.
No. MLTox provides research and preclinical decision support. Outputs remain hypotheses and evidence aids for qualified review, experimental follow-up, and applicable regulatory obligations.
Integration and evaluation
The REST API, CLI, and Python library support programmatic scoring and report retrieval today. A direct ELN, LIMS, registry, or scientific data-platform connector should begin as one named handoff with explicit identity, metadata, version, ownership, and validation requirements.
Identify the source system, molecular identity, required metadata, assessment version, review owner, and destination. Treat connector development and write-back as scoped work until demonstrated.
A functional champion tests the decision while an independent validator, Informatics, IT, Security, Quality, Legal, and Procurement review operation.