MLTox Small Molecule Safety

The decision

Safety evidence loses leverage when it arrives after the chemistry is fixed

Discovery teams often have more compounds than assay capacity and several tools that disagree. The useful question is which result should change synthesis, design, testing, or advancement, and why.

Prioritize compounds with the whole profile attached

mltox keeps toxicity, potency, ADME/PK, physicochemistry, analogue evidence, and uncertainty separate so one attractive score cannot hide a weak model or poor drug profile.

Choose which compounds to synthesize or advance.
Identify structures to redesign or deprioritize.
Focus an assay, analog review, or expert escalation.
Keep unsupported predictions out of the decision.

Real MLTox portal overview for a small-molecule assessment — The current portal keeps the compound, endpoint summary, and report context in one review surface.

Endpoint coverage

Twelve endpoints, reviewed one decision at a time

Coverage and performance vary by endpoint and model version. Each result retains model evidence, applicability, uncertainty, and intended-use limits.

Selected endpoint

Hepatotoxicity

Potential to cause drug-induced liver injury (DILI).

Decision question: Could a liver-injury signal change compound priority or trigger focused review?
Evidence to inspect: Calibrated hazard state, conformal set, applicability domain, structural drivers, and measured analogues.

Boundary: this is a screening classification for scientific prioritization. It does not replace the endpoint-specific experimental study.

Hazard × dose

Ask not only whether, but at roughly what dose

Three direct structure-based dose readouts, plus a tiered systemic POD and a derived IVIVE cardiac margin, add magnitude to the hazard profile. Precision follows the evidence supporting each model.

Magnitude map

Read direct estimates and derived context differently

Acute oral LD50Direct estimate

Approximate rat oral mg/kg value, broad interval, and GHS category.
hERG IC50Direct estimate

Channel-block concentration, blocker class, and uncertainty range.
ER AC50Direct estimate

Estrogen-receptor activity concentration and potency tier.
Systemic POD + cardiac IVIVETiered context

Coarse bands remain visibly separate from direct model readouts.

Applicability gate. The platform withholds estimates outside recognized chemistry instead of extending the scale beyond its evidence.

ADME/PK and druglikeness

Ask whether the compound can behave like a drug

Seven regressors and eight classifiers cover absorption, distribution, metabolism, and excretion. Computed physicochemical and medicinal-chemistry rules sit beside the predictions.

Selected disposition domain

Absorption

Can the compound cross key barriers and support oral uptake?

Caco-2 permeability
Human intestinal absorption
PAMPA permeability
Oral bioavailability and P-gp inhibition

Each result keeps its endpoint metric, uncertainty range, and applicability decision beside the value.

Current MLTox ADME and physicochemistry module with predicted values, endpoint validation metrics, uncertainty ranges, domain gates, and computed druglikeness properties — **Every disposition value carries its own evidence.** Predicted ADME/PK outputs, validation metrics, ranges, and directly computed physicochemical rules remain visibly separate.

Move from a score to inspectable evidence

The report gives a chemist, toxicologist, or reviewer several independent ways to challenge the prediction before acting.

Nearest-analogue read-across

Inspect similar training compounds, ECFP4 similarity, domain status, and measured outcomes rather than trusting a global score alone.

Approved-drug percentiles

Place the whole profile against 3,365 approved drugs while fading weak or out-of-domain spokes.

OECD QMRF and QPRF

Export model-level and prediction-level documentation as PDF or editable DOCX; the files support review but do not guarantee regulatory acceptance.

Batch, API, and grounded questions

Screen up to 300 submitted entries per request, integrate through REST/CLI/Python, and ask questions grounded only in the report.

A comparator, not a verdict. The profile shows where the candidate sits against 3,365 approved drugs and fades weak or out-of-domain spokes.

Current MLTox Ames endpoint result showing calibrated probability, conformal set, model AUC, feature drivers, and a measured nearest analogue — **Challenge the call locally.** A measured nearest analogue supports review of the result; it does not validate the prediction by itself.

Workflow

One assessment, four review steps

Move from a controlled structure to an endpoint-level decision without losing the evidence, uncertainty, model version, or report context.

Four review gates covering intended use, controlled molecular identity, endpoint evidence, and an export or experimental handoff — Four review gates keep the intended use, controlled identity, endpoint evidence, and final action connected. Prototype artwork with illustrative data.

Define the use
Name the compound set, discovery stage, endpoints, comparator, decision owner, and action that may change.
Submit a structure
Draw it, search a public name, paste SMILES/InChI/MOL, or upload a .smi or .sdf library.
Inspect the whole report
Review hazards, potency, ADME/PK, physicochemistry, analogue evidence, uncertainty, and domain status.
Compare, export, or feed back
Rank the library, export PDF/JSON/QMRF/QPRF, or submit a confirmed small-molecule outcome for retraining.

Decision confidence

Separate a confident-looking score from a defensible decision

mltox keeps four different questions visible so reviewers can tell whether a prediction should inform the decision, trigger more evidence, or be withheld.

Prototype visual: Three matched endpoint records representing supported, ambiguous, and unsupported predictions, each with local evidence, uncertainty, applicability, and a distinct next step. — Three matched endpoint records representing supported, ambiguous, and unsupported predictions, each with local evidence, uncertainty, applicability, and a distinct next step. Prototype artwork with illustrative data.

01
Does the endpoint model earn trust?
Start with endpoint-specific evaluation, calibration, dataset design, and the exact model version rather than the displayed probability alone.
02
Is this prediction decisive?
Inspect the 90% split-conformal set. Coverage is marginal across the data distribution, not a guarantee for this molecule.
03
Does the model recognize this chemistry?
Check applicability, nearby evidence, scaffold coverage, and conflicts before carrying the result into a series decision.
04
What action does the evidence support?
Record whether to advance, redesign, run a targeted assay, seek expert review, or abstain, together with the reason.

Team workflow

Built for a cross-functional discovery decision

The record is shared. Each discipline keeps its own judgment and responsibility.

Medicinal chemistry

Compare and redesign

Bring endpoint findings into series and design reviews.

DMPK and discovery safety

Focus evidence

Prioritize assays and challenge uncertain results.

Computational toxicology

Review methods

Inspect versions, applicability, calibration, and disagreement.

Project toxicology and informatics

Carry the rationale

Connect candidate reviews, reports, and scientific systems.

Focused evaluation

Prove one discovery decision before expanding the platform

Start with one representative compound series and a separate prospective set. Compare mltox with current tools, experts, and assay plans.

Discuss a small-molecule pilot

Measures that matter

Useful disagreement with the current process.
Review time and evidence inspection burden.
Appropriate escalation of unsupported cases.
Whether the result improves the next action.
A required exit: adopt, revise and rerun, or stop.