In Silico Toxicity Prediction
Catch liabilities early.

Stereochemistry kept as signal, and every model checked against real chemistry — not dataset artefacts.

Book a Demo All Modules

L Lead AI Toxicity workbenchIllustrative profile

05 / SCREENING EVIDENCEReview required

Toxicity endpoints

Human-health and environmental signals.

EndpointResultSource

Acute oral toxicity135 mg/kgWebTEST · OPERA · VEGA

Mutagenicity (Ames)NegativeVEGA · WebTEST · CTX

CarcinogenicityReviewVEGA · CTX

READ THE EVIDENCE

A result with its limits

The demo carcinogenicity models disagree near their domain boundary. That uncertainty stays in the profile.

Source tools retainedDomain-aware reviewScientist-led decision

Explore each stage of the screening profile.Workbench demo values · no live analysis

Supported workflowToxicity & ADMET Prediction

Why teams come
to us for this.

Late toxicity burns time and budget. Screening earlier only helps if the data and the model are scientifically defensible.

Respiratory toxicity: ROC-AUC 0.900 and recall 0.894
External validation on 193 compounds: 85.5% accuracy, 90.4% recall
Stereochemistry improved every model and dataset tested

FROM CHEMISTRY TO A DEFENSIBLE DECISION

A prediction is the beginning.
The evidence takes it further.

Molecular detail, prepared data and visible uncertainty.
Connected all the way to scientific review.

Stereochemistry retained

3D structure kept as signal — different forms of a molecule behave differently.

Switch the configuration to see what is preserved

MOLECULAR IDENTITYChiral detail retained

ConfigurationAConnectivityUnchanged

✓

Same connections. Different orientation.

Keep the spatial distinction in the molecular representation.

Schematic illustration of the research capability; no toxicity comparison implied.

Prepared toxicity datasets

Standardised molecular data with a defensible benchmark split.

Inspect the records that need attention

DATA PREPARATION

A record you can trace.

Example records

⌬MOL-001Structure standardisedReady

⌬MOL-002Stereo detail retainedReady

⌬MOL-003Duplicate identityReview

⌬MOL-004Endpoint label missingReview

⌬MOL-005Provenance attachedReady

IdentityQuality checksBenchmark split

REPORTED EXTERNAL VALIDATION

193compounds

Keep the benchmark
outside the training set.

85.5%Accuracy

90.4%Recall

Respiratory toxicity research reported on this site. Separate from the example records shown here.

Interpreted model behaviour

Checks the model responds to real chemistry, not dataset artefacts.

Explore the evidence behind each endpoint

ENDPOINT EVIDENCEWorkbench demo

Classification

Review required

VEGACTXEvidence review

The demo tools disagree near the domain boundary. A single summary label must not hide that disagreement.

Confidence: Low
Applicability: Borderline

01 / Source02 / Evidence type03 / Interpretation

Honest prediction limits

Computational evidence for candidate selection — never a clinical conclusion.

See how a data gap changes the next step

APPLICABILITY CHANGES THE DECISION

Make uncertainty part of the result.

Near the domain boundary, confidence and model disagreement become essential context for the reviewer.

NEXT STEPInvestigate before advancing

Built for

Head of R&DResearch leadershipData science and AI leadership

The rest of the Research & Development

Supported workflow

Therapeutic

A shortlist you can defend.

Supported workflow

Diagnostics

Simulate 1%. Predict the rest.

Want ADMET running on your data?

Tell us the process or prediction you need. We'll map it to a module, a deployment and an honest timeline.

Bring your challenge.
Let’s find the right next step.

Request a Free Data Assessment