Prerequisite: Load The Molecule From ChemrytIQ

Before opening ChemrytIQ-ToxPred, search the molecule in ChemrytIQ by SMILES, InChI, molecule name, or CAS number. Confirm the correct molecule on the ChemrytIQ page, then open the required Chemryt app from that same molecule context so the selected structure is loaded into the app automatically.

What It Does

ChemrytIQ-ToxPred combines browser heuristic and XAI runtime with server-side GNN inference through toxpred APIs. It supports toxicity and ADMET triage for molecules drawn or searched in ChemrytIQ.

Toxicity prediction

Scores alerts and endpoint risks such as Ames, hERG, hepatotoxicity, CYP, BBB, P-gp, and absorption-related signals.

Server GNN groups

Uses endpoint groups such as Tox21, ClinTox, ToxRefDB POD, BBBP, HIV, and PubChem qHTS where available.

SELFIES and XAI

Shows token hotspots, candidate perturbations, safe-space suggestions, and endpoint movement explanations.

Quick Tutorial

Load a molecule in ChemrytIQ and open ToxPred from the child module panel.
Confirm the input source, molecular string representations, and current structure preview.
Run toxicity prediction and review endpoint matrix, ADMET summary, toxicophore alerts, and heuristic explanations.
Review server-side GNN panels where available, including endpoint group tabs, confidence, applicability-domain notes, and ontology caveats.
Use SELFIES token hotspots, safe-space candidates, morph points, or bioisosteric suggestions to explore risk-reducing edits.
Feed toxicity signals into MCCP or use them to prioritize experimental toxicology assays.

Main Areas

Area	What to review	When to use it
Input and capabilities	ChemrytIQ molecule, SMILES, SELFIES, supported endpoint groups, and ML capability status.	Use before running prediction.
Endpoint results	Toxicity matrix, ADMET categories, GNN groups, confidence, and applicability-domain notes.	Use to triage safety risk.
Explainability	Toxicophores, token pressure, endpoint deltas, morph paths, and safe-space analogs.	Use to understand and reduce risk.

ML Model / Computation Used

Model or method	What it predicts	Implementation details
ToxPred GNN registry	Toxicity and ADMET endpoint groups including Tox21, BBBP, ClinTox, HIV, PubChem Tox qHTS, and ToxRefDB POD families.	Registry uses PyTorch checkpoints named toxpred_gnn_best.pt with task, metric, threshold, applicability-domain, and calibration JSON files. Standardization uses RDKit cleanup/largest organic fragment; atom feature count is 37 and bond feature count is 12.
Tox21 GINE v2 example model	Tox21 nuclear-receptor and stress-response endpoints.	GINE architecture with hidden dim 128, 4 layers, dropout 0.15, scaffold split, best epoch 36; reported test mean ROC-AUC about 0.84 and mean average precision about 0.48.
Heuristic/XAI runtime	Toxicophore alerts, token hotspots, endpoint deltas, and safe-space analog suggestions.	Browser heuristics and explainability views complement the server-side GNN outputs; SIDER is marked withheld because of indication/reporting-bias caveats.

Good Practice

ToxPred is early triage with guarded trained-model support. It is not regulatory toxicology evidence; confirm safety decisions with validated assays and qualified toxicology review.

Reference Used

This Tutorial page mirrors the ChemrytIQ reference module: ChemrytIQ-ToxPred.