01
Graph representation
Molecular structure encoded through SMILES must be transformed into graph data that a GCN or GAT model can process.
PRJ—04Predictive toxicology
Tox21
A molecular-toxicity exploration tool built around graph neural networks, SMILES input, and twelve prediction endpoints.
(Overview)
Molecular toxicity is complex. Exploring a prediction should not be.
Tox21 provides an approachable interface for investigating in-silico toxicity predictions. Users can enter SMILES notation, work with a molecule library, or submit a batch for analysis across twelve endpoints.
Type
Machine-learning study
Role
ML & product engineering
Focus
Molecules · GNN · visualization
Status
Experimental product
(Introduction)
Tox21 turns graph-neural-network output into an exploratory interface for comparing molecules, endpoints, and toxicity signals.
(Problem)
The model speaks in graphs and endpoints. The user needs an understandable molecular result.
Toxicity prediction involves specialized representations, imbalanced labels, multiple endpoints, and uncertainty. A useful interface has to preserve those constraints without becoming inaccessible.
02
Multiple endpoints
One molecule may produce different signals across twelve toxicity targets, making a single binary label insufficient.
03
Scientific context
Predictions need clear limits: the result is an in-silico aid, not a substitute for laboratory validation.
(Approach)
Build an exploration layer around the model instead of exposing a bare prediction endpoint.
The workflow was shaped around how someone investigates a molecule: provide or select a structure, run prediction, compare endpoint signals, and retain enough context to interpret the result responsibly.
Parse SMILES structures01
Build molecular graph features02
Predict twelve endpoints03
Present uncertainty and context04
(Solution)
One interface for single molecules, batch analysis, and endpoint comparison.
The product joins a DeepChem prediction pipeline with a restrained Next.js interface. It supports direct SMILES input, categorized molecule discovery, comparison, and batch CSV workflows.
01
SMILES input
Users can submit a molecular representation directly for toxicity analysis.
02
Twelve endpoints
The result preserves the multi-target nature of the Tox21 dataset.
03
Molecule library
Example compounds can be filtered by categories such as drugs, hormones, solvents, and pesticides.
04
Batch CSV
Multiple compounds can be prepared for a broader analysis workflow.
05
Molecule comparison
Results can be evaluated side by side instead of in isolated screens.
06
Responsible framing
The interface clearly positions output as an in-silico prediction requiring further validation.
(Outcome)
A specialized ML pipeline turned into a product people can explore without hiding its limits.
Tox21 brought graph-based modeling, chemical representations, API behavior, and product interface decisions into one focused experiment.
The outcome is not a laboratory verdict. It is a clearer way to inspect predicted toxicity signals and understand where deeper investigation should begin.