Toxicity Predictions for Androgen and Estrogen Receptor Chemicals via SMILES

A descriptor-free deep learning model predicts chemical toxicity for androgen and estrogen receptors directly from SMILES strings. The models were trained on 1,664 and 1,529 chemicals with experimental binary labels for AR and ER, respectively, achieving accuracy scores of 0.75 and 0.81. Francesca Cutropia published this work on figshare in 2026, introducing an explainable AI approach for substructure-level interpretability.

Use Cases

• Predicting chemical toxicity for androgen and estrogen receptors based on SMILES string inputs.
• Training explainable AI models for mechanistic insight into receptor-mediated effects.
• Benchmarking descriptor-free deep learning approaches against traditional molecular descriptor methods in toxicology.

Strengths

• Model performance metrics are provided, including accuracy of 0.75 ± 0.08 for AR and 0.81 ± 0.05 for ER.
• Training data sizes are explicitly stated: 1,664 chemicals for AR and 1,529 for ER.
• The model framework is descriptor-free, operating directly on SMILES strings for automatic feature extraction.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the underlying dataset is unknown, which may limit suitability assessment.
• The dataset is small in scale at 186.8 KB, indicating limited scope.

Provenance

Source

figshare

Collection Method

Models trained on publicly available, high-quality datasets with experimental binary labels.

Freshness

Last updated 2026-06-03 23:43:27; freshness should be verified.