SSL-GCN

SSL-GCN predicts chemical toxicity using semi-supervised learning (SSL) and graph convolutional neural networks (GCN) to improve toxicity assessment of compounds for drug discovery.

Key Features:

• Semi-Supervised Learning (SSL): Leverages both labeled and unlabeled data to improve model performance beyond supervised learning alone.
• Graph Convolutional Neural Network (GCN): Captures molecular topology and connectivity by operating on graph representations of chemical compounds.
• Mean Teacher Algorithm: Trains a student model with consistency regularization against a teacher model to stabilize learning from unlabeled data.
• Molecular Graph Representation: Processes molecular graphs where nodes represent atoms and edges denote bonds.
• Training Data Ratio: Demonstrates optimal unannotated-to-annotated training data ratios ranging from 1:1 to 4:1.
• Performance Metrics: Evaluated on the Tox21 dataset achieving an average ROC-AUC score of 0.757 across twelve toxicological endpoints.
• Comparative Performance: Shows a reported 6% improvement over traditional supervised GCN models and conventional machine learning methods and surpasses built-in DeepChem ML methods.

Scientific Applications:

• Chemical Toxicity Prediction: Predicts toxicological properties of chemical compounds to support safety assessment in drug discovery.
• High-Throughput Screening Support: Reduces reliance on resource-intensive in vitro and in vivo experiments by enabling computational screening.
• Other Chemical Property Prediction: Can be extended to additional chemical property prediction tasks using large existing chemical databases.

Methodology:

Uses a GCN architecture that processes molecular graphs (nodes = atoms, edges = bonds) and is trained with the Mean Teacher semi-supervised algorithm to balance labeled and unlabeled data (optimal unannotated:annotated ratio 1:1–4:1); evaluated on the Tox21 dataset with reported ROC-AUC 0.757 across twelve endpoints and a 6% improvement versus supervised GCN and conventional ML, also surpassing DeepChem ML methods.