Chemical toxicity prediction based on semi-supervised learning and graph convolutional neural network.

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Chemical safety is crucial in drug discovery.
• Traditional toxicity testing is time-consuming, costly, and raises ethical concerns.
• Limited annotated data hinders machine learning model performance in chemical toxicology.

Purpose of the Study:

• To develop a Graph Convolutional Neural Network (GCN) model for chemical toxicity prediction.
• To leverage semi-supervised learning (SSL) with the Mean Teacher (MT) algorithm to enhance model performance using unannotated data.
• To improve upon existing machine learning methods in predicting toxicological endpoints.

Main Methods:

• Proposed a Graph Convolutional Neural Network (GCN) architecture.
• Employed the Mean Teacher (MT) algorithm for semi-supervised learning (SSL).
• Trained and evaluated models on the Tox21 dataset for twelve toxicological endpoints.

Main Results:

• The SSL-GCN model achieved an average ROC-AUC score of 0.757 on the test set.
• Demonstrated a 6% improvement over GCN models trained with supervised learning and conventional ML methods.
• Showcased superior performance compared to built-in DeepChem ML methods.

Conclusions:

• Semi-supervised learning significantly enhances chemical toxicity prediction by effectively utilizing unannotated data.
• The optimal ratio of unannotated to annotated data ranges from 1:1 to 4:1.
• SSL-GCN offers a promising approach for chemical property prediction, applicable to large chemical databases.