A Benchmark Study of Graph Models for Molecular Acute Toxicity Prediction

Area of Science:

• Computational toxicology
• Cheminformatics
• Machine learning

Background:

• Organic compounds are widely used, necessitating efficient acute toxicity assessment.
• Reducing animal testing and human labor in toxicity evaluation is a significant goal.
• Graph models offer a promising avenue for predicting chemical toxicity.

Purpose of the Study:

• To evaluate the performance of five distinct graph models for acute toxicity prediction.
• To identify the most effective graph model for diverse aquatic toxicity datasets.
• To explore the explainability of the best-performing model.

Main Methods:

• Application of five graph models: message-passing neural network, graph convolution network, graph attention network, path-augmented graph transformer network, and Attentive FP.
• Testing on four acute toxicity datasets: fish, *Daphnia magna*, *Tetrahymena pyriformis*, and *Vibrio fischeri*.
• Performance evaluation based on prediction error and model explainability.

Main Results:

• Attentive FP achieved the lowest prediction error across all four toxicity tasks.
• The model demonstrated superior predictive performance compared to the other four graph models.
• Attention weights from Attentive FP enabled the creation of atomic heatmaps for enhanced interpretability.

Conclusions:

• Attentive FP is a highly effective graph model for predicting the acute toxicity of organic compounds.
• The model's explainability features, derived from attention weights, are valuable for toxicological insights.
• Graph-based approaches, particularly Attentive FP, show significant potential in reducing traditional toxicity testing methods.