ToxMPNN: A deep learning model for small molecule toxicity prediction.

Area of Science:

• Computational chemistry
• Toxicology
• Drug discovery

Background:

• Machine learning (ML) shows promise for predicting small molecule toxicity, but data limitations hinder performance.
• Models trained on single toxicity endpoints often yield unsatisfactory results.
• A comprehensive dataset with 27 toxic endpoints across seven classes was curated.

Purpose of the Study:

• To develop an accurate toxicity predictive model for small molecules.
• To improve the performance and stability of toxicity prediction models.
• To assess the utility of graph-based deep learning for toxicity assessment.

Main Methods:

• Developed ToxMPNN, a toxicity predictive model utilizing the message passing neural network (MPNN) architecture.
• Integrated a dataset of toxic small molecules with multiple endpoints and added marketed drugs as negative samples.
• Employed graph-based deep learning (DL) algorithms for toxicity prediction.

Main Results:

• ToxMPNN demonstrated superior performance in capturing toxic molecular features, achieving an ROC_AUC score of 0.886 on the Toxicity_drug dataset.
• The inclusion of marketed drugs as negative samples enhanced predictive performance and model stability for the Common-Toxicity task.
• Graph-based DL algorithms proved effective in toxicity prediction.

Conclusions:

• ToxMPNN is a trustworthy and effective tool for assessing small molecule toxicity.
• The developed model can aid in the efficient development of new drugs.
• Comprehensive datasets and advanced DL architectures improve toxicity prediction accuracy.