Equivariant Graph Neural Networks for Toxicity Prediction.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Predictive toxicity modeling is vital for efficient drug discovery, reducing late-stage failures.
• Current machine learning models often rely on 2D molecular representations, limiting accuracy.
• 3D molecular structures offer a more natural and potentially accurate representation for predictive tasks.

Purpose of the Study:

• To investigate the efficacy of equivariant graph neural networks (EGNNs) for predicting molecular toxicity.
• To evaluate the performance of the equivariant transformer (ET) model using 3D molecular structures.
• To explore the relationship between molecular total energy and toxicity prediction.

Main Methods:

• Utilized the equivariant transformer (ET) model within the TorchMD-NET framework.
• Trained and evaluated the model on eleven diverse toxicity datasets from MoleculeNet, TDCommons, and ToxBenchmark.
• Incorporated attention weight analysis to enhance model interpretability.

Main Results:

• The ET model demonstrated strong performance in toxicity prediction, achieving accuracies comparable to state-of-the-art methods.
• EGNNs successfully learned relevant 3D molecular representations correlating with toxicity.
• No significant relationship was found between a molecule's total energy and its toxicity prediction.

Conclusions:

• EGNNs provide a powerful approach for toxicity prediction by leveraging 3D molecular geometry.
• The ET model offers a reliable and explainable method for integrating molecular conformers into drug discovery pipelines.
• Future applications of EGNNs are promising for larger and more diverse datasets in toxicity assessment.