Drug toxicity prediction model based on enhanced graph neural network.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Drug toxicity prediction is crucial but challenging for traditional machine learning.
• Existing Graph Neural Networks (GNNs) do not fully capture complex molecular interactions.

Purpose of the Study:

• To develop an enhanced GNN algorithm for improved drug toxicity prediction.
• To incorporate multi-view node features and multi-scale attention for better molecular representation.

Main Methods:

• Proposed an enhanced GNN with multi-view node features and preprocessed adjacency matrix.
• Implemented a pooling technique, normalization, and activation layer.
• Applied multi-scale attention to learn intricate relationships within molecular graphs.

Main Results:

• Achieved high ROC-AUC scores on binary classification tasks (e.g., 0.92 for DILI, 0.845 for Carcinogens).
• Demonstrated strong performance on multi-task (ToxCast: 0.691 ROC-AUC) and regression tasks (LD50 MSE: 0.896, hREG MSE: 0.766).
• Outperformed Graph Convolution Network, Graph Attention Network, Graph Isomorphism Network, and others on all tested datasets.

Conclusions:

• The enhanced GNN effectively predicts drug toxicity by capturing complex molecular features.
• The proposed method offers a significant advancement over existing GNNs for toxicity assessment.
• This approach holds promise for accelerating safer drug discovery.