Adaptive Self-Attention Graph Pooling for Drug-Target Affinity Prediction.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Drug-target affinity (DTA) prediction is crucial for drug discovery and precision medicine.
• Existing graph neural network (GNN) pooling methods struggle with molecular and protein structural diversity, causing information loss.
• Adaptive pooling is needed to handle varying data structures effectively.

Purpose of the Study:

• To develop an adaptive graph pooling mechanism for improved DTA prediction.
• To create a multi-modal framework integrating GNNs and Transformers for DTA analysis.
• To enhance model interpretability and computational efficiency in DTA prediction.

Main Methods:

• Proposed the Adaptive Self-Attention Graph Pooling (ASAGPooling) mechanism with a learnable pooling ratio.
• Developed the ASAG-DTA framework, integrating GNNs and Transformers.
• Modeled molecular graphs, protein contact maps, SMILES, and FASTA sequences.

Main Results:

• ASAGPooling achieved competitive DTA prediction accuracy (MSE = 0.186 on Davis).
• The ASAG-DTA framework demonstrated adaptability, interpretability, and computational efficiency.
• ASAGPooling eliminates manual pooling ratio tuning and allows visualization of key residues/atoms.

Conclusions:

• ASAGPooling provides a practical, lightweight alternative for DTA prediction, especially in resource-constrained virtual screening.
• The proposed method enhances model interpretability and reduces complexity compared to existing approaches.
• ASAG-DTA offers a valuable tool for accelerating drug discovery pipelines.