Turbocharging protein binding site prediction with geometric attention, inter-resolution transfer learning, and...

Area of Science:

• Computational biology
• Drug discovery
• Structural bioinformatics

Background:

• Accurate identification of small molecule binding sites in proteins is crucial for drug discovery.
• Existing deep learning methods for binding site prediction face limitations in efficiency, information loss, and data utilization.

Purpose of the Study:

• To develop an improved deep learning model for predicting protein binding sites at both pocket and residue resolutions.
• To address limitations of current methods by enhancing architecture, reducing post-processing information loss, and maximizing data utilization.

Main Methods:

• A novel model architecture layering geometric self-attention units on 3D Convolutional Neural Network (CNN) outputs.
• Computation units configured as residues and pockets, rather than voxels, to minimize information loss.
• Inter-resolution transfer learning and homology-based data augmentation to maximize data source utilization.

Main Results:

• The proposed method significantly outperformed state-of-the-art baselines in both pocket and residue resolution binding site prediction.
• Ablation studies confirmed the effectiveness of the proposed architecture, transfer learning, and data augmentation.
• Case study on human serum albumin demonstrated superior capability in identifying multiple binding sites.

Conclusions:

• Introduced a novel computational method for binding site prediction with practical applications and strong performance.
• The developed model architecture, transfer learning, and augmentation strategies offer valuable components for future research in the field.