Enhanced information cross-attention fusion for drug-target binding affinity prediction

Area of Science:

• Artificial Intelligence
• Computational Chemistry
• Drug Discovery

Background:

• Deep learning significantly enhances drug discovery efficiency, particularly in predicting drug-target affinity (DTA).
• Traditional DTA prediction methods often neglect crucial 3D structural and physicochemical properties of proteins and compounds.

Purpose of the Study:

• To develop a novel deep learning model, CAFIE-DTA, for improved drug-target affinity prediction.
• To incorporate 3D protein structural information (curvature, electrostatic potential) and compound physicochemical features into DTA prediction.

Main Methods:

• Proposed CAFIE-DTA model utilizes Delaunay triangulation for protein surface curvature approximation and APBS software for electrostatic potential calculation.
• Employs cross multi-head attention to fuse protein sequence, physicochemical information, and 3D structural properties.
• Integrates graph-based and physicochemical features of compounds using a similar attention mechanism.

Main Results:

• CAFIE-DTA demonstrated superior performance on the Davis and KIBA benchmark datasets compared to existing methods.
• Achieved significant improvements in R-squared and Mean Squared Error (MSE) on both datasets.
• Outperformed traditional models relying solely on 2D structures and sequence data for DTA prediction.

Conclusions:

• CAFIE-DTA offers a more accurate and comprehensive approach to drug-target affinity prediction.
• The integration of 3D structural and physicochemical information is crucial for advancing DTA prediction models.
• The developed model shows promise for accelerating drug discovery and development pipelines.