PLMCA: A General Multimodal Protein-Ligand Cross-Attention Framework for Pocket Identification and Binding Affinity...

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Accurate drug-target binding affinity (DTA) prediction is essential for efficient drug discovery.
• Integrating diverse data types (sequence, structure, physicochemical properties) remains a challenge.
• Experimental assay conditions can introduce noise and batch effects.

Purpose of the Study:

• To develop a multimodal framework, PLMCA, for accurate DTA prediction.
• To unify various data representations including protein sequence embeddings, 3D geometric features, physicochemical descriptors, and ligand molecular graphs.
• To incorporate experimental assay conditions to improve prediction robustness.

Main Methods:

• PLMCA utilizes a multimodal protein-ligand cross-attention mechanism.
• It integrates protein sequence embeddings from two language models.
• It incorporates 3D geometric features, physicochemical descriptors, ligand molecular graphs, and ChEMBL experimental assay conditions.

Main Results:

• PLMCA achieves competitive or superior performance compared to state-of-the-art methods on the PDBbind21 dataset for Kd and Ki prediction across various splits.
• On the ChEMBL_mini dataset, PLMCA yields R² values of 0.531 (IC50), 0.635 (Kd), and 0.519 (Ki).
• PLMCA demonstrates strong protein binding pocket prediction, achieving an AUPR of up to 0.655 in the unseen-protein setting.

Conclusions:

• PLMCA effectively integrates multimodal data for enhanced DTA prediction.
• The framework shows promise in both affinity prediction and binding pocket identification.
• PLMCA offers a robust approach to address challenges in computational drug discovery.