MGF-DTA: A Multi-Granularity Fusion Model for Drug-Target Binding Affinity Prediction

Area of Science:

• Computational Biology
• Drug Discovery
• Bioinformatics

Background:

• Drug-target affinity (DTA) prediction is crucial for efficient drug discovery.
• Existing methods struggle with limited drug multi-modal data, protein sequence length constraints, and inadequate feature extraction.

Purpose of the Study:

• To develop an advanced model, MGF-DTA, for improved drug-target binding affinity prediction.
• To address limitations in multi-modal drug information and multi-scale feature capture in protein sequences.

Main Methods:

• Utilized ChemBERTa-2 for deep semantic feature extraction from SMILES strings and integrated with molecular fingerprints via gated fusion.
• Employed residual fusion to combine global protein embeddings (ESM-2) with local k-mer and PCA features.
• Implemented a hierarchical attention mechanism for multi-granularity feature extraction from both drug and protein sequences.

Main Results:

• MGF-DTA demonstrated superior performance compared to mainstream methods on the Davis, KIBA, and BindingDB datasets.
• Ablation studies confirmed the significant contribution of individual model components.
• Case studies highlighted the model's robust generalization capabilities.

Conclusions:

• The MGF-DTA model effectively enhances drug-target affinity prediction by leveraging multi-modal drug information and multi-granularity protein features.
• This approach offers a promising advancement for computational drug discovery and development.