CoAff-DTI: Fine-grained drug-target interaction prediction using pre-trained language models and affinity-guided...

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Accurate drug-target interaction (DTI) prediction is crucial for efficient drug discovery.
• Pre-trained language models (PLMs) excel at molecular and protein representations but struggle with fine-grained biochemical interactions.
• Existing global embedding methods limit predictive accuracy and biological interpretability of DTIs.

Purpose of the Study:

• To develop an advanced deep learning framework, CoAff-DTI, for enhanced multi-scale interaction modeling in DTI prediction.
• To address the challenge of underrepresented localized interaction patterns in DTI prediction.
• To improve both the accuracy and biological interpretability of DTI prediction models.

Main Methods:

• CoAff-DTI employs a token-level decomposition strategy to generate pharmacophore- and residue-level representations from global embeddings.
• An Affinity-Guided Cross-Attention (AGCA) module explicitly models fine-grained interactions between ligand substructures and protein residues.
• An Affinity-Gating Fusion (AGF) module dynamically integrates cross-modal features for enhanced DTI prediction.

Main Results:

• CoAff-DTI consistently outperformed state-of-the-art methods across multiple benchmark datasets.
• Attention-based visualizations demonstrated improved model interpretability.
• Learned attention patterns in CoAff-DTI effectively aligned with experimentally verified protein binding regions.

Conclusions:

• CoAff-DTI offers a robust framework for accurate and interpretable DTI prediction.
• The model's multi-scale interaction modeling capabilities advance computational drug discovery.
• CoAff-DTI's ability to capture localized features represents a significant improvement over conventional global embedding approaches.