Enhancing Local Functional Structure Features to Improve Drug-Target Interaction Prediction.

Area of Science:

• Computational chemistry
• Bioinformatics
• Artificial intelligence in drug discovery

Background:

• Molecular simulation is vital for drug discovery but computationally expensive.
• Existing deep learning models for drug-target interaction (DTI) often overlook critical local structural features.
• Identifying these local features is key to understanding molecular recognition and improving screening.

Purpose of the Study:

• To develop a deep learning framework, LoF-DTI, that explicitly represents and couples local functional structure features for enhanced DTI prediction.
• To improve the efficiency and accuracy of virtual screening in drug discovery.
• To provide interpretable insights into drug-target binding mechanisms.

Main Methods:

• Drugs represented as molecular graphs (from SMILES) and targets as sequences.
• Utilized a Jumping Knowledge (JK) enhanced Graph Isomorphism Network (GIN) for drug feature extraction.
• Employed residual CNNs with N-mer statistics for target motif capture.
• Integrated a Gated Cross-Attention (GCA) module for atom-to-residue interaction learning.

Main Results:

• LoF-DTI achieved competitive performance across multiple DTI prediction benchmarks.
• The framework demonstrated improved early retrieval rates crucial for virtual screening.
• Case studies successfully identified known functional binding sites, validating the model's interpretability.

Conclusions:

• LoF-DTI effectively prioritizes local features in both encoding and interaction for accurate DTI prediction.
• The model offers mechanism-aware guidance, potentially streamlining the drug design pipeline.
• LoF-DTI represents a significant advancement in leveraging deep learning for accelerated drug discovery.