COLDLNA: Enhancing long-range node features extraction to improve robust generalization ability of drug-target...

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Deep learning significantly advances drug-target affinity (DTA) prediction.
• Existing models often fail to fully leverage drug molecular graphs and long-range protein features, impacting accuracy.

Purpose of the Study:

• To introduce COLDLNA, a novel model for robust DTA prediction.
• To improve the utilization of drug structure and protein sequence information in DTA prediction.

Main Methods:

• COLDLNA utilizes a Long-range Node Attention Module for drug structure representation.
• A Convolutional Attention Module extracts long-range protein sequence information to identify binding sites.

Main Results:

• COLDLNA reduced Mean Squared Error (MSE) by 12.2% on the Davis dataset and 11.5% on the KIBA dataset compared to GraphDTA.
• The model demonstrated strong generalization on Human and C. elegans datasets, including cold-start scenarios.

Conclusions:

• COLDLNA offers a more effective approach to DTA prediction by integrating advanced graph and sequence feature extraction.
• The model's performance and generalization capabilities show promise for accelerating drug discovery and development.