MultiKD-DTA: Enhancing Drug-Target Affinity Prediction Through Multiscale Feature Extraction.

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Traditional drug discovery is costly, time-consuming, and faces safety challenges.
• Manual screening of drug molecules against protein targets is slow and limited.
• Deep learning offers a promising approach to predict drug-target interaction (DTI) affinities.

Purpose of the Study:

• To introduce an innovative deep learning architecture for enhanced prediction of DTI affinities.
• To overcome the limitations of traditional manual screening methods in drug discovery.

Main Methods:

• Utilized graph neural networks (GNNs) and multiscale convolutional networks for molecular graph feature extraction.
• Employed pre-trained ESM-2 large models and bidirectional long short-term memory (BiLSTM) networks for protein sequence encoding.
• Integrated molecular and protein embeddings using a fusion module for affinity score computation.

Main Results:

• The proposed deep learning model demonstrated superior performance compared to existing methods.
• The model achieved enhanced prediction accuracy on two public DTI datasets.
• The architecture effectively combines GNNs, large protein models, and attention mechanisms.

Conclusions:

• The developed deep learning framework significantly improves the prediction of DTI affinities.
• This approach holds substantial promise for accelerating pharmaceutical agent discovery.
• The model represents an innovative advancement in computational drug discovery.