Contextual Representation Learning With ResNet Refinement and Deep Classification for Enhanced Drug Target...

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in drug discovery

Background:

• Accurate prediction of drug target interactions (DTIs) is crucial for drug discovery and repositioning.
• Existing computational methods struggle with the complexity of proteins and drugs, limiting their generalization capabilities.
• Predicting DTIs is vital for computer-aided drug design, especially for complex diseases involving multiple targets.

Purpose of the Study:

• To develop a robust deep learning framework to enhance DTI prediction accuracy.
• To effectively capture contextual and biochemical features from protein and drug representations.
• To improve drug discovery and repurposing through enhanced DTI prediction.

Main Methods:

• Leveraging pre-trained BERT-based language models for contextual embeddings from protein and drug sequences.
• Refining modality-specific representations using a ResNet-based subnetwork to preserve biochemical characteristics.
• Integrating refined embeddings and using a deep feedforward neural network for DTI prediction.

Main Results:

• Consistent performance improvements over baseline methods on four benchmark DTI datasets (DrugBank, Caenorhabditis elegans, BindingDB, GPCR).
• Significant F1-score gains of ~6.6% on GPCR and 3.7% classification accuracy increase on BindingDB.
• A 6.7% performance gain on an independent drug repurposing dataset, demonstrating robustness and practical applicability.

Conclusions:

• The proposed framework effectively captures contextual and structural information for improved DTI prediction.
• Enhanced prediction accuracy and generalization capabilities highlight the model's robustness.
• The framework shows practical applicability for accelerating drug discovery and drug repurposing efforts.