TextDTI: A Multimodal Context Representation Learning Framework for Drug-Target Interaction Prediction

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Predicting drug-target interactions (DTIs) is vital for efficient drug discovery.
• Deep learning, especially Large Language Models (LLMs), shows potential for encoding drug (SMILES) and protein sequence data.
• Integrating diverse data modalities for DTI prediction remains a significant challenge.

Purpose of the Study:

• To propose TextDTI, a novel multimodal framework for predicting drug-target interactions.
• To effectively integrate sequential and structural representations of drugs and proteins.
• To enhance the accuracy and robustness of DTI prediction models.

Main Methods:

• Utilizing Pretrained Language Models (PLMs) to generate functional descriptions for protein sequences.
• Encoding drug SMILES and protein functional descriptions using separate LLMs.
• Fusing drug and target features via convolutional and graph-based modules.
• Classifying DTIs using a multilayer perceptron, enhanced by adversarial learning and contrastive loss.

Main Results:

• TextDTI demonstrates superior performance in DTI prediction across multiple datasets.
• The model achieves robust results in both single-domain and cross-domain prediction settings.
• Experimental validation confirms the effectiveness of the proposed multimodal approach.

Conclusions:

• TextDTI offers a powerful new approach for predicting drug-target interactions.
• The multimodal framework successfully integrates diverse data types for enhanced prediction.
• This work advances the application of LLMs in computational drug discovery.