MDTips: a multimodal-data-based drug-target interaction prediction system fusing knowledge, gene expression profile,...

Area of Science:

• Bioinformatics
• Computational Chemistry
• Drug Discovery

Background:

• Traditional experimental screening of drug-target interactions (DTIs) is resource-intensive.
• Computational DTI models leverage knowledge graphs, chemical notations, and genomic data for drug discovery.
• A unified framework for multimodal data fusion in DTI prediction is needed.

Purpose of the Study:

• To develop a multimodal fusion DTI prediction system integrating heterogeneous data sources.
• To enhance the accuracy and robustness of DTI prediction models.
• To facilitate drug repurposing and discovery through advanced computational methods.

Main Methods:

• Developed MDTips, a system fusing knowledge graphs, gene expression profiles, and drug/target structural information.
• Employed deep learning-based encoders, including Attentive FP and Transformer, for feature extraction.
• Validated performance against traditional chemical descriptors and state-of-the-art prediction models.

Main Results:

• MDTips demonstrated accurate and robust performance in DTI predictions.
• Multimodal fusion learning effectively incorporated information from diverse data aspects, improving model performance.
• Deep learning encoders outperformed traditional methods; MDTips surpassed existing state-of-the-art models.
• Successfully reverse-screened candidate targets for 6766 drugs, aiding drug repurposing.

Conclusions:

• MDTips provides a powerful, multimodal approach for DTI prediction.
• The system's ability to integrate diverse data enhances its utility in drug discovery pipelines.
• MDTips offers a valuable tool for identifying novel drug-target associations and facilitating drug repurposing.