A deep learning method for predicting molecular properties and compound-protein interactions.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Predicting molecular properties and compound-protein interactions (CPIs) are crucial for drug design and virtual screening.
• Deep learning methods have shown significant promise in various computational challenges.
• There is a need for efficient deep learning models to predict both molecular properties and CPIs accurately.

Purpose of the Study:

• To propose a novel deep learning method, MG-S, for predicting both molecular properties and CPIs.
• To leverage integrated chemical structure and sequence information for enhanced prediction accuracy.
• To provide a versatile computational tool for early-stage drug design and discovery.

Main Methods:

• Developed the MG-S deep learning model integrating compound structure and protein sequence information.
• Combined topological structure and sequence fingerprint information for feature extraction.
• Validated the model on established datasets for molecular property prediction (BACE, P53, hERG) and CPI prediction (Human, C. elegans, KIBA).

Main Results:

• MG-S demonstrated superior performance in molecular property prediction on the P53 dataset.
• The model achieved consistently good results on BACE and hERG datasets.
• MG-S exhibited impressive performance in CPI prediction on the KIBA dataset, outperforming state-of-the-art models in AUC, Precision, and MCC.
• The model showed higher performance, better classification ability, and faster convergence.

Conclusions:

• The MG-S model is an effective deep learning approach for predicting molecular properties and CPIs.
• Its ability to integrate diverse data types leads to improved prediction accuracy.
• MG-S offers a valuable tool for accelerating lead compound discovery in drug design and development.