Compound-protein interaction prediction by deep learning: Databases, descriptors and models.

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Screening compound-protein interactions (CPIs) is vital for identifying drug candidates.
• Traditional methods like high-throughput screening (HTS) face limitations in cost and efficiency.
• Deep learning (DL) presents a promising alternative for rapid and economical CPI prediction.

Purpose of the Study:

• To provide a comprehensive review of deep learning-based methods for compound-protein interaction (CPI) prediction.
• To survey existing databases, compound and protein representations, and state-of-the-art DL models.
• To identify current challenges and future trends in DL for CPI prediction.

Main Methods:

• Summarizing popular databases for small molecules, proteins, and binding complexes.
• Outlining classical and modern representations for compounds and proteins.
• Introducing state-of-the-art deep learning models and analyzing their prediction performance.

Main Results:

• Deep learning models demonstrate significant potential for accurate and efficient CPI prediction.
• The review categorizes DL models based on their design paradigms.
• Performance analysis highlights the strengths and weaknesses of various DL approaches.

Conclusions:

• Deep learning significantly enhances the efficiency and reduces the cost of CPI screening.
• Addressing current challenges in data, interpretability, and model generalizability is crucial.
• Future research should focus on practical applications and developing robust DL models for drug discovery.