Improving Compound-Protein Interaction Prediction by Self-Training with Augmenting Negative Samples.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Identifying compound-protein interactions (CPIs) is vital for drug discovery.
• Experimental validation of CPIs is costly and time-consuming.
• Computational methods, particularly machine learning, are used for CPI prediction but suffer from data imbalance due to a lack of negative samples.

Purpose of the Study:

• To develop a self-training method for augmenting credible and informative negative samples.
• To improve the performance and generalizability of machine learning models for CPI prediction, especially on external datasets.
• To provide guidelines for enhancing CPI predictions using real-world data.

Main Methods:

• Developed a self-training approach to generate negative samples for imbalanced CPI datasets.
• Evaluated model performance against conventional methods for addressing data imbalance.
• Analyzed the impact of pseudo-labeling thresholds on model generalizability.

Main Results:

• The proposed self-training method significantly improved model performance compared to conventional approaches.
• Performance gains were particularly notable when tested on external datasets, indicating enhanced generalizability.
• Augmenting samples with ambiguous prediction scores during self-training proved beneficial for model generalizability.

Conclusions:

• The developed self-training method effectively addresses data imbalance in CPI prediction.
• This approach enhances model generalizability, crucial for real-world drug discovery applications.
• The study offers practical guidelines for improving computational CPI prediction accuracy and efficiency.