Improving classification of correct and incorrect protein-protein docking models by augmenting the training set

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in bioinformatics

Background:

• Protein-protein interactions are crucial for biological processes.
• Experimental determination of 3D structures is time-consuming and costly.
• Computational protein-protein docking generates numerous poses, creating imbalanced datasets unsuitable for machine learning.

Purpose of the Study:

• To address the challenge of imbalanced datasets in protein-protein docking.
• To develop a data augmentation method for improving machine learning model training.
• To enhance the accuracy of protein-protein docking pose prediction.

Main Methods:

• Developed a weak supervision-based data augmentation technique named hAIkal.
• Utilized hAIkal to increase the volume of labeled training data.
• Trained and evaluated multiple machine learning classifiers on the augmented dataset.

Main Results:

• The hAIkal method significantly increased the labeled training data for docking classifiers.
• The best-trained classifier achieved 81% accuracy and a 0.51 Matthews' correlation coefficient on the test set.
• The developed classifier outperformed existing state-of-the-art scoring functions in protein-protein docking.

Conclusions:

• Weak supervision and data augmentation, as implemented in hAIkal, are effective for improving protein-protein docking.
• The enhanced classifiers provide a more accurate prediction of protein interaction poses.
• This approach offers a viable computational solution to complement experimental structure determination.