Multi-task learning to leverage partially annotated data for PPI interface prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in biology

Background:

• Protein-protein interactions (PPIs) are vital for cellular functions.
• Predicting residues involved in PPIs from protein sequences is challenging.
• Limited availability of residue-level PPI interface annotations hinders deep learning approaches.

Purpose of the Study:

• To develop a deep learning strategy to predict protein-protein interaction sites.
• To address the challenge of limited and missing data in PPI interface annotations.
• To improve the prediction of functional properties of proteins using multi-task learning.

Main Methods:

• Implemented a multi-task learning model architecture adapted to handle missing data in the cost function.
• Included related learning tasks: secondary structure prediction, solvent accessibility, and buried residue prediction.
• Extended the dataset with structural feature data without additional PPI annotations.

Main Results:

• The multi-task learning strategy significantly outperformed single-task approaches.
• The multi-task strategy effectively learned from extended datasets with structural features, even without extra PPI annotations.
• Performance comparable to a single-task learner on all data was achieved by the multi-task learner using only one-eighth of the PPI annotations.

Conclusions:

• Multi-task learning is beneficial for datasets with partially annotated functional properties, such as PPI interfaces.
• This approach enhances prediction accuracy despite limited training data.
• The strategy effectively leverages related biological tasks to improve performance on the primary prediction task.