Multitask Matrix Completion for Learning Protein Interactions Across Diseases

Area of Science:

• Computational biology
• Infectious disease research
• Bioinformatics

Background:

• Viruses interact with host proteins for replication, crucial for understanding infectious diseases.
• Related viruses often share biological similarities, suggesting potential for joint modeling.
• Predicting pathogen-host protein interactions is vital for developing antiviral strategies.

Purpose of the Study:

• To develop a multitask learning method for jointly modeling human protein interactions with three related viruses: Hepatitis C, Ebola, and Influenza A.
• To improve the accuracy of predicting these complex interactions.
• To provide interpretable insights into virus-specific and general interaction characteristics.

Main Methods:

• A multitask matrix completion-based model was employed.
• The model incorporates a shared low-rank structure and task-specific sparse structures.
• This approach jointly analyzes interactions across different viral tasks.

Main Results:

• Achieved 7 to 39 percentage point improvement in predictive performance over existing state-of-the-art models.
• Demonstrated the model's ability to reveal general and specific interaction-relevant characteristics of the viruses.
• The developed code is publicly available for further research.

Conclusions:

• Multitask learning effectively models interactions between human proteins and related viruses.
• The proposed method offers significant performance gains in predicting these interactions.
• The model provides valuable insights for understanding infectious disease mechanisms.