DeepHVI: A multimodal deep learning framework for predicting human-virus protein-protein interactions using protein...

Area of Science:

• Virology
• Computational Biology
• Immunology

Background:

• Understanding human-virus protein-protein interactions is crucial for deciphering viral pathogenesis, immune responses, and disease spread.
• Existing methods for identifying these interactions are often limited in scope or efficiency.

Purpose of the Study:

• To develop a novel multimodal deep learning framework for systematic prediction of human-virus protein-protein interactions.
• To enhance the accuracy and efficiency of identifying biologically relevant protein partnerships.

Main Methods:

• Integration of high-confidence experimental datasets with a multimodal deep learning approach.
• Utilizing complementary tasks: binary classification for interaction prediction and conditional sequence generation.
• Leveraging protein language models and multimodal fusion techniques.

Main Results:

• The framework demonstrated improved accuracy in predicting biologically relevant human-virus protein interactions.
• Application to severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) identified novel candidate interactions, including those not present in the training data.
• Several predicted interactions were corroborated by independent studies.

Conclusions:

• The developed deep learning framework offers a powerful tool for rapid, cost-effective discovery of protein interactions critical for public health.
• Predictions provide valuable insights into potential therapeutic targets for antiviral drugs and vaccines, contributing to pandemic preparedness.
• This approach facilitates the design of novel interventions against emerging infectious diseases.