An adversarial scheme for integrating multi-modal data on protein function.

Area of Science:

• Cellular Biology
• Bioinformatics
• Computational Biology

Background:

• Understanding the cell's hierarchical structure requires integrating multi-scale data on subcellular organization.
• Existing methods struggle to comprehensively combine diverse data types like protein sequence, interactions, and localization.

Purpose of the Study:

• To develop a novel multi-modal generative model, MIRAGE, for integrating diverse biological data.
• To create a joint embedding space capturing complex relationships between protein sequence, protein-protein interactions, and protein localization.
• To enable the generation of missing data modalities and improve downstream biological predictions.

Main Methods:

• Developed MIRAGE, a multi-modal generative adversarial network.
• Integrated protein sequence, protein-protein interaction, and protein localization data.
• Learned a joint embedding space representing multi-modal biological data.

Main Results:

• Achieved superior performance in protein function prediction compared to existing methods.
• Demonstrated enhanced capabilities in protein complex detection.
• Successfully constructed a hierarchical map of subcellular organization in HEK293T cells.
• Recovered known protein assemblies across multiple scales.

Conclusions:

• MIRAGE effectively integrates diverse biological data modalities.
• The model provides a powerful framework for deciphering cellular hierarchy and subcellular organization.
• MIRAGE advances the prediction of protein function and complex detection, aiding biological discovery.