Contextual AI models for single-cell protein biology.

Area of Science:

• Computational biology
• Genomics
• Systems biology

Background:

• Understanding protein function and interactions is crucial for molecular therapies.
• Existing algorithms struggle to model protein interactions across diverse biological contexts.
• Cell type and tissue specificity are key determinants of protein behavior.

Purpose of the Study:

• To introduce Pinnacle, a novel geometric deep learning approach for generating context-aware protein representations.
• To leverage a multi-organ single-cell atlas to train Pinnacle on contextualized protein interaction networks.
• To enable accurate modeling of protein interactions within specific cellular and tissue environments.

Main Methods:

• Developed Pinnacle, a geometric deep learning framework.
• Utilized a multi-organ single-cell atlas comprising 156 cell type contexts across 24 tissues.
• Generated 394,760 context-aware protein representations.
• Evaluated Pinnacle's performance on downstream tasks including 3D structure-based representation enhancement and drug effect investigation.

Main Results:

• Pinnacle's embedding space effectively captures cellular and tissue organization, enabling zero-shot retrieval of tissue hierarchy.
• Pretrained protein representations demonstrated adaptability for enhancing immuno-oncological protein interaction resolution and investigating drug effects.
• Pinnacle outperformed state-of-the-art models in identifying therapeutic targets for rheumatoid arthritis and inflammatory bowel diseases.
• Identified cell type contexts with superior predictive capability compared to context-free models.

Conclusions:

• Pinnacle provides a powerful new tool for generating context-specific protein representations.
• The approach facilitates large-scale, context-aware predictions in biological systems.
• Pinnacle enhances the discovery of therapeutic targets and understanding of disease mechanisms across different cell types and tissues.