A foundation model for learning genetic associations from brain imaging phenotypes.

Area of Science:

• Computational Biology
• Neuroscience
• Genetics

Background:

• Neurological disorders have complex causes, making it difficult to link multiomics data with standard methods.
• Interpretable associations between genetic markers and brain imaging phenotypes are crucial for understanding disease etiology.

Purpose of the Study:

• To introduce COMICAL, a contrastive learning framework for identifying many-to-many associations between multiomics data and brain imaging-derived phenotypes.
• To leverage transformer-based encoders and self-supervised learning for joint omics representation.

Main Methods:

• COMICAL employs transformer-based encoders with custom tokenizers for joint learning of omics representations.
• A modality-agnostic approach utilizes self-supervised learning and cross-modal attention to identify complex associations.
• The UK Biobank dataset was used to validate the method's effectiveness.

Main Results:

• COMICAL identified significant associations between genetic markers and imaging-derived phenotypes across various neurological disorders.
• The model demonstrated the ability to predict diseases and unseen clinical outcomes from learned representations.
• The approach successfully generated interpretable, many-to-many associations.

Conclusions:

• COMICAL offers a powerful new approach for dissecting the genetic underpinnings of neurological disorders using multiomics data.
• The framework facilitates the discovery of novel biomarkers and prediction of disease trajectories.
• Open-source code and pretrained weights are available for transfer learning and further research.