Learning heritable multimodal brain representation via contrastive learning.

Area of Science:

• Neuroimaging
• Genetics
• Machine Learning

Background:

• Magnetic resonance imaging (MRI)-derived phenotypes (IDPs) have facilitated the discovery of genomic loci linked to brain structure and function.
• Current IDPs often rely on single imaging modalities, potentially limiting comprehensive genetic discovery by overlooking cross-modal information.

Purpose of the Study:

• To introduce a multimodal contrastive learning framework for deriving heritable brain representations from paired T1- and T2-weighted MRIs.
• To enhance genetic discovery by integrating complementary information across different MRI modalities.

Main Methods:

• Developed a momentum-based contrastive learning framework utilizing paired T1- and T2-weighted MRI data.
• Applied the framework to derive shared representations across modalities, contrasting with single-modality reconstruction methods.

Main Results:

• The multimodal approach improved predictions of traditional IDPs, age, and brain disorders.
• Genome-wide association studies (GWAS) on learned representations showed significantly higher overlap of genetic loci across modalities.
• Identified shared protein and drug targets from GWAS loci, providing biological insights.

Conclusions:

• The proposed framework effectively learns shared representations across brain imaging modalities.
• These representations exhibit enhanced anatomical and genetic coherence, improving the integration of neuroimaging and genetic data.
• This multimodal approach offers a promising avenue for more comprehensive genetic discovery in neuroscience.