Integrating whole-body MRI-derived body composition with multi-omics data in the UK Biobank study

Sammanfattning

Medical imaging and multi-omics data provide complementary information about human physiology and disease. In this thesis, plasma proteomics (n≈9,000) and metabolomics (n≈95,000) from the UK Biobank were used to predict future body composition features derived from whole-body Dixon MRI scans (n≈95,000) acquired up to 15 years later. To extract body composition features, the MRI images were segmented using the automated AI-based segmentation method VIBESegmentator with additional post-processing to extract volume and fat fraction measurements for 83 different tissues and organs. Sex-stratified significant univariate associations between omics and body composition features were estimated with Pearson r-values with an absolute value of up to 0.70 in proteins and up to 0.39 in metabolites. The variables with significant associations were subsequently used to model multivariable associations using Partial Least Squares regression, evaluated using R2 values from repeated cross-validation. The sex-stratified multivariable analysis exhibited stronger associations for proteomics than for metabolomics, for both volumetric and fat fraction features. The results showed strong associations between multi-omics data and future body composition, particularly for fat depots where the visceral adipose tissue (VAT) exhibited R2 values of up to 0.53 and subcutaneous adipose tissue (SAT) exhibited values of up to 0.61, and the volumes of the spleen and liver with values of up to 0.57 and 0.41 respectively.