Abstract MP18: Comparative Assessment of Proteomics, Phenotypic, and Genomic Data on Visceral Adipose Tissue Volume Using Machine Learning in UK Biobank: Advancing Precision Medicine

• DOI: 10.1161/circ.149.suppl_1.mp18
• Publication Date: 2024-05-16T14:05:30Z
• Authors: Zahra Azizi; Trisha Gupte; Pik Fang Kho; Kevin Nzenkue; Jiayan Zhou; Rodrigo Guarischi...; Daniel Panyard; Ming-Li Chen; Fahim Abbasi; Shoa L CLARKE; Philip S Tsao; Themistocles L As...
• Keywords: 03 medical and health sciences; 0302 clinical medicine

Background: Adiposity is a key modulator for Insulin resistance (IR); a strong risk factor for type 2 diabetes and cardiovascular diseases. The blood proteome is a crucial indicator of biological processes associated with complex disorders. In this study, we investigate the associations between plasma proteomics and visceral adipose tissue volume (VAT), and their intersection with phenotypic and genomic data to identify critical proteins for adiposity linked to IR.

Methods: We examined UK biobank participants with proteomic data generated using the Olink proximity extension antibody assay (PEA). Interactions between, polygenic scores, biomarkers, socio-behavioral, and fitness factors were investigated to identify the most important proteins associated with VAT (Figure). Multi-step machine learning algorithms including separate models for each feature type, followed by a combined model were employed. We used R2 as an indicator of the variance of VAT explained by the features in each model and reported SHAP values for the top 30 features.

Results: We analyzed a cohort of 5,342 participants (females: 52.2%, mean age: 54.8 ± 7.9 years). Significant variation of VAT with plasma proteins was observed. Polygenic scores, phenotypic, proteomics, and the combined model explained 1%, 62%, 64%, and 67% of the variation in VAT, respectively. The top proteins in the combined model were CDHR2, ACY1, PRSS8, PSPN, LPL, PLAT, FABP4, COL4A1, CPM, CTSD, SSC4D, PON3, CDH15, DSG2, CNTN3, IGFBP2, CCL16, ADM< WFDC12, OXT, CNTN5, PRCP, ADGRG2, and LEP. The model that included only the top 30 features explained 66% of the variation in VAT.

Conclusion: A small number of plasma proteins measured with the PEA explain a significant portion of the variance of VAT beyond non-proteomic measures. Further validation and incorporation of additional protein-phenotype models may enable a single-source, personalized measurement of adiposity that is easily accessible and may facilitate further understanding the biology of IR and adverse health outcomes.