Machine learning across multiple imaging and biomarker modalities in the UK Biobank improves genetic discovery for liver fat accumulation

Abstract

Metabolic dysfunction-associated steatotic liver disease (MASLD), liver with more than 5.5% fat content, is a leading risk factor for chronic liver disease with an estimated worldwide prevalence of 30%. Though MASLD is widely recognized to be polygenic, genetic discovery has been lacking primarily due to the need for accurate and scalable phenotyping, which proves to be costly, time-intensive and variable in quality. Here, we used machine learning (ML) to predict liver fat content using three different data modalities available in the UK Biobank: dual-energy X-ray absorptiometry (DXA; n = 46,461 participants), plasma metabolites (n = 82,138), and anthropometric and blood-based biochemical measures (biomarkers; n = 262,927). Based on our estimates, up to 29% of participants in UKB met the criteria for MASLD. Genome-wide association studies (GWASs) of these estimates identified 15, 55, and 314 loci associated with liver fat predicted from DXA, metabolites and biomarkers, respectively, totalling 321 unique independent loci. In addition to replicating 9 of the 14 known loci at genome-wide significance, our GWASs identified 312 novel loci, significantly expanding our understanding of the genetic contributions to liver fat accumulation. Genetic correlation analysis indicated a strong correlation between ML-derived liver fat across modalities (rgranging from 0.85 to 0.96) and with clinically diagnosed MASLD (rgranging from 0.74 to 0.88), suggesting that a majority of the newly identified loci are likely to be relevant for clinical MASLD. DXA exhibited the highest precision, while biomarkers demonstrated the highest recall, respectively. Overall, these findings demonstrate the value of leveraging ML-based trait predictions across orthogonal data sources to improve our understanding of the genetic architecture of complex diseases.

Genetics-driven Risk Predictions with Differentiable Mendelian Randomization

“…This feature allowed us to verify that DMR learns plausible nonlinearities for diabetes biomarkers. DMR's flexibility, facilitated by its PyTorch implementation, promises broad applicability, from trait embeddings derived from unsupervised models [27,28] to focused biomarker discovery, such as to define new anthropometric biomarkers [29] or blood clinical scores [30].…”

HistoGWAS: An AI Framework for Automated and Interpretable Genetic Analysis of Tissue Phenotypes

“…Finally, the approach of integrating advanced AI tools for the encoding and decoding high-content phenotypes with multivariate statistical analysis of latent spaces is applicable to a wide array of high-content phenotypic modalities and use cases. For example, the approach implemented in HistoGWAS can facilitate the study of genetic effects on morphological changes visible through non-invasive medical imaging [8][9][10][11][12][13][14] , and enable the assessment of the impacts of genetic and chemical perturbations in high-content screens of in vitro systems 62 . This is particularly relevant in advanced systems like organoids, enabling a detailed description and visualization of complex phenotypic changes resulting from induced perturbations.…”

“…provided insights into the functional mechanisms of loci implicated in human diseases 1 through the identification of key genes and metabolites influenced by disease loci [1][2][3][4][5][6][7] . Extending these analyses to medical imaging-derived traits has further illuminated the intermediate effects of genetic loci [8][9][10][11][12][13][14][15] . Despite these advances, comprehensive genetic analyses of the rich tissue and cellular details available in histological images are lacking.…”