Deep learning-based predictive modeling of PNPLA3 variant carriers using radiomics data from the UK Biobank

Background Steatotic liver disease (SLD) represents a prevalent public health concern.Genetic factors play a pivotal role in SLD's development, especially the patatin-like phospholipase domain-containing protein 3 (PNPLA3) gene, with its most prominent variant rs738409_G. Identification of rs738409_G carriers based on genome analysis is expensive. Therefore, we propose a novel approach to discern rs738409_G carriers utilizing liver magnetic resonance imaging (MRI) data.

Methods We selected UK Biobank patients for which both data on PNPLA3 rs738409_G status (N_homozygous=2041, N_non-carrier=26756) and MRI data was available. To extract the liver fat fraction we calculated r2*-corrected water and fat images based on the IDEAL post-processing technique . The liver was subsequently segmented by training a UNet model and radiomics features were extracted. We then trained a Random Forest classifier using extracted radiomics features for a matched group of non-carriers and homozygotes to predict homozygous carriers and non-carriers.

Results The data was divided into the fatty liver group (1018 controls and carriers, respectively) and normal liver fat group (803 controls and carriers each) based on the fat content (threshold 5%) and the prediction achieved an AUROC of 0.62 and 0.54 on the independent test sets respectively, which indicates that the livers of SLD patients contain more features to differentiate between homozygous carriers and non-carriers.

Conclusion Utilizing a well-characterized and large cohort, we demonstrate a novel approach to identify PNPLA3 rs738409_G carriers. Our approach introduces a streamlined and accessible method for identifying carriers, which holds great promise for advancing the field of SLD diagnostics and personalized medicine.

Publication History

Article published online:
23 January 2024