Application of Attention and Graph Transformer-Based Approaches for RNA Biomarker Discovery in Metabolically-Associated Fatty Liver Disease (MAFL/NASH)

Abstract

The prevalence of nonalcoholic fatty liver disease (NAFLD) and nonalcoholic steatohepatitis (NASH) in the United States has reached epidemic proportions, increasing the risk of liver cirrhosis and cancer. Current methods of diagnosis for NAFLD/NASH are invasive and costly, motivating the need for genetic "RNA" biomarkers detectable in a blood sample. In this study, explainable artificial intelligence (XAI) techniques are employed to increase the interpretability of the deep learning models in detecting the potential mRNA biomarker candidates for NAFLD/NASH. Nine RNA datasets ([~]1000 patients) with NAFLD/NASH were collected from the Gene Expression Omnibus. After conducting a differential gene expression analysis to reduce the dimensionality of the expression data, single-head and multi-head attention models were compared to baseline machine learning models in their ability to classify patients as NAFLD/NASH/healthy. XAI methods, including L1 regularization on baseline models and analysis of the internal attention matrix of the attention models, were utilized to identify biomarker candidates based on the relative importance of genes. The attention models achieved superior performance (accuracy: 67.5%) compared to the baseline models (Negative Binomial Linear Discriminant Analysis-62.64%; Poisson Linear Discriminant Analysis with Power Transformation - 58.24%). The top 17 and top 20 XAI-identified biomarkers with the baseline machine learning algorithms and the attention-based models respectively were then evaluated in lab. Preliminary data from in-lab validation confirmed upregulation of MT-ND3, HLA-B, APOC-1, and APOL-1 in NAFLD/NASH patients. Attention models have shown promise in identifying expression-based mRNA biomarkers and accurately diagnosing patients with NAFLD/NASH.