Machine learning using genotype and gene-expression data identifies alterations of genes involved in infection susceptibility, antigen presentation and cytokine signalling as key contributors to JIA risk prediction.

Abstract

Background: Previous genome-wide association studies (GWAS) have identified numerous genetic loci associated with juvenile idiopathic arthritis (JIA). However, the functional impact of these variants-particularly on tissue-specific gene expression-and which regulatory interactions make the greatest relative contribution to JIA risk remain unclear. Identifying these key single-nucleotide polymorphism (SNP)-gene-tissue combinations can help prioritise targets for future functional studies and therapeutic interventions.

Method: We performed two-sample Mendelian randomisation (2SMR) using spatial expression quantitative trait loci (eQTLs) from nine tissue-specific gene-regulatory networks as instrumental variables (IVs). We also identified JIA-associated SNPs from previous GWAS and mapped their spatial eQTL effects across 49 human tissues. These SNP sets were then used as features in a Lasso-regularised logistic regression model to predict JIA disease status. The model weight magnitudes served as proxies for each SNP's contribution to JIA risk. We evaluated the robustness of our model's feature ranking across 50 cross-validation runs.

Results: The top-ranked SNPs included rs7775055, which tags the human leukocyte antigen (HLA) class II haplotype DRB1*0801-DQA1*0401-DQB1*0402, and rs6679677, a non-coding variant that is in 100% linkage with with a coding variant in PTPN22. IVs for genes implicated in infection-related immune processes (eg, MSH5, MICA and LINC01149) also made significant contributions to JIA risk. We additionally identified a spatial eQTL (rs10849448) that upregulated the cytokine signalling gene LTBR across all 49 tissues. Overall, our model highlighted the roles of genes involved in antigen presentation, infection susceptibility and cytokine signalling.

Conclusion: By applying a machine learning approach to rank SNP-gene-tissue contributions to JIA risk, our findings offer insights into the genetic mechanisms underlying JIA pathogenesis. Future experimental validation could facilitate new therapeutic targets for the treatment or prevention of JIA.