Machine learning nominates the inositol pathway and novel genes in Parkinson’s disease

Genome-wide association studies (GWAS) have nominated many variants associated with complex traits. In Parkinson’s disease (PD), the most recent GWAS revealed 90 independent risk variants across 78 genomic loci. Although many single-nucleotide polymorphisms (SNPs) are in novel genomic loci, well-established PD genes discovered many years ago, such as LRRK2, PINK1, DJ-1, SNCA, GBA1, PRKN and MAPT still account for the vast majority of research on Parkinson’s disease. Several disadvantages of GWAS limit additional functional analyses. First, above 90% of all GWAS significant SNPs are in noncoding regions. These SNPs are often passenger variants due to complex linkage disequilibrium (LD). Second, the causal gene associated with the causal SNPs remains unclear in most GWAS loci. To overcome these challenges, downstream GWAS analyses were established with the aim of identifying causal genes within GWAS loci. This involves techniques such as fine-mapping and colocalization methods to nominate causal SNPs, as well as transcriptome-wide association studies to nominate gene-trait associations. These models use LD structure, and gene expression panels to discover causal SNPs/genes. While these methods may propose causal variants and genes, additional biological evidence is generally required to pair causal variants with causal genes. Using multi-omic analyses, one can integrate a diverse range of comprehensive cellular and biological datasets such as genomic, transcriptomic and epigenetic datasets and use platforms such as Open Targets Genetics (https://genetics.opentargets.org/) to perform systematic analyses of gene prioritization across all publicly available GWASs. Although powerful, Open Targets Genetics lacks disease-specific tissues relevant to PD such as dopaminergic neurons and microglia. Using a similar approach, we may discover additional pathways and genetic targets involved in PD. In this study, we leveraged PD-relevant transcriptomic, epigenomic and other datasets in our gradient boosting model (Figure 1). We trained this model on well-established PD genes to nominate causal genes from PD GWAS loci.

Genomics, Transcriptomics, Epigenetics