A multi-network approach to Alzheimers Disease gene prioritization complements cis-regulatory prioritization with molecular quantitative trait loci

Gene prioritization within mapped disease-risk loci from genome-wide association studies (GWAS) remains one of the central bioinformatic challenges of human genetics. This problem is abundantly clear in Alzheimers Disease (AD) which has several dozen risk loci, but no therapeutically effective drug target. Dominant strategies emphasize alignment between molecular quantitative trait loci (mQTLs) and disease risk loci, under the assumption that cis-regulatory drivers of gene expression or protein abundance mediate disease risk. However, mQTL data do not capture clinically relevant time points or they derive from bulk tissue. These limitations are particularly significant in complex diseases like AD where access to diseased tissue occurs only in end-stage disease, while genetically encoded risk events accumulate over a lifetime. Network-based functional predictions, where bioinformatic databases of gene interaction networks are used to learn disease-associated gene networks to prioritize genes, complement mQTL-based prioritization. The choice of input network, however, can have a profound impact on the output gene rankings, and the optimal tissue network may not be known a priori. Here, we develop a natural extension of the popular NetWAS approach to gene prioritization that allows us to combine information from multiple networks at once. We applied our multi-network (MNFP) approach to AD GWAS data to prioritize candidate genes and compared the results to baseline, single-network models. Finally, we applied the models to prioritize genes in recently mapped AD risk loci and compared our prioritizations to the state-of-the-art mQTL approach used to functionally prioritize genes within those loci. We observed a significant concordance between the top candidates prioritized by our MNFP method and those prioritized by the mQTL approach. Our results show that network-based functional predictions are a strong complement to mQTL-based approaches and are significant to the AD genetics community as they provide a strong functional rationale to mechanistically follow-up novel AD-risk candidates. Author SummaryRisk genes give us insight into the failing molecular mechanisms that drive disease phenotypes. However, these risk genes are several layers of complexity removed from the emergent phenotypes they are influencing, the p-value that denotes their risk status gives little insight into the functional implications of that risk, and it is not clear when that risk gene may be having its effect. Methods like network-based functional prediction start to address several of these limitations by contextualizing risk genes in their broader genetic neighborhood within disease-relevant tissues. For complex diseases like Alzheimers, there are many possible relevant tissues incorporating everything from individual brain cell types to whole lobes of the brain. The work in this paper expands upon the traditional network-based functional prediction approach by considering a genes connections in multiple relevant tissue networks to prioritize candidate genes. Unlike traditional genetic risk studies, this kind prioritization benefits the Alzheimers genetics community as it provides a strong functional rationale to mechanistically follow-up on novel gene candidates.