Finding Genetic Overlaps Among Diseases Based on Ranked Gene Lists

Abstract

To understand disease relationships in terms of their genetic mechanisms, it is important to study the common genetic basis among different diseases. Although discoveries on pleiotropic genes related to multiple diseases abound, methods flexibly applicable to various types of datasets generated from different studies or experiments are needed to gain big pictures on the genetic relationships among a large number of diseases. We develop a set of genetic similarity measures to gauge the genetic overlap between diseases, as well as several estimators of the number of overlapping disease genes between diseases. These methods are based on ranked gene lists so that they could be flexibly applied to different types of data. We first investigate the performance of the genetic similarity measure for evaluating the similarity between human diseases in simulation studies. Then we apply the method to diseases in the OMIM database. We show that our proposed genetic measure achieves superior performance in explaining phenotype similarities between diseases compared to simpler methods. Furthermore, we identified common genes underlying the genetic overlap between disease pairs. With an example of five vision-related diseases, we demonstrate how our methods can provide insights into the relationships among diseases based on their shared genetic mechanisms.

A novel similarity score based on gene ranks to reveal genetic relationships among diseases

“…We carry out simulations to evaluate the performance of SimSIP and the existing rank-based methods WeiSumE Ã (Chen, Zhou & Sun, 2015), OrderedList (Yang et al, 2006), FES 0.001 , and FES 0.01 (Serra, Romualdi & Fogolari, 2016), which are weighted sum of the number of overlapping genes between the top part of ranked gene lists, and Euclidean distance EucD between the original values, which is a method without ranking.…”

“…For convenience, genes 1 to d are assumed to be associated with disease 1 and genes d − o + 1 to 2d − o are associated with disease 2, where d ≥ o. As performed by Chen, Zhou & Sun (2015), two normal distributions N(0,σ 2 ) and N(1,σ 2 ) are employed to generate the rank of each gene. Specifically, N(1,σ 2 ) is for the associated genes and N(0,σ 2 ) for the remaining n − d non-associated genes, in either disease 1 or 2.…”

“…For example, Serra, Romualdi & Fogolari (2016) introduced similarity measure FES β (fraction enrichment sum) and set the exponentially decaying parameter β as 0, 0.001, and 0.01. Chen, Zhou & Sun (2015) adopted the default series of β values in the R package OrderedList and proposed a robust statistic minimum p value: min b p b , where parameter β can be calculated by setting a minimum weight and a default series of positions. To avoid arbitrariness and manual intervention of parameter β selection, Chen, Zhou & Sun (2015) finally found a parameter-free similarity measure WeiSumE Ã with good performance to detect overlaps among ranked gene lists in simulations, which is the weighted sum that normalizes the number of overlapping genes on the top genes of two ranked gene lists by its expectation.…”

“…Chen, Zhou & Sun (2015) adopted the default series of β values in the R package OrderedList and proposed a robust statistic minimum p value: min b p b , where parameter β can be calculated by setting a minimum weight and a default series of positions. To avoid arbitrariness and manual intervention of parameter β selection, Chen, Zhou & Sun (2015) finally found a parameter-free similarity measure WeiSumE Ã with good performance to detect overlaps among ranked gene lists in simulations, which is the weighted sum that normalizes the number of overlapping genes on the top genes of two ranked gene lists by its expectation. Generally, existing rank-based similarity measures are mostly based on the overlaps among ranked gene lists and may lead to information loss owing to fixed cutoff positions in the ranked gene lists, the uncertainty of results because of the arbitrariness of the cutoff position, or heavy computational burden.…”

“…By exploring the summation of MAG over all genes, we develop a novel similarity score SimSIP with simpler form and light computation burden to gauge genetic overlap among diseases based on gene ranks instead of intersections between the top part of ranked gene lists. To show the superiority of SimSIP, we firstly conduct a series of simulation studies to demonstrate the performance of SimSIP compared to some existing similarity measures based on ranked gene lists (WeiSumE Ã (Chen, Zhou & Sun, 2015), OrderedList (Yang et al, 2006), FES 0.001 , FES 0.01 (Serra, Romualdi & Fogolari, 2016)) and Euclidean distance EucD under various scenarios. Secondly, we apply SimSIP to analyze the gene expression data of cancers in The Cancer Genome Atlas database and find that it sheds light on the genetic relationships among cancers.…”