Flexible and efficient count-distribution and mixed-model methods for eQTL mapping with quasar

Identifying genetic variants that affect gene expression, expression quantitative trait loci (eQTLs), is a major focus of modern genomics. Today, various methods exist for eQTL mapping, each using different statistical and methodological approaches. However, it is unclear which approaches lead to better performance, and challenges, particularly scalability as datasets continue to increase in size, remain. Here, we introduce quasar, a flexible and efficient C++ software program for eQTL mapping. Compared to existing eQTL mapping methods, quasar implements a wider variety of statistical models, including the linear model, Poisson and negative binomial generalised linear models, linear mixed model and Poisson and negative binomial generalised linear mixed models. Methodologically, we introduce and implement a simple, analytic approximation to the score test variance in mixed models. Furthermore, we highlight that difficulties with accurately estimating the negative binomial dispersion parameter, previously identified in the context of RNA-seq differential expression analysis, also apply to eQTL mapping. Therefore, quasar implements the Cox-Reid adjusted profile likelihood which enables unbiased estimation of the negative binomial dispersion parameter. We assess quasar’s performance and compare it to three existing eQTL mapping methods: apex, jaxQTL and tensorQTL, on the OneK1K dataset. We demonstrate that quasar’s output agrees with established methods where their models aligns but that quasar is between 5 to over 45 times faster. We exploit the range of models implemented in quasar to compare statistical models for eQTL mapping without confounding by implementation. We find that: count-based models have higher power, that mixed models do not show better performance in a dataset without substantial relatedness, and that the adjusted profile likelihood improves Type 1 error control when using the negative binomial distribution. Additionally, we investigate the relative performance of Poisson and negative binomial mixed models and the use of different approaches for gene-level FDR control. Overall, quasar provides a performant and versatile program for eQTL mapping and we nominate the negative binomial GLM model, incorporating adjusted profile likelihood dispersion estimation, as the statistical model with the best performance.