Dissecting Genomic Determinants of Positive Selection with an Evolution-Guided Regression Model

Abstract

In evolutionary genomics, it is fundamentally important to understand how characteristics of genomic sequences, such as gene expression level, determine the rate of adaptive evolution. While numerous statistical methods, such as the McDonald-Kreitman test, are available to examine the association between genomic features and the rate of adaptation, we currently lack a statistical approach to disentangle the independent effect of a genomic feature from the effects of other correlated genomic features. To address this problem, I present a novel statistical model, the MK regression, which augments the McDonald-Kreitman test with a generalized linear model. Analogous to the classical multiple regression model, the MK regression can analyze multiple genomic features simultaneously to infer the independent effect of a genomic feature, holding constant all other genomic features. Using the MK regression, I identify numerous genomic features driving positive selection in chimpanzees. These features include well-known ones, such as local mutation rate, residue exposure level, tissue specificity, and immune genes, as well as new features not previously reported, such as gene expression level and metabolic genes. In particular, I show that highly expressed genes may have a higher adaptation rate than their weakly expressed counterparts, even though a higher expression level may impose stronger negative selection. Also, I show that metabolic genes may have a higher adaptation rate than their non-metabolic counterparts, possibly due to recent changes in diet in primate evolution. Overall, the MK regression is a powerful approach to elucidate the genomic basis of adaptation.

Strong evidence for the adaptive walk model of gene evolution in Drosophila and Arabidopsis

“…To jointly estimate the effect of the potential confounding factors, we applied a recently developed method that extends the MK test with a generalized linear model [ 57 ]. This approach disentangles the effects of each factor on the rate of adaptive substitutions per nucleotide site.…”

“…However, this method does not model the DFE and hence cannot account for segregating slightly deleterious mutations, which can bias estimates of the rate of adaptive substitutions [ 58 ]. Hence, following the approach suggested in Huang [ 57 ], we removed sites for which the derived allele frequency was below 50% to minimize any potential bias. Despite the large reduction in the data set, this analysis revealed a significant effect of gene age ( S3 Table ; data available in S19 – S22 Data tables).…”

“…We performed the MK regression analysis by comparing polymorphic and divergence sites between nonsynonymous and 4-fold degenerated sites in both species. As this method does not account for the effects of weak selection [ 57 ], we analysed only sites with an allele frequency above 50%. To assess the relative contribution of each variable to the regression model, we compared the likelihood pseudo-R 2 following the Cox and Snell method [ 59 ].…”

Mixture Density Regression reveals frequent recent adaptation in the human genome

“…Higher number of PPIs has been also associated with higher positive selection in the chimpanzee lineage according to the McDonald-Kreitman (MK) test. However, this factor does not seems to be a determinant of positive selection when other genomic factors are simultaneously considered using a MK regression approach [46]. Given the inconsistency with previous evidence and the low strength of the association, the weak negative relationship between selection and the number of PPIs found in this study should be taken with caution.…”

“…However, this factor does not seems to be a determinant of positive selection when other genomic factors are simultaneously considered using a MK regression approach [46].…”