Distinguishing Positive Selection From Neutral Evolution: Boosting the Performance of Summary Statistics

Abstract

Summary statistics are widely used in population genetics, but they suffer from the drawback that no simple sufficient summary statistic exists, which captures all information required to distinguish different evolutionary hypotheses. Here, we apply boosting, a recent statistical method that combines simple classification rules to maximize their joint predictive performance. We show that our implementation of boosting has a high power to detect selective sweeps. Demographic events, such as bottlenecks, do not result in a large excess of false positives. A comparison to other neutrality tests shows that our boosting implementation performs well compared to other neutrality tests. Furthermore, we evaluated the relative contribution of different summary statistics to the identification of selection and found that for recent sweeps integrated haplotype homozygosity is very informative whereas older sweeps are better detected by Tajima's p. Overall, Watterson's u was found to contribute the most information for distinguishing between bottlenecks and selection.

Learning Natural Selection from the Site Frequency Spectrum

“…We note that our feature set bears some conceptual similarity to that of Lin et al (2011). With the exception of the iHH features, Lin et al (2011) effectively learn from a small set of weighted linear combinations of the scaled SFS.…”

“…In fact, our analysis shows that the small difference in observed power is explained by the iHH features ( Figure S4). Furthermore, the fundamental nature of our feature set enables our framework to potentially learn linear combinations that are not captured by the set used in Lin et al (2011). In addition, we emphasize that although SFselect-s and Lin et al (2011) show similar power in the parameter-specific case, in practice the parameters of a selective sweep are seldom known.…”

“…Although there have been recent applications of machine learning to SFS-and LD-based summary statistics for inferring selection (Pavlidis et al 2010;Lin et al 2011), to the best of our knowledge, this study represents the first attempt to apply supervised learning directly to the scaled SFS to this end. In addition, whereas most supervised learning approaches are inherently specific to parameters of the training data, here we propose a way to overcome this by leveraging common attributes in the learned models of selection.…”

A Fast Estimate for the Population Recombination Rate Based on Regression

“…This explains why we use boosting in connection with regression here. In Lin et al (2011) on the other hand, Y has been taken as a binary variable with 0 indicating neutral evolution and 1 positive selection. This led to a problem of classification.…”

“…A promising method for estimating the best possible model is boosting (Bühlmann and Hothorn 2007), a machine learning method, which can be used both for classification and regression. In another population genetic setup, Lin et al (2011) implemented boosting successfully to distinguish positive selection and neutral evolution, by solving a binary classification problem.…”