A New Likelihood-based Test for Natural Selection

Abstract

We present a new statistic for testing for neutral evolution from allele frequency data summarised as a site frequency spectrum, which we call the relative likelihood neutrality test or ρ. Classical methods of testing for natural selection, such as Tajima's D and its relatives, require the null model to have constant population size over time and therefore can confound demographic change with natural selection. ρ can directly incorporate a null hypothesis reflecting general demographic histories. It has a natural Bayesian interpretation as an approximation to the log-probability of the null model, given the data. We use simulations to show that ρ has greater power than Tajima's D to detect departure from neutrality for a range of scenarios of positive and negative selection. We also show how ρ can be adapted to account for sequencing error. Application to the ACKR1 (FYO) gene in humans supported previous studies inferring positive selection in sub-Saharan populations which were based on inter-population comparisons. However, we did not find the signal of selection to be maximal in the region of the FY*O or Duffy-null allele in these populations. We also applied ρ to investigate in greater detail a region on the 2q11.1 band of the human genome that has previously been identified as showing evidence of selection. This was done for a range of populations: for the European populations we incorporated a demographic history with a bottleneck corresponding to the putative out of Africa event. We were able to localise signals of selection to some specific regions and genes. Overall, we suggest that ρ will be a useful tool for identifying genomic regions that may be subject to natural selection.