Widespread selection against deleterious mutations in the Drosophila genome

We have developed a computational approach to simultaneous genome-wide inference of key population genetics parameters: selection strengths, mutation rates rescaled by the effective population size and the fraction of viable genotypes, solely from an alignment of genomic sequences sampled from the same population. Our approach is based on a generalization of the Ewens sampling formula, used to compute steady-state probabilities of allelic counts in a neutrally evolving population, to populations subjected to selective constraints. Patterns of polymorphisms observed in alignments of genomic sequences are used as input to Approximate Bayesian Computation, which employs the generalized Ewens sampling formula to infer the distributions of population genetics parameters. After carrying out extensive validation of our approach on synthetic data, we have applied it to the evolution of the Drosophila melanogaster genome, where an alignment of 197 genomic sequences is available for a single ancestral-range population from Zambia, Africa. We have divided the Drosophila genome into 100-bp windows and assumed that sequences in each window can exist in either low- or high-fitness state. Thus, the steady-state population in our model is subject to a constant influx of deleterious mutations, which shape the observed frequencies of allelic counts in each window. Our approach, which focuses on deleterious mutations and accounts for intra-window linkage and epistasis, provides an alternative description of background selection. We find that most of the Drosophila genome evolves under selective constraints imposed by deleterious mutations. These constraints are not confined to known functional regions of the genome such as coding sequences and may reflect global biological processes such as the necessity to maintain chromatin structure. Furthermore, we find that inference of mutation rates in the presence of selection leads to mutation rate estimates that are several-fold higher than neutral estimates widely used in the literature. Our computational pipeline can be used in any organism for which a sample of genomic sequences from the same population is available.