Abstract
Synonymous mutations do not alter proteins but undergo natural selection in many species. For Drosophila melanogaster, reports on selection strength vary from undetectable to surprisingly strong. Here we apply a new method to estimate the population selection coefficient (2 Ns) for all 134 ordered pairs of synonymous codon changes. The method uses ratios of site frequency spectra (SFS) for codon changes to neutral changes, and does not depend on divergence data, or codon frequencies, and is relatively insensitive to demographic history. Results indicate that natural selection on synonymous codons is weak, with |2 Ns|<2.07 for all pairs of codons and |2 Ns|<1 for 64% of codon changes. Despite being derived solely from polymorphism data, codon fitness estimates are strongly correlated with observed codon frequencies. A selection-mutation-drift model based on our 2 Ns estimates accurately predicts codon usage, while a model based on mutation alone fails. Codon fitnesses correlate strongly with a measure of codon frequency covariation among genes, and codons with large frequency differences between high- and low-expression genes have high fitness values. Finally, we detect clear signs that selection favors codon changes that stabilize mRNA secondary structure. By avoiding divergence data, which have multiple sources of additional variance, and focusing solely on allele frequencies, a clear picture emerges of selection on codon usage. The convergence of multiple independent lines of evidence (codon frequencies, expression-dependent usage, covariation patterns, and mRNA structure) validates this polymorphism-based approach and provides a coherent framework for understanding selection on synonymous site evolution.