Disentangling the process of speciation using machine learning

Abstract

Historically, investigations into the processes driving speciation have largely been isolated from systematic investigations into species limits. Recent advances in sequencing technology have led to a rapid increase in the availability of genomic data, and this, in turn, has led to the introduction of many novel methods for species delimitation. However, these methods have been limited to divergence-only scenarios and have not attempted to evaluate complex modes of speciation, such as those that include gene flow during early stages of divergence (sympatric speciation) or population size changes (founder effect speciation). To address this shortcoming, we introduce delimitR, an approach that enables biologists to infer species boundaries and evaluate the demographic processes that may have led to speciation. delimitR uses the binned multidimensional Site Frequency Spectrum and a machine-learning algorithm (Random Forests) to compare speciation models. We use simulations to evaluate the accuracy of delimitR. When comparing models that include lineage divergence and gene flow for three populations, error rates are near zero with recent divergence times (<100,000 generations) and a modest number of Single Nucleotide Polymorphisms (SNPs; 1,500). When applied to a more complex model set (including divergence, gene flow, and population size changes), error rates are moderate (~0.15 with 10,000 SNPs), and misclassifications are generally between highly similar models. We also evaluate the utility of delimitR using three previously published datasets and find results that corroborate previous findings. Our analyses indicate that delimitR can serve as an important conceptual bridge uniting various investigations into the process of speciation.