Come on feel the noise – from metaphors to null models

The metaphor of ‘speciation islands’ (Turner et al., 2005) has dominated speciation research in the last decade. It invokes a particular (and plausible) feedback between divergent selection, migration and recombination (Barton & Bengtsson, 1986) and has led to a general re-evaluation of the role of gene flow in speciation. This, together with the fall in sequencing costs, has spawned an industry of studies that scan the genomes of closely related species for outliers of divergence (usually measured in relative terms as FST). Undoubtedly, outlier scans have contributed to the discovery of spectacular examples of reproductive barrier loci in particular in Heliconius butterflies and cichlid fish (e.g. Nadeau et al., 2012; Malinsky et al., 2015). Yet, given the flood of genomic studies, one may wonder why we have not learned more about the speciation process and the genes and genetic architectures involved in the buildup of reproductive isolation. Despite its appeal as a metaphor, the idea of speciation islands has proven frustratingly difficult to relate to sequence data in a concrete way. Studies invoking ‘speciation islands’ as an explanation for outliers of divergence abound (see Wolf & Ellegren, 2017 for a review), and a great deal of effort has been devoted to ‘follow up’ on such outliers with genetic mapping or experimental studies. However, there have been few attempts to relate patterns of sequence diversity and divergence to the underlying population-level processes in a quantitative way. Ravinet et al. (2017) give a careful review of the demographic and selective processes involved in the build-up of reproductive isolation and the complex ways they interact with each other to shape diversity and divergence along the genome. One of their main conclusions is that a meaningful interpretation of the genomic landscape of speciation must account for both the background demography and the heterogeneity in basic genome properties, such as gene density and the rate of recombination and mutation. Ravinet et al. (2017) also stress that divergence and diversity are highly stochastic and that incomplete lineage sorting (ILS) ‘[. . .] increases the variance of genomic divergence estimates making it difficult to identify true outliers and also potentially introducing false positives’. Their conclusion, however, is oddly ambiguous: ‘[. . .] incorporating demographic history in tests for selection is difficult as incorrect specification of the history, potentially generated by ILS patterns increases error rates’. They go on to argue that ‘[. . .] approaches that do not use demographic models may be preferable in some cases, although these too are prone to bias’. I agree about the difficulty of the task, but – contrary to Ravinet et al. (2017) – I would argue that model-based inference is the only hope for understanding the genome signatures of speciation. It is of course true that we are a long way from being able to fit a full, mechanistic model of the speciation process to genomic data. As Ravinet et al. (2017) make clear, even an extremely simplified cartoon of speciation at the genomic level must necessarily be complex and include demography, heterogeneity in background selection and recombination, specifics about the number and distribution of barrier loci and about how andwhen selection has acted on them. Perhaps even more worryingly, we currently have no general understanding of how much information about past demography and selection is actually contained in genome data. In other words, even if we had a perfect method for extracting all the relevant signal, it is not clear how much detail we would be able to infer about a particular speciation history from genomic data alone. Clearly, the information contained in sequence variation is finite, whereas the space of potential speciation scenarios is not. The fact that even very simple demographic histories for a single population have recently been shown to be nonidentifiable from the site frequency spectrum (Terhorst & Song, 2015; Lapierre et al., 2017) is a pertinent reminder of the inherent limits of the information in sequence data. Thus, Ravinet et al. (2017) are right to emphasize the value of incorporating independent information in the form of recombination and background selection maps. The idea that all the demographic and selective processes that shape a particular genomic landscape of speciation could be captured in a single model is daunting at best and infeasible at worst. However, a much more realistic and very worthwhile starting point for speciation genomics would be to ask how well a particular genomic landscape can be explained by simple null models. Thanks to algorithmic improvements in coalescent simulations (Kelleher et al., 2016), we now have – for the first time – the ability to efficiently generate genomic landscapes of divergence under the full ancestral process of coalescence and recombination (and even condition such simulations on a recombination map) and for any demographic scenario. To give a concrete example, consider the simplest possible null model of speciation: a strictly allopatric split without any selection or heterogeneity in recombination. Coalescent simulations under this model give an immediate feel for the noise inherent in divergence and diversity Correspondence: Konrad Lohse, Institute of Evolutionary Biology, University of Edinburgh, King’s Buildings, Edinburgh EH9 3FL, UK. Tel.: +44(0)1316507335; fax: +44(0)1316506564; e-mail: konrad.lohse@ed.ac.uk