Low concordance of short-term and long-term selection responses in experimental Drosophila populations

Experimental evolution is becoming a popular approach to study genomic selection responses of evolving populations. Computer simulation studies suggested that the accuracy of the signature increases with the duration of the experiment. Since some assumptions of the computer simulations may be violated, it is important to scrutinize the influence of the experimental duration with real data. Here, we use a highly replicated Evolve and Resequence study in Drosophila simulans to compare the selection targets inferred at different time points. At each time point approximately the same number of SNPs deviated from neutral expectations, but only 10 % of the selected haplotype blocks identified from the full data set could be detected in the first 20 generations. Those haplotype blocks that emerged already after 20 generations differ from the others by being strongly selected at the beginning of the experiment and displaying a more parallel selection response. Consistent with previous computer simulations, our results confirm that only Evolve and Resequence experiments with a sufficient number of generations can characterize complex adaptive architectures.

The number of candidate SNPs is inflated as a result of linkage disequilibrium in the experi-92 mental populations (Nuzhdin and Turner (2013); Tobler et al. (2014)). To account for 93 non−independence of candidate SNPs we used a window based approach. We split the main 94 chromosomes (X, 2 and 3) into non-overlapping windows of 5,000 SNPs that are segregating in 95 all generations and replicates. We chose SNPs instead of base pairs as window size measure to 96 allow for variation in SNP density along the genome. To determine if a given window contains 97 more candidate SNPs than expected, we sampled the same number of random SNPs as candi-98 date SNPs in this window (1,000 iterations). "Candidate windows" contained at least as many 99 candidate SNPs as the 99 t h percentile of randomly sampled SNPs. Applying the procedure in-100 dependently to candidate SNPs from all time points provides time point specific candidate win-101 dows ( Figure S1). We evaluated the similarity of two time points with the Jaccard index (for both 102 candidate SNPs, and candidate windows).

103
The number of candidate SNPs in a window is a summary statistic which ignores the signifi-104 cance of the candidate SNPs. If a signal is robust between two time points, we expect the same 105 p-value based ranking of candidate SNPs. Thus, we also evaluated whether candidate SNPs 106 in a given window had a similar relative significance. For each candidate window we created 107 a ROC-like curve (similar to Jakšić and Schlötterer (2016)) by ranking the candidate SNPs by 108 their p-values -the most significant SNP was assigned rank 1 -and calculating the overlap in 109 top-ranked SNPs between two time points. and 96 % of the reconstructed haplotype blocks could be confirmed. This suggests that recon-116 structed haplotype blocks provide a reliable set of linked candidate SNPs.

117
Taking advantage of this additional confirmation of the candidate SNPs in a selected haplotype 118 block we developed a third measure of similarity between time points. We determined the frac-119 tion of candidate SNPs comprising a haplotype block that were also discovered at a given time 120 point (haplotype block discovery rate, HADR) using the poolSeq R-package (Taus et al. (2017)). 121 We note that inference of selected haplotype blocks at each generation does not provide a good 122 alternative to HADR, as the ability to cluster SNPs into haplotype blocks is dependent on the  Early Detected HAplotype blocks (EDHAs) 126 We applied hierarchical clustering (Pollard and Laan (2005) remarkable that a similar number of candidate SNPs was detected at each time point (Table S4).

194
This may imply that earlier time points harbor more false positives, but it is also possible that  More than 27,000 candidate SNPs can be identified at each time point (Table S4) (Table S5-S6) and the inclusion of rare SNPs into the analysis ( Figure S5). We propose that 213 meta-analyses of E&R data should be performed on the level of windows, or probably based on 214 selected haplotype blocks to avoid false negatives due to the high stochasticity of SNP-based Figure 3: Less than 5% of candidate SNPs in generation 60 are detected consistently at every generation. The bars depict the fraction candidate SNPs (purple) and candidate windows (yellow) at generation 60, which are candidates in all subsequent generations (e.g. 40.1% of generation 60 candidate SNPs are candidates in generation 50 and 40). Candidate windows are more consistent than candidate SNPs. Figure  S5 depicts the ratios for candidate sets that are not restricted to SNPs segregating in all generations and time points.

241
We thank the members of the Institut für Populationsgenetik for fruitful discussion and support.