Eusociality influences the strength of negative selection on insect genomes

Abstract

While much of the focus of sociobiology concerns identifying genomic changes that influence social behaviour, we know little about the consequences of social behaviour on genome evolution. It has been hypothesized that social evolution can influence the strength of negative selection via two mechanisms. First, division of labour can influence the efficiency of negative selection in a caste-specific manner; indirect negative selection on worker traits is theoretically expected to be weaker than direct selection on queen traits. Second, increasing social complexity is expected to lead to relaxed negative selection because of its influence on effective population size. We tested these two hypotheses by estimating the strength of negative selection in honeybees, bumblebees, paper wasps, fire ants and six other insects that span the range of social complexity. We found no consistent evidence that negative selection was significantly stronger on queen-biased genes relative to worker-biased genes. However, we found strong evidence that increased social complexity reduced the efficiency of negative selection. Our study clearly illustrates how changes in behaviour can influence patterns of genome evolution by modulating the strength of natural selection.

1. Introduction

Understanding the processes underlying the evolution and elaboration of sociality is a central goal of sociobiology. Eusocial behaviour evolved multiple times in insects and is characterized by overlapping generations, cooperative brood care and reproductive division of labour [1–4]. It is commonly theorized that eusociality evolved by positive selection on mutations that promote altruistic behaviours [5–9]. It is not surprising then that the majority of molecular population genetic studies of social insects focus only on positive selection [10–16]. While understanding how changes in DNA adaptively drive changes in social behaviour is clearly important, there is a growing interest in understanding how the evolution of social behaviour can influence patterns of genome evolution [17]. Some phenotypes, such as those associated with eusociality, can have substantive consequences on demography and effective population size [17,18], and these changes in effective population size can have profound consequences on the evolutionary landscape in ways that may influence molecular evolution and genome architecture and complexity [19,20]. In other words, the evolution of social behaviour can potentially feed back to influence genome evolution [17,18]. Evaluating this bidirectional link between eusociality and the genome is critical for both disentangling the molecular causes and consequences of eusocial evolution, and for understanding if and how behaviour's influence on the genome could have contributed to the fitness of incipient eusocial lineages.

One plausible way through which sociality can influence patterns of genome evolution is by altering the efficiency of negative selection. The evolution of eusociality has been hypothesized to influence the efficiency of negative selection in a caste- and lineage-specific way. Theory predicts that, all other factors being equal, direct selection on queen traits is more efficient than indirect selection on worker traits [21] because of the stronger association between fitness and the transmission of alleles in the former. This hypothesis directly predicts a relaxation of negative selection on genes with indirect fitness effects (i.e. genes underlying worker traits associated with helping) than genes with direct fitness effects (i.e. genes affecting queen traits). Additionally, eusociality is hypothesized to reduce the effective population size (N_e) [22–29] because of overlapping generations, often extreme sex ratios and reproductive skews (e.g. hundreds to thousands of sterile workers relative to a small number of reproductives), and male-production by workers. A corollary of this concept is that the greater the reproductive skew found within a colony, the greater the reduction in effective population size. The effective population size is a central parameter in population genetics that directly influences the strength of genetic drift, thereby influencing the efficiency of natural selection [30]—a process that can influence genome structure and complexity [19,20]. Genetic drift in populations with small N_e can overwhelm natural selection allowing deleterious mutations to ‘drift’ up to higher frequencies. This hypothesis has been indirectly supported by elevated rates of amino acid evolution in four eusocial species relative to solitary insects [25], although we note that other evolutionary forces can be responsible for this signature, including positive selection on beneficial mutations [31].

We used available population genomic datasets for the advanced eusocial honeybee Apis mellifera [10], the primitively eusocial bumblebee Bombus impatiens [12], the primitively eusocial paper wasp Polistes dominula [11] and the advanced eusocial fire ant Solenopsis invicta [32], in conjunction with existing population transcriptomic data for several other social and solitary insects [25], to investigate the relationship between caste and social complexity on the strength of negative selection. We quantified the strength of negative selection using the distribution of fitness effects approach [25,33–38], which relies on the principal that negative selection skews the allele frequency distribution towards rare variants. This involved analysing within-species polymorphism data, where the strength of negative selection at 0-fold (i.e. sites at which all changes are non-synonymous), 5′ and 3′ untranslated regions (UTRs), intronic, and intergenic sites can be estimated by comparing their allele frequency distributions to fourfold sites (i.e. those at which all changes are synonymous) predicted to be neutral [25,33–38]. We used our population genomic datasets to test the following hypotheses: (i) negative selection on worker traits is relaxed relative to negative selection acting on queen traits, and (ii) negative selection is relaxed as a function of social complexity in insects.

2. Material and methods

(a). Population genomics data

We used published datasets for 49 A. mellifera scutellata workers [10,39], 10 B. impatiens drones [12], 10 P. dominula workers [11] and 40 S. invicta males [32]; these datasets were derived from paired-end Illumina genome sequencing. While the honeybee A. mellifera is sometimes assumed to be domesticated, we note that prior population genetic analyses do not provide any evidence for a ‘domestication bottleneck’ [40,41] and, for the analysis here, we purposefully chose a sample of pure bees from Africa—a population that is rarely used for commercial beekeeping [10]. We used a bioinformatics pipeline to detect single nucleotide polymorphisms (SNPs) using GATK 3.8 (modules used: HaplotypeCaller to obtain gvcf files from bam files, GenotypeGVCFs to merge gvcf files, SelectVariants to select SNPs, VariantFiltration to mask and filter out indels and ambiguous sites) [42]. We used GATK 3.8's filter recommendations to remove sites with low quality (50 < QUAL < 5000; 2 < QD < 40) and depth (SNPs are within 1.5× of interquartile range for depth). Tri-allelic sites were also removed for this analysis. After variant calling, we used SnpEff [43] to annotate SNPs as: 0-fold, 4-fold, 5′ UTR, 3′ UTR, intronic or intergenic. Genes with warning for incomplete transcripts, no start codon or multiple stop codons were removed.

(b). Analyses of negative selection

We used the distribution of fitness effects method in DFE-Alpha to estimate the proportion of sites experiencing negative selection [44]. This analysis only requires within-species polymorphism data, and our analysis was carried out independently for honeybees, bumblebees, paper wasps and fire ants. To reduce the impact of demography on population genetic inference, we used DFE-Alpha's recommendation of conducting a modelling run on the neutral (i.e. fourfold synonymous mutations) site frequency spectrum to estimate demographic and mutation parameters first, before running the entire dataset to estimate negative selection using the demographic and mutation parameters estimated from the first run [44]. This process proceeded independently for each species. We used SnpEff's annotations to determine fourfold synonymous SNPs used as a benchmark for neutrally evolving regions in DFE-Alpha. As recommended, we used the folded allele spectrum for our analysis [44] to estimate the proportion of sites under different bins of negative selection. The folded site frequency spectrum represents the distribution of counts of minor alleles for each dataset calculated over all segregating sites (i.e. the number of sites with 1, 2, 3, 4, etc. minor alleles). For the analysis presented in figure 3, we obtained estimates of the proportion of 0-fold sites experiencing strong negative selection from published data for Halictus scabiosae, Messor barbarus, Reticulitermes grassei, Culex pipiens, Melitaea cinxia and Drosophila melanogaster [25,45]; these estimates were also derived from DFE-Alpha using the same approach as we used for Apis, Bombus, Polistes and Solenopsis.

Figure 3.

Social complexity predicts the strength of negative selection in insects. We found a significant association between social complexity and the proportion of 0-fold sites under strong negative selection in 10 insect species after correcting for phylogeny. (Online version in colour.)

(c). Caste-biased genes

Similar to other studies of molecular evolution in social insects [10–12], we used gene expression datasets to define caste-biased genes (i.e. genes that are upregulated in specific castes). We directly used published lists of differentially expressed genes or proteins in queens versus workers that are significant after correcting for multiple testing (false discovery rate (FDR) < 5%) for B. impatiens [12], P. dominula [11], S. invicta [46] and A. mellifera [47]. We also analysed two A. mellifera RNAseq datasets to identify queen-biased and worker-biased genes, as these two studies did not publish such lists as supplementary data [48,49]. For these two datasets, we generated caste-biased gene lists using DESeq2 [50], and we included any gene in the analysis with greater than 14 reads across all samples and used un-normalized counts following standard practices [50]. Genes were considered differentially expressed if their log2 fold change was greater than 1 and their FDR (Benjamini–Hochberg)-corrected p-value was less than 0.05.

(d). Statistical analyses

Most statistical analyses were performed using Scipy.stats in Python [51]. Following established methods [36], we bootstrapped the results of DFE-Alpha 5000 times to generate 95% confidence intervals to carry out statistical comparisons. When comparing DFE-Alpha estimates of negative selection between groups, we first subtracted the means of the two groups for each bootstrap run, and then estimated the 95% confidence interval for differences in bootstrapped means [52]; differences in means were deemed significant if the 95% confidence interval of the bootstrapped mean differences does not contain zero [52].

For studying the relationship between social complexity and the strength of negative selection, we used previously published information on colony sizes of social species (electronic supplementary material, table S1), and corrected for the effect of phylogeny using the method of phylogenetically independent contrasts [53] as implemented using the function pic in the R [54] package phytools [55]. To do so, we constructed a phylogeny for all 10 taxa using eight single-copy orthologous nuclear genes from OrthoDB [56]: Cdc37 (FBgn0011573), mEFG1 (FBgn0263133), hh (FBgn0004644), exd (FBgn0000611), Hsc70Cb (FBgn0026418), Hr39 (FBgn0261239), CG5504 (FBgn0002174), and wls (FBgn0036141). Amino acid sequences for these genes in the 10 species were aligned using PRANK [57] and the resulting alignments were used to construct a phylogenetic tree based on maximum likelihood as implemented in MEGA [58]; both analyses were carried out using default parameters. We carried out the phylogenetically independent contrasts analysis using both raw proportion data (r² = 0.561; p = 0.02) and arcsine-transformed proportion data (r² = 0.552; p = 0.02); both analyses were consistent and statistically significant.

3. Results

(a). Patterns of negative selection in eusocial genomes

We used previously published population genomic datasets for A. mellifera [10], B. impatiens [12], P. dominula [11] and S. invicta [32] to estimate the strength of negative selection at 0-fold, 5′ UTRs, 3′ UTRs, intronic and intergenic sites; overall, our dataset included mutations in the vast majority of annotated genes in these four species (N = 19 855 for fire ants, 10 662 for honeybees, 14 794 for bumblebees and 11 768 for paper wasps). We used data from fully degenerate fourfold sites as a neutral benchmark. The strength of negative selection was estimated as a categorical variable with the following bins: effectively neutral evolution (N_es = 0–1), weak (N_es = 1–10), moderate (N_es = 10–100) and strong negative selection (N_es > 100). Consistent with Drosophila and human genomes [34,45,59], we found that 0-fold sites experienced the strongest negative selection in all four eusocial species (figure 1; electronic supplementary material, table S2). Polistes dominula (figure 1a) had 72% of their 0-fold sites under strong negative selection (N_es > 100), followed by B. impatiens (62%, figure 1b), A. mellifera (44%, figure 1c) and S. invicta (43%, figure 1d). All species experienced some level of negative selection on 5′ UTRs and 3′ UTRs (figure 1; electronic supplementary material, table S2). Intronic and intergenic sites evolved neutrally in most species expect for S. invicta, where 20–25% of these sites experienced some form of negative selection (figure 1).

Figure 1.

Negative selection in four eusocial insect genomes. The strength of negative selection (N_es) in P. dominula (a), B. impatiens (b), A. mellifera (c), and S. invicta (d) at intergenic, intronic, 5′ and 3′ UTR, and 0-fold sites. N_es bins of 0–1 represent effectively neutral evolution, 1–10 represent weak negative selection, 10–100 represent moderate negative selection and greater than 100 represent strong negative selection. The error bars represent the 95% confidence interval of the means obtained using bootstrapping. (Online version in colour.)

(b). Negative selection on queens versus workers

We tested the hypothesis that genes upregulated in queens (i.e. queen-biased genes) experience stronger negative selection compared to genes that are upregulated in workers (i.e. worker-biased genes) because selection acts on the former directly but on the latter indirectly [21]. We tested this hypothesis using three published expression datasets from honeybees and one each for bumblebees, paper wasps and fire ants. Across all but one of the datasets, we found no evidence of relaxed selection on worker-biased genes (figure 2; electronic supplementary material, table S3). This included comparisons of adult fire ants [46], adult and larval (L2 + L4) honeybees [47,49], adult bumblebees [60] and adult paper wasps [61]. Only a single dataset was consistent with the hypothesis: in honeybee L4 larvae [48], the proportion of 0-fold sites in queen-biased genes with strong negative selection was significantly higher relative to worker-biased genes (figure 2; electronic supplementary material, table S3).

Figure 2.

Caste expression does not lead to a relaxation of negative selection on workers under most circumstances. Estimates of strong negative selection (N_es > 100) at 0-fold sites in queen-biased and worker-biased genes. The error bars represent the 95% confidence interval of the means, obtained using bootstrapping. n.s. and * indicate non-significant, or significant differences in means between queens and workers, respectively. N indicates number of differentially expressed genes for each caste for each study. (Online version in colour.)

(c). Social complexity predicts relaxed constraint in insects

Social complexity was associated with relaxed negative selection in all four species. For example, the species with the greatest caste divergence and the largest colonies (S. invicta) had the smallest proportion of 0-fold sites under strong negative selection (43.5% [42.6–45.9%]), while the species with the least caste divergence and smallest colonies (P. dominula) had the highest proportion of 0-fold sites (72.5% [71.2–73.8%]) experiencing strong negative selection (figure 1). To further explore the hypothesis that social complexity impacts the strength of negative selection, we supplemented our data with published estimates of negative selection on 0-fold sites from three additional eusocial insects [25]: the sweat bee H. scabiosae, the harvester ant Mes. barbarus and the termite R. grassei, and three solitary insects [25,45]: the common house mosquito C. pipiens, the glanville fritillary Mel. cinxia and the fruit fly D. melanogaster. For each species, we developed the following parameter as a proxy for social complexity: log10(1 + no. workers colony⁻¹). Solitary insects thereby have a score of zero and increasingly positive values indicate eusocial species with increasingly larger colonies. We then asked if increasing social complexity leads to a relaxation of strong negative selection on 0-fold sites. After correcting for phylogeny, we found a significant negative correlation between social complexity and the proportion of 0-fold sites experiencing strong negative selection (figure 3, r² = 0.561; p = 0.02).

4. Discussion

We used an allele frequency spectrum approach to quantify the strength of negative selection acting on several insect genomes to understand how eusociality influences patterns of molecular evolution. We first tested the hypothesis that indirect selection on worker phenotypes is less efficient than direct selection on queen phenotypes [21]. Following previous studies [11,12], we used published RNA and protein expression datasets to define sets of genes that are upregulated in workers (worker-biased) relative to queens, and vice versa (i.e. queen-biased). Despite adequate power to detect predicted differences between queen-biased and worker-biased genes (electronic supplementary material, table S4), we did not find strong support for this hypothesis (figure 2; electronic supplementary material, table S3), with only one out of six expression datasets showing a pattern that is consistent with relaxed negative selection on worker-biased genes. We found evidence of weaker indirect selection on worker-biased genes only in L4 larvae in honeybees. Caste differences in the strength of negative selection were not prevalent in primitively eusocial bumblebees and paper wasps (figure 2), which is perhaps not surprising given that: (i) workers in both species contribute to the production of males, (ii) worker Polistes retain the ability to mate and produce daughters, and (iii) relaxed selection on alleles with indirect effects on siblings are less pronounced in monandrous versus polyandrous species [21]. The paper wasp, bumblebee and fire ant studied herein all have an effective mating number of approximately 1 (i.e. monandrous), while honeybees are highly polyandrous [62]. It is important to note that the prediction of relaxed negative selection on worker-biased versus queen-biased genes only holds assuming ‘all else being equal’ [21]. It is possible that this assumption—queen-biased and worker-biased genes are essentially equivalent with the exception of their effect (direct versus indirect, respectively)—rarely occurs in social insects, or that it only transiently occurs during specific developmental stages. As such, it may be operationally difficult to test this hypothesis, especially in eusocial species where workers can produce haploid sons, thereby allowing selection to act directly on their traits under some circumstances.

Next, we tested the hypothesis that eusocial complexity is correlated with patterns of relaxed negative selection [22–24]. The effective population size is a key parameter that influences the strength of genetic drift, which, in turn, influences the ability of natural selection to fix beneficial alleles and remove deleterious ones [30]. In social insects, it is the number of reproductives that influences N_e. For example, consider a colony of 40 000–80 000 honeybees; such a colony requires a tremendous amount of environmental resources but only contains a single queen and a few hundred drones. Assuming a finite carrying capacity, we would expect eusocial species with large colony sizes to have relatively lower effective population size than eusocial species with small colony sizes. Our analysis of fire ants, honeybees, bumblebees and paper wasps was consistent with this hypothesis. Fire ants have the largest colony sizes but experienced the weakest negative selection on 0-fold sites, while paper wasps had the smallest colony sizes but experienced the strongest negative selection on 0-fold sites.

We supplemented our analysis with comparable estimates of negative selection on 0-fold sites from three additional eusocial species and three solitary species. To study how eusociality influences the strength of negative selection, we needed an appropriate metric to capture differences in social behaviour and complexity across all 10 insect species. A recent categorical scale was proposed for comparative analyses of solitary and social species [15] with values ranging from 0 to 3: 0 indicates solitary behaviour, while 1–3 indicate increasing complex eusocial societies. However, this scale does not directly capture the importance of colony size to social complexity [63], which is of primarily interest in our study. We, therefore, decided to quantify social complexity as log(1 + no. workers colony⁻¹). This scale is intuitive and quantitative: solitary species have a score of zero and social species with larger colonies score higher than species with smaller colonies. We found a strong negative association between our measure of social complexity and the proportion of 0-fold sites experiencing strong purifying selection. While correlation does not imply causation, our results are consistent with the hypothesis that sociality reduces N_e [22–29], and the well-established relationship between reductions in N_e and a genome-wide relaxation of selection [19,20,24,64]. Interestingly, there appears to be an association between recombination and social complexity; advanced eusocial insects tend to have higher rates of recombination relative to solitary insects [18,65,66]. The causes of this association have yet to be discovered, but—all other factors being equal—higher recombination is expected to enhance the efficiency of negative selection [67]. Unfortunately, we lacked estimates of cross-over rates for the vast majority of species in figure 3 to jointly estimate the effects of social complexity and recombination rate on negative selection. However, we note that the observed negative association between social complexity and relaxed selection on 0-fold sites (figure 3) may have actually been larger if recombination rates did not positively covary with social complexity.

While there has been much interest in quantifying patterns of molecular evolution that led to the rise of eusociality in insects, we know little about how eusociality can ‘feed back’ to alter patterns of molecular and genome evolution. We uncovered a close relationship between social complexity and relaxed negative selection—a phenomenon that is believed to be key for the evolution of phenotypic plasticity in social insects [68] and other organisms [69]. Going forward, it will be increasingly important to disentangle the genomic changes underlying the evolution of sociality, from those caused by social evolution via its influence on N_e. Moreover, it will be interesting to study how relaxed selection could have contributed to the elaboration of sociality following the evolution of caste.

Data accessibility

We used previously published and deposited genomic and transcriptomic data. Refer to articles cited herein for data accession numbers.

Authors' contributions

M.A.I. and A.Z. designed the research. K.A.D. and B.A.H. provided datasets. M.A.I., K.A.D., B.A.H. and A.Z. performed the analyses. M.A.I. and A.Z. wrote the paper.

Competing interests

The authors declare no competing interests.

Funding

Our study was funded by a Discovery Grant from the Natural Sciences and Engineering Research Council of Canada (NSERC) and a York University Research Chair in Genomics to A.Z.