We used the distribution of putatively deleterious alleles carried by individuals to infer the relative strength of mutation load across a natural hybrid complex. This approach allowed us to test the prediction that mutation load is reduced in hybrids that are locally adapted to intermediate environments, a pattern that confounds our ability to test this prediction phenotypically.
Proportion of alleles predicted to be deleterious
Approximately 13% of nonsynonymous minor alleles segregating in the interior spruce hybrid complex were flagged as deleterious (using the threshold PROVEAN score recommended by [16]). This estimate is not far from other such estimates in wild populations, including about 20% in Arabidopsis thaliana and rice [17], about 12% in Helianthus annuus [23] and about 28% in Populus trichocarpa [25]. Here, we refrain from comparing numbers or proportions of deleterious alleles per individual (reflecting the extent of mutation load) with those estimated for other taxa, due to differences in methods and cautions explained below. However, inbreeding experiments have long suggested that conifers suffer from relatively high mutation load [32, 34, 35, 51, 52].
We recommend caution when interpreting absolute numbers and absolute proportions of deleterious alleles estimated bioinformatically in this and other studies. Choi et al. [16] suggest that PROVEAN has approximately an 80% specificity (meaning that 20% of all true negative SNPs tested would give false positive results). Surprisingly, this and other studies have found a lower proportion of variants that are flagged as deleterious (a group that should include true positives + false positives) than are expected or even possible given the estimated specificity of PROVEAN (or the other similar programs that have been used) [17, 20, 22, 23, 25]. Finding fewer positive results than predicted by the expected false positive rate is difficult to explain unless the stated specificity of the PROVEAN method is estimated in error.
To accurately estimate specificity, one must test the focal program on a set of known neutral alleles and calculate the number of false positives generated. However, confirming neutrality is very difficult. Studies estimating specificity (including Choi et al. [16]) tend to deem disease-causing variants as truly deleterious and variants not known to cause disease as truly neutral. We predict that many of the variants assumed to be truly neutral are actually weakly deleterious, with effects too small to detect phenotypically or through functional assays, but large enough to be kept at low frequency by selection. If true, this would help to explain why studies like ours often find fewer false positives than predicted by specificity estimates. While this issue calls into question direct interpretation of bioinformatically estimated numbers and proportions of deleterious alleles, such estimates are still useful for relative comparisons across groups within studies (e.g., across the interior spruce hybrid complex) and across studies that use the same PROVEAN score threshold or for which the programs used have similar sensitivity and specificity, estimated in a similar way.
The efficacy of PROVEAN predictions in spruce
PROVEAN and similar programs predict that alleles are deleterious when they occur in conserved amino acid positions and when the amino acid replacement is either a relatively rare or causes a substantial biochemical change at the site [12,13,14,15,16]. However, in some cases such alleles may not be truly deleterious. These alleles may represent genetic innovations that are globally beneficial to the focal species, or they may be beneficial in particular environments and neutral or deleterious in others. In support of the PROVEAN predictions, however, we find strong evidence that alleles predicted to be deleterious are at lower allele frequencies on average than those not predicted to be deleterious, suggesting that predicted deleterious alleles are enriched for truly deleterious alleles. We also find that non-synonymous mutations that were not predicted to be deleterious were at lower allele frequencies than synonymous mutations, most likely reflecting the presence of false negatives in that category. Nonetheless, similar evidence that predicted deleterious alleles are enriched for truly deleterious alleles has recently been found in other systems as well, including Arabidopsis [17], maize [20], barley and soybean [22], sunflower [23] and humans [19, 21, 24], suggesting that that these approaches are useful for identifying deleterious alleles underlying mutation load.
Genotyping error presents an underappreciated complication to studies identifying deleterious alleles bioinformatically. Because we expect deleterious alleles to be at low frequencies, we cannot use a minor allele frequency cutoff when filtering SNPs to help eliminate rare genotyping errors, as is typically done in studies of the genetics of adaptation. Because genotyping errors are not real alleles that are exposed to selection, they may appear to be alleles resulting in more severe biochemical changes and thus be called deleterious more often than real alleles. This represents an alternative explanation for the commonly reported pattern that predicted deleterious alleles are at lower frequencies. However, when we eliminated apparent alleles observed only a single time (the frequency class most likely to contain genotyping errors), we still found strong evidence that predicted deleterious alleles are at lower allele frequencies. This suggests that while rare genotyping errors are almost certainly present in the dataset, they are not driving the pattern, and instead, many real deleterious alleles have been identified.
Ideally, we would like to confirm that predicted deleterious alleles indeed have a negative effect on a phenotypic fitness proxy. Because most deleterious alleles are of small effect and are at low allele frequencies, detecting their individual phenotypic effects requires prohibitively massive sample sizes. Here, we tested for cumulative effects of deleterious alleles (i.e. the proportion of alleles that are deleterious, or the proportion of loci that are homozygous deleterious) on the total biomass of seedling individuals, a proxy for juvenile fitness, and we were not able to detect an effect of either variable beyond the effects of ancestry proportion itself. Because deleterious alleles tend to be strongly differentiated among species, their patterns are tightly correlated with those of ancestry. This has perhaps led to low power for detecting an additional effect of mutation load on our phenotypic fitness proxy. Moreover, seedling biomass is only one small component of total fitness and the correlation between this trait and fitness may not be strong. Finally, our sample of putatively deleterious alleles is likely only a small fraction of those that exist in the genome, and many may be outside of coding regions, which our approach could not target. Zhang et al. carried out a similar test in Populus trichocarpa, and they did detect a significant effect of the proportion of putatively deleterious homozygous alleles on plant height after accounting for distance from the range center and population structure using principal components analysis [25]. Other studies have also found that genes associated with complex traits or genes with known functional effects are enriched for bioinformatically predicted deleterious variants [17, 20, 25].
Relative amount of mutation load in white spruce and Engelmann spruce
Our results suggest that Engelmann spruce carries a greater mutation load than white spruce. Engelmann spruce individuals tend to be burdened by more deleterious alleles (in both heterozygous and homozygous state) due to both higher diversity and higher frequencies of all rare alleles, including deleterious ones. Mitochondrial DNA haplotype diversity is also considerably greater in Engelmann spruce than in white spruce where their ranges overlap (JC Degner, unpublished data). Furthermore, isozyme diversity increases with latitude in Engelmann spruce, and it is highest where its range overlaps with white spruce, a pattern attributed to hybridization [29]. On the other hand, white spruce in this area are at the leading edge of a rapid and long distance range expansion [26, 27]. Therefore, their relatively low diversity may be due to serial founder events that have taken place during range expansion. It may seem counterintuitive that the species with greater genetic diversity has greater mutation load. This pattern could be explained by anything that maintains a larger species-level population size or a higher mutation rate. In particular, if local population sizes are low and there is low dispersal between local populations, then drift is strong relative to selection within local populations, allowing neutral and deleterious alleles to drift locally to high frequencies, while low dispersal maintains a high level of genetic diversity at the species level [53]. Engelmann spruce is adapted to high elevations and their range is currently fragmented on mountain tops [28]. It is plausible (although by no means certain) that during glaciation, refugial population sizes were small with low dispersal between them, contributing to the pattern we observe.
While with increasing ancestry from Engelmann spruce, individuals have more deleterious alleles on average, we also find that the same individuals have proportionately fewer deleterious alleles relative to non-deleterious alleles (both in total and in homozygous state). In other words, patterns of deleterious alleles vary with ancestry proportion when correcting for patterns in non-deleterious alleles. First, this provides evidence that patterns in deleterious alleles are distinguishable from those of demography (as represented by non-deleterious alleles). Second, it may provide evidence that selection is less efficient at removing deleterious variants in more white spruce-like populations, due to weaker purifying selection and/or stronger genetic drift. Given that these white spruce populations are at the leading edge of a recent, long-distance range expansion [26, 27], these results may be a signature of serial founder effects during range expansion (i.e. expansion load) [54]. However, further work is needed to test for expansion load in white spruce. Similarly, [55] found that while African American individuals contained more non-deleterious and deleterious variants than European American individuals, on average, they had a lower proportion of deleterious variants, likely resulting from both increased drift due serial founder effects and a decreased strength of purifying selection in populations that expanded out of Africa [24]. Note though, that several other relevant human studies have been done, with a range of results (e.g. [21, 24, 55, 56]. Together they are forming a detailed picture of the effects of the out-of-Africa expansion on deleterious variation in humans. Also importantly, while there are some similarities between our results and human results, there are also differences. These are likely due in part to significant differences in demographic history between the systems.
Relative amount of mutation load in parental species and their hybrids
Hybrids are intermediate relative to parental species for the proportion of alleles that are deleterious. Thus, if most deleterious alleles have additive effects, then hybrids should have an intermediate mutation load relative to parental species. However, evidence suggests that deleterious alleles tend to be (at least partially) recessive [2, 7,8,9]. Here, we find that hybrids have a lower proportion of loci that are homozygous for deleterious alleles than either parental species. While this pattern is also expected (and was found) for non-deleterious alleles, showing a decrease in the homozygosity of deleterious alleles provides direct evidence for the possibility of complementation in hybrids. Complementation is the mechanism underlying the dominance hypothesis of heterosis [3, 4]. Even if complementation of deleterious alleles does not have a large enough effect to result in heterosis (i.e., hybrid fitness > parent fitness), it may still contribute to higher fitness in hybrids than they would otherwise have and therefore, have a significant impact on the outcomes of hybridization and on the stability of hybrid zones. When hybrids also benefit from local adaptation to their own environment, decreased mutation load due to complementation of deleterious alleles may give them a fitness advantage as well, though its effects on phenotypic fitness proxies would be inseparable from those of local adaptation. Studying mutation load at the genetic level has allowed us to infer that interior spruce hybrids, which are locally adapted to environmental conditions that are intermediate to the parental species (Engelmann and white spruce), also benefit from reduced mutation load due to complementation of deleterious alleles, given that most deleterious alleles are recessive.