Our results based on data from genetic crosses indicate that when similar traits evolve independently in different lineages, the probability that the same genes are used is estimated to be 0.32, on average. The probability estimated from candidate gene studies is 0.55, on average. One explanation for such a high probability of repeated use of the same genes is that at the time of adaptation, effect sizes and availability of beneficial mutations were strongly biased towards a small number of genes. This result can be summarized by the ‘effective’ number of genes, equivalent to the number of loci available if all have effects of equal magnitude and the same probability of fixation. For example, imagine that in two ancestral populations there are n
equivalent genes underlying a trait under selection in which new advantageous mutations might occur and fix. If a major effect mutation occurs and fixes in one gene in one of the populations, the probability is 1/n
that a second population experiencing similar selection fixes a mutation in the same gene. In this case, a probability of gene reuse of 0.32 corresponds to an n
of 1/0.32 = 3.1 effective genes. A probability of gene reuse of 0.55 corresponds to an n
of 1/0.55 = 1.8 effective genes. This rough calculation is simplistic, because real genes do not have equivalent effects. In addition, it does not indicate the cumulative number of genes that might contribute if parallel or convergent evolution is repeated many times. Nevertheless, the high probabilities of gene reuse estimated from published data indicate that the effective number of genes used in parallel and convergent phenotypic adaptation is typically small. If the causes of this low number can be elucidated, then genetic evolution may indeed be somewhat predictable [7
It is difficult to judge how surprising these estimates of effective number of genes are without knowing the total number of genes available in which mutations would cause similar phenotypic changes. Some data are available to assess this. Of the six genes of the Eda-
signalling pathway, mutations in most of which produce a similar phenotype in mammals, only two have been found to be associated with lateral plate variation in threespine stickleback: Eda
and the receptor Edar
]. Similarly, Streisfeld & Rausher [11
] noted that changes to any of the nine enzymes of the anthocyanin biosynthetic pathway would alter pathway flux and produce a change in intensity of flower pigmentation. In accord, 37 spontaneous mutations affecting floral pigment intensity have been detected in five of these nine genes, predominantly in coding regions (another 32 mutations affecting floral pigment intensity occurred in transcription factors). However, in all seven cases in which evolved differences in pigment intensity were mapped, the fixed changes mapped to transcription factors that regulate the pathway genes rather than to the genes themselves, indicating a strong fixation bias away from coding mutations in pathway proteins [11
]. In five of these seven cases, the changes occurred in a gene encoding an R2R3 Myb
transcription factor. Such findings suggest that the number of genes used and reused in adaptive evolution is a small subset of available genes. A host of factors may lead to much higher probabilities of certain genes being involved in phenotypic adaptation than others, including amounts of standing genetic variation, differences in mutation rates or mutation effect sizes, pleiotropic constraints, linkage relationships and epistatic interactions with the genetic background [5
Any explanations for such a high probability of repeated use of the same genes must also explain why this probability declines as more distantly related taxa are compared. First, the high probability of repeated use of the same genes by young, closely related populations might result in part because they have access to the same pool of standing genetic variation [14
], an option not available to more distantly related taxa. Second, as lineages diverge, not only do the specific genes that affect the phenotypic trait diverge in sequence, but the genetic backgrounds with which they interact diverge as well. Hence, the biases that favour use of some genes over others during repeated phenotypic evolution themselves should evolve, in which case we would expect the probability of repeated use of the same genes to decline with time and genetic divergence. The probability that changes to the same genes produce similar phenotypic changes is also likely to be reduced the more widely divergent the lineages [58
], unless gene functions are highly conserved.
Repeated evolution can be divided into two types: parallel evolution, whereby evolution begins from the same starting point, and convergent evolution, whereby evolution begins at different starting points. Arendt & Reznick [17
] argued that from a genetic perspective there is no clear distinction between parallel and convergent evolution. We found that average PS of genes underlying parallel phenotypic evolution was greater than that underlying convergent evolution (). The reasons are probably similar to those described for the effect of node age, since points representing parallel evolution have younger node ages than points representing convergent evolution. If evolution is biased towards some genes over others, populations beginning from the same ancestral genome will more likely share these biases than populations beginning from divergent genomes. However, there is no sudden break in the probability of gene reuse between parallel and convergent evolution (). The distinction is one of degree rather than of kind.
Our estimates based on candidate genes are higher than those based on genetic crosses. Although not statistically significant in our analysis, the difference suggests that the calculated probability of repeated use of the same gene depends on the methods used to detect it. If the difference is real, what are the possible reasons? Whereas genetic cross methods allow us to estimate the contributions of all genes (or at least all genes of moderate to major effect) to repeated phenotypic evolution, the candidate gene approach allows us to determine only whether a specific gene of interest makes a contribution in each case. This essentially lowers the bar for a positive outcome in the case of candidate genes, because the probability of reuse of a gene of interest between two taxa is likely to be higher than the proportional shared use of genes when all mapped genes are considered. Another reason is that the candidate gene method might be more strongly affected by publication bias than estimates based on genetic crosses. We suspect that studies which fail to confirm a role for a candidate gene are more likely to go unreported than results from mapping studies, which produce noteworthy findings if evidence for genes is found anywhere in the genome. On the other hand, estimates of PS that take magnitude of effect into account (i.e. those based on genetic crosses) are prone to higher sampling error, which will tend to cause a downward bias in estimates for the probability of gene reuse.
Caution is warranted when interpreting our results because of numerous judgements and uncertainties inherent to a meta-analysis involving heterogeneous data collected from various organisms, traits and genes. In some cases, we considered overlapping QTL to be reuse of the same ‘gene’, although we may eventually learn that different genes within the QTL underlie repeated phenotypic evolution. While this and other factors already described, such as publication bias, may cause us to overestimate the probability of gene reuse, still other factors may cause an underestimation. For example, our definition of repeated genetic evolution treats paralogous genes as different (several examples were present in our dataset, including the paralogous genetic basis of convergent evolution of caerulein skin toxin in frogs [59
], of digestion of foregut-fermenting bacteria in leaf-eating colobine monkeys and ruminant artiodactyls [58
], and of red flower colour in Mimulus
]). In addition, multiple populations within a single named species were represented by only a single data point in our analysis (the average) to prevent rampant parallel genetic evolution within any one species from unduly affecting the results. Finally, the number of studies on which we have based our analyses is not large, which also adds uncertainty to these results. Despite these uncertainties, our aim here has been to stimulate thinking about these issues and to move towards a quantitative understanding of repeated genetic evolution, which we have attempted with the best available information.
As we accumulate more studies of the genetics underlying repeated phenotypic evolution in natural populations, we will be better able to estimate the probability of the same genes being used. In turn, this will enhance our ability to ask what factors explain variability in genetic parallelism and convergence. For example, broader sampling may allow us to ask whether there is a difference in probability of gene reuse between loss-of-function and gain-of-function traits, or between genes of major and minor effect. Improved knowledge of the biochemical functions and pathway positions of genes will allow us to address whether genes that influence a greater number of other genes in developmental pathways are more or less likely to underlie repeated phenotypic evolution than genes acting at terminal points in the pathways [63
]. Knowledge of mutations will allow us to address how properties such as dominance contribute to the probability they will repeatedly underlie evolution of a phenotype [44
]. Further tests are required of the mechanisms proposed to underlie the high rate of reuse of the same genes, such as pleiotropy and mutation bias [11
]. In the future, it will be interesting to compare our estimates with probabilities of gene reuse from whole-genome sequences of populations adapting to similar environments. We feel that studies starting from purely genetic and genomic approaches must incorporate steps to understand the phenotypic effects of the genetic changes detected. This will be important to determine whether parallel genomic signatures resulted from selection on the same phenotypic traits in different populations, and to determine the mechanisms of selection. Likewise, studies of phenotypic evolution should be followed through to its genetic basis to gain a better understanding of the consequences of repeated phenotypic evolution at the level of genes and mutations. With solid connections between phenotypes and genotypes, repeated phenotypic evolution provides a powerful way to study the predictability of genetic changes underlying adaptive evolution.