Evolvability of physiological and biochemical traits: evolutionary mechanisms including and beyond single-nucleotide mutation

Despite the vulnerability and limitations of life processes, organisms exploit seemingly inhospitable environments and perform remarkable feats. The biochemical and physiological mechanisms that underlie these extraordinary capacities have been a major focus of experimental biologists, increasingly using the toolkits of molecular biology and genome-enabled research. Indeed,such research is increasingly revealing the inner workings of these mechanisms, as many of the other contributions to this review volume demonstrate. But how have these mechanisms themselves come into being?

Many biologists, if pressed, invoke `adaptation by natural selection' as an answer to this question. For some biologists, the answer stops there. How else could the extraordinary biochemical–physiological traits that so obviously enable organisms to function in diverse environments and are so clearly `tuned' to each organism's environment have arisen except by natural selection (Fig. 1)?Consolidating this assessment are numerous examples of evolutionary convergence and parallelism, wherein functionally equivalent or identical mechanisms have arisen multiple times in response to similar selective pressures.

Other biologists will go further, and emphasize consistency with the fundamental principles of natural selection: (a) individuals in a population vary in genotype; (b) individuals in a population consequently vary in phenotype (including biochemical and physiological traits); (c) individuals in a population consequently vary in fitness; i.e. reproductive output. Hence genotypes that encode higher fitness will come to predominate in a population. Documentation of points (a) and (b) is now abundant both within and among populations, species and higher taxa, and documentation of point (c) in natural populations is growing (Endler,1986; Feder et al.,2000; Feder and Mitchell-Olds,2003; Van Straalen and Roelofs, 2006). Four examples of diverse vintage are considered below.

(1) In human populations (Ingram,1963; Williams et al.,2005), mutation of nucleotide 17 from A to G in the sequence encoding the beta-subunits of the Hemoglobin A protein converts a glutamine to a valine. As a result, hemoglobins form fibrils via interactions due to the valine. Consequently, the shape of the hemoglobin-containing erythrocyte changes. Individuals homozygous for the mutation develop sickle-cell disease, and have lower fitness. Individuals heterozygous for the mutation, by contrast, have enhanced resistance to malaria(Plasmodium infection) by some as-yet-unknown mechanism, which might include an increased immune response, filtering of infected erythrocytes by the spleen, and/or outright inhospitability to the Plasmodium. This is a classic example of a balanced polymorphism, wherein the advantages to the heterozygote offset the disadvantages to the homozygote mutant.

(2) The fish Fundulus heteroclitus inhabits brackish water habitats along the eastern coast of North America, and so experiences enormous variation in temperature. A key component of the adaptation to this thermal gradient has been evolution of the ldh-b gene, encoding one component of the enzyme lactate dehydrogenase (LDH). Specifically, northern populations express more of this gene's product than southern populations, which may compensate for the cooler average temperatures of the northern populations. Evolved differences between alleles are both in protein function (i.e. in coding sequence) and in gene expression (i.e. in the proximal promoter region). Indeed, as initially reported(Segal et al., 1996), the north–south variation in proximal promoter sequence is sufficient for corresponding differences in gene expression.

(3) Beach mice (Peromyscus) inhabit beaches with diverse sand coloration, and mismatches of fur and sand color enhance predation on mice(Hoekstra et al., 2006). Thus mouse populations on dark substrates have repeatedly evolved melanic coloration. In one case, a single-nucleotide variant (changing arginine 65 to cysteine) of the melanocortin-1-receptor gene, Mc1r, is strongly associated with melanism; this gene's product is key to the production of dark-pigmented fur. In other cases, no such genetic variation is evident,suggesting that melanism has arisen through at least two different genetic routes.

(4) Brewer's yeast (Saccharomyces cerevisiae) makes ethanol via the alcohol dehydrogenase enzyme ADH1, which it then feeds into the tricarboxylic acid (TCA) cycle via the enzyme ADH2. Thomson et al. (Thomson et al., 2005)have theorized that ancestral yeast had a single ADH that, much like vertebrate LDH and lactate, recycled NADH for use in glycolysis by producing,in this case, ethanol; ethanol could readily be lost to the environment before it poisoned the yeast. When fleshy fruits arose and Saccharomycescolonized them, yeast then evolved enhanced biosynthetic capacity for ethanol(and tolerance for it), which enabled yeast to triumph over microbial competitors in the fruit. If so ADH2, which is specialized for ethanol catabolism, would only have been needed post hoc. Using maximium likelihood techniques (Thornton,2004), the authors inferred the nucleotide substitutions that must have transformed the common ancestor of the ADH1- and ADH2-encoding genes into their present state, synthesized the corresponding enzyme, and confirmed that it was catalytically specialized for ethanol synthesis. Importantly, this work signifies the possibility of reconstructing the sequence of single-nucleotide mutations that might transform one protein to another (see also Thornton, 2004).

These are only four of many examples of such work(Hochachka and Somero, 2002; Watt and Dean, 2000), but are noteworthy for several reasons. First, each exemplifies an unambiguous linkage among gene, trait, whole-organism function or performance, and fitness. Second, each involves alternative single nucleotide polymorphisms (SNPs),alleles or haplotypes segregating in natural populations. This variation could plausibly have arisen through the processes of single-nucleotide mutation:replication error, or damage to a base due to chemical agents or ionizing radiation followed by imperfect repair. Plausibly, multiple rounds of such random mutagenesis could have produced even the more complex haplotypes that are segregating in the latter three examples. Indeed, SNPs are common, account for a substantial fraction of phenotypic variation within populations, and are the basis for the coming era of pharmacogenomics and personalized medicine. No less a personage than Ernst Mayr (Mayr and Provine, 1998), p. 20 stated that `...gene mutations are the only raw material of evolution' [as cited by Ryan(Ryan, 2006)], with which many biologists concur. But is single-nucleotide mutagenesis the only process generating genetic variation that is consequential for fitness in natural populations? And is this process sufficient to account for the remarkably complex adaptations of living things?

Although Charles Darwin understood neither heredity nor mutation in detail,he clearly anticipated that the origin of complex adaptations through the accumulation of small changes would challenge his theory(Darwin, 1859). Also well in advance of the genomic era, Richard Goldschmidt proposed that, because mutations of small effect would be insufficient to yield complex adaptations,rare mutations of large effect were needed(Goldschmidt, 1960). [As Ernst Mayr (Mayr, 1963) again pointed out, this proposition was itself implausible – and where were these mutations of large effect and how did they arise?] Even today, the`irreducibility of complexity' (Behe,1996) is a major talking point in the creationism and `Intelligent Design' political campaigns against the science of evolution – alongside the absence of fossils documenting `missing links'. But if `genes are living fossils' (E. Gaucher, personal communication), the molecular paleontology of modern genomics and its successor sciences have now revealed numerous mechanisms that can create massive variation in genomes upon which evolution can act. First, however, what are the evolutionary issues that these new discoveries may resolve?

According to Falconer (Falconer,1981), the change in the mean value of a trait from generation to generation is the product of two variables: the intensity of natural selection and the genetic variance of the trait. Intensity of selection is a variable readily comprehensible to comparative physiologists, and needing little substantiation. Nonetheless, Falconer's equation states that no matter how intense selection may be, evolutionary change cannot ensue unless adequate genetic variation is present; all else being equal, evolvability is proportional to genetic variability. Evolvability, however, is seemingly both an independent and a dependent variable. Ongoing selection can purge a population of genetic variability if the intensity of purifying selection exceeds the ability of mutation and other variational processes to generate it– unless, that is, something shields genetic variability from selection. How, then, can genetic variability either be generated de novo or shielded from the purgative of selection? And how can genetic variability be generated such that at least some of it is neither so minor as to impede evolution nor so extreme as to be deleterious?

A mechanical engineer could, in principle, transform an automobile into a powered boat by adding a hull, rudder and propeller, and by making appropriate adjustments in the transmission and steering mechanisms. Transformation of the same engine into a machine for powered flight is likewise feasible (cf. the Wright brothers, 1902), or of a propeller-driven aircraft into a jet aircraft or rocket. A computer engineer could upgrade the performance of a computer by installing a new operating system, input/output devices, and firmware. In such cases, human engineers have luxuries unavailable to evolving organisms. First,the engineers can readily power down the machine they are transforming, make changes and restart it; living things can perform the counterpart operation only with great difficulty (but see embryogenesis and metamorphosis). Second,engineers can obtain an extraordinary variety of alternative parts, or invent novel parts if necessary; if relying on simple mutation, evolving organisms can proceed only by modifying the genes already at hand(Long et al., 2003). Third, an engineer can conceive that, for example, to transform an automobile into a powered boat, requires simultaneous changes in the exoskeleton, propulsive system, steering system, etc., and implement these changes simultaneously. If relying on simple mutation, evolving organisms must either await numerous successive mutations, none of which enhances fitness without the others, or an incredibly improbable bout of massive simultaneous mutation. Weinreich et al.(Weinreich et al., 2006)addressed a relevant situation experimentally: bacteria in which five mutations confer maximal resistance against an antibiotic. These five mutations could occur in any of 5! (120) possible orders, of which only 18 successively increase antibiotic resistance. The remainder are prohibited because at least one intermediate stage decreases antibiotic resistance. Finally, an engineer can discard failed attempts and begin again, whereas organisms in which mutation decreases fitness substantially are conceivably doomed evolutionarily. How, then, does a flightless organism evolve flight, a terrestrial organism an aquatic existence, or any less-adapted organism the extraordinary features that comparative physiologists have discovered?

The flowering of interactomics, network biology, and systems biology has revealed how single-nucleotide substitutions (and/or much larger mutations;see below) may in principle yield transformations of large effect, and thereby create adaptation. Although network biology is still in its infancy, several points are clear. Like any system, biological systems can be characterized as networks of interacting components, be they genes and transcription factors,messengers and signalling pathways, or species in a biological community(Barabâasi, 2003). Not all components in a biological network are equal; some are relatively highly connected (so-called `hubs') and some less so, with the number of connections obeying a power-law distribution. This `scale-free' pattern differs from random networks, in which the number of connections is a Poisson distribution. Not all links among components are of equal strength(Csermely, 2006). Finally, the structure of the network is consequential for its function and its resistance to perturbation (i.e. robustness). While a more extensive general discussion of networks is impossible here (but see, for example, Barabaasi and Csermely),several features are critical to evolvability.

Because all components in a network are connected, perturbing one component can perturb many – especially if that component is a hub. Examples are numerous: single-gene knock-outs can be lethal, as can pharmacological inhibition of a single enzyme or signaling molecule, and removal of a single`keystone species' can devastate a natural community. More benignly, a small number of transcription factor or signaling molecules can create a cascade of impacts downstream that effect growth, development, and regulation. Such organization may be inevitable for complex systems(Csete and Doyle, 2004): the necessity for integrating numerous complex systems requires `common currencies' such as ATP and the major second messengers.

Networks are heterogeneous in the strength and number of connections. Some sub-organizations occur multiple times and are highly connected internally,but are somewhat interchangeable with their counterparts. These are modules,biological equivalents of the individual telephones in a telephone network,individual circuit boards in a computer or parts in a machine, or individual computers in a network. In both development and intracellular signaling, for example, specific modules repeatedly appear in diverse pathways(Carroll, 2005; Pereira-Leal et al., 2006; Pereira-Leal and Teichmann,2005).

Networks are more like roller coasters (in which removal of an average single strut has little impact on function) than like houses of cards (in which removal of an average card causes the entire structure to collapse)(Feder, 1996). An exhaustive account of how, at every level of biological organization, change in or removal of components does not affect the performance of the whole has recently been published (Wagner,2005). Alternative components whose substitution has no impact are said to occupy the same `functional space', and these functional spaces are remarkably extensive. For example, although some single-gene knockouts can be harmful or lethal (see above), the vast majority of genes can be deleted with no apparent ill effect (Feder and Walser,2005).

Combining these features yields a recipe for evolvability. Because of redundancy and robustness, mutation can tinker with numerous individual components without jeopardizing the whole. Because of modularity, a mutated module can be substituted for an original module(Carroll, 2005; Pereira-Leal et al., 2006; Pereira-Leal and Teichmann,2005). Because of hierarchical organization, some small changes can trigger major changes downstream. Thus, while one or more single-nucleotide mutations in the average gene may seldom suffice to generate major adaptive novelty, such mutations in key genes in the right module at the right time may have enormous impacts. Recently there have been several reviews or similar theses advanced (Kirschner and Gerhart, 2005; Carroll,2005; Carroll et al.,2005; Davidson and Erwin,2006).

The most relevant formalization of this thinking is the duplication–degeneration/complementation model of Force, Lynch and colleagues (Fig. 2), originally applied to duplicate genes (Force et al.,1999) but recently extended to networks in general(Force et al., 2005). It reasons that when a component is duplicated [and recall that many components and modules are already redundant], only one duplicate is necessary to maintain the pre-existing function. When the second component undergoes change, it may either degenerate (as selection need not maintain it in view of the duplicate) or take on a differing function (like the gene encoding ADH2,discussed above). The differing function can either be an entirely new one(neofunctionalization, as when change results in a gene's promoter acquiring a new response element), or the two duplicates can each take on part of the function of the original component (subfunctionalization, as exemplified by the genes encoding ADH1 and ADH2). Once refunctionalized, the former duplicate may then be preserved by selection. Again because of hierarchical organization, refunctionalization of a single component can trigger major changes downstream. Perhaps the most notorious example of this final point concerns the homeobox or Hox genes, which have undergone duplication/refunctionalization in development and now specify whether a body segment grows a limb and/or wing and/or haltere and/or other structure,depending on which particular Hox paralog is active in that segment(Carroll et al., 2005). The Hox paralogs have acquired the ability to respond appropriately for their position in the developing embryo via refunctionalization of their promoters to bind transcription factors that vary in concentration from anterior to posterior. Additional spectacular examples involve the origin of armor plating in stickleback fish (e.g. Colosimo et al., 2005) and coloration in insects (e.g. Gompel et al.,2005).

Interestingly, it has been suggested that the typical network structure of living things is itself a facilitator of evolvability, at least in theory(Oikonomou and Cluzel, 2006). These authors modeled the outputs of networks with equal numbers of components but with different topologies, random versus scale-free (i.e. resembling many biological networks), using Boolean threshold dynamics to compare the output with a target function. Emulating natural selection, they then randomly mutated the components of the two networks, and selected for or against mutants according to their resemblance to the target function. In this simulation, the scale-free networks evolved more rapidly and regularly toward the target function than did the random networks.

Duplication–degeneration/refunctionalization takes time and, as noted above, natural selection is fully able to purge temporarily unsuccessful intermediates from a population before conditions favor their preservation and/or subsequent evolution. Evolutionary capacitors are mechanisms that shield variation from the purgative of selection. The term was proposed by the editors of the journal Nature as the title of the first report that a molecular chaperone, Hsp90, might act in this way(Rutherford and Lindquist,1998).

Molecular chaperones are proteins (or other compounds) that interact specifically with other `client' proteins, normally in the context of aggregation deterrence, protein folding and/or degradation; many chaperones are heat-shock or stress proteins. Hsp90 is a peculiar chaperone in that it is normally present and at relatively high concentrations in eukaryotic cells,where it interacts with numerous client proteins. It is thus not specific for a particular client. Hsp90 is also a typical chaperone in that it will bind proteins not in their native (i.e. fully and normally folded) state and assist in their proper folding. Rutherford and Lindquist hypothesized that, because of these properties, Hsp90 could bind mutant gene products that if unchaperoned would yield dramatic phenotypes – and the ensuing loss of their encoding genes due to selection. Hsp90 would thus deter the elimination of these mutations. Upon environmental stress, however, the profusion of stress-damaged proteins in the cell would titrate Hsp90 away from these mutant gene products, allowing their phenotypes to emerge. Hypothetically, if these emergent novel phenotypes were beneficial, they could then be preserved by selection (Fig. 3). Although no beneficial mutation has yet been discovered to have arisen by this mechanism in natural or experimental evolution (see Mayr's criticism of Goldschmidt,above), Lindquist and colleagues have now assembled numerous data consistent with most components of this hypothesis(Cowen and Lindquist, 2005; Queitsch et al., 2002; True et al., 2004), and their work has stimulated broad interest in evolutionary capacitors.

In hindsight, many features of living things are evolutionary capacitors or render genetic variation cryptic (Gibson and Dworkin, 2004; Masel,2006). These include ploidy >1, recombination, segregation and sex. Like chaperones, these can protect gene products or combinations thereof from selection and suddenly reveal them. Likewise, chromatin configuration(Kornberg and Lorch, 1999; Segal et al., 2006),epigenetic modification (e.g. gene imprinting, methylation), alternative splicing and phenotypic plasticity in general, can all suppress/reveal genetic variation. Alternative splicing is especially spectacular in this regard, as a single gene can encode >38 000 splice variants(Schmucker et al., 2000). Indeed, evidence is mounting that natural selection can manipulate splicing to express variants appropriate for a given environment(Marden, 2006). Finally, the intronic nature of eukaryotic genes provides an additional site for genetic variation. Mattick and colleagues have hypothesized that introns have enabled the evolution of encoding of RNAs that themselves do not encode proteins, and that these non-coding RNAs are in turn responsible for the complexity of eukaryotes (Mattick,2003).

Now that techniques are available that can assess and compare most genes in multiple genomes, we are recognizing that largescale (i.e. involving many nucleotides at once, if not entire genomes) mutations, in combination with network features and evolutionary capacitance, may play a correspondingly major role in the genesis of biological variation, especially the variations that comparative physiologists call `adaptations'. Unlike single-nucleotide mutation, which unarguably remains an important component of the evolutionary toolkit, these readily yield plausible scenarios of large-scale evolutionary change and origin of new features (Long et al., 2003); termed `genomic creativity'(Ryan, 2006).

The standard genetic machinery of DNA replication, recombination and segregation is not perfect. Not only does this yield single-nucleotide mutations, but also duplications of larger fractions of the genome if not entire genomes. There are several excellent introductions to the topic(Hurles, 2004; Zhang, 2003), which emphasize that gene duplications are about as frequent as single-nucleotide mutations. As a result 30–65% of the genes in eukaryotic genomes are duplicate[table 1 in Zhang (Zhang,2003)]. Lynch and Conery(Lynch and Conery, 2000)estimated that one in a hundred genes is duplicated and fixed every million years. Thus, gene duplication is not a rare occurrence.

Duplications of nucleotide sequences arise primarily by two distinct mechanisms (Hurles, 2004; Zhang, 2003). In the first(tandem or segmental duplication), unequal recombination occurs between homologous sequences at two places in the genome (within the same chromatid,between sister chromatids, between different chromosomes). In the second(retroposition), a transcribed mRNA sequence is reverse-transcribed, and the resultant DNA inserted into a chromosome. Thus retroposition (a) can be into any accessible chromatin in the genome and is not limited to homologous sequence, (b) only duplicates a transcribed sequence, which contains no promoter sequence and need not contain introns, (c) duplicates genes or operons one at a time; in contrast, tandem duplications may be of sequences containing many genes and intergenic regions. Genome-wide studies reveal the extraordinary impact of these processes. Our own species is remarkable for numerous large segmental duplications(Bailey and Eichler, 2006), as is evident in whole-genome displays [see figs 2 and S4 of Bailey et al.(Bailey et al., 2002)]. The Drosophila genome contains hundreds of genes lacking introns but otherwise close matches of intron-containing genes on other chromosomes(Betran et al., 2002; Emerson et al., 2004; Long et al., 2003), evidently the result of retroposition.

Errors in segregation during meiosis can also result in extra copies of entire chromosomes or, the utmost in gene duplication, of the entire genome. Indeed, whole genome duplication has clearly occurred in yeast and plants, and two rounds of duplication occurred in the chordate lineage(Dehal and Boore, 2005).

Hybridization also has long been recognized as a potential source of genomic change. By definition, it involves the combination of two distinct genomes, typically due to the failure of reproductive isolative mechanisms. Although perhaps not an everyday occurrence, novel species have clearly arisen in nature as a result (reviewed by Ryan,2006).

Much more frequently, species acquire new genes by lateral or horizontal gene transfer from another species, sometimes involving a biological vector and sometimes simply by uptake of DNA(Bushman, 2002). The likely magnitude of this transfer is such that it has been concluded that no real`tree of life' exists because all `branches' have interchanged genes and hence are genealogically indistinct (Doolittle,1999). Again, the combination of genomic information and bioinformatic tools establishes how extensive this interchange is [see fig. 3 in Kunin et al. (Kunin et al.,2005)].

Finally, few species of eukaryotes are, strictly speaking, single species,but actually are multispecies consortia of symbiotes and their genomes(Margulis and Sagan, 2002). Whatever one's reaction to the more controversial aspects of Margulis's thesis, interactions of symbiont genomes and host genomes have clearly brought about the evolution of what physiologists classically regard as adaptations,such as the bacteriocyte of aphids (Moran and Degnan, 2006), the light organ of squids(Koropatnick et al., 2004),and the rhizosphere of plant roots (Marx,2004).

Some DNA sequences encode the capacity to move within or among genomes. A first class of mobile element moves by encoding a mRNA that is reverse-transcribed and inserts elsewhere in the genome; these are aptly named retrotransposons. A second class either copies itself or excises from its original site and inserts elsewhere, but remains DNA throughout. Those unfamiliar with transposable elements tend vastly to underestimate their impact on evolutionary change (Kazazian,2004); some (V. Walbot, personal communication; M. Evgen'ev,personal communication) attribute perhaps 40–90% of all evolutionary change to transposable elements. The simplest impact of these elements is to interrupt the sequence of their current (and often former) site in the genome;i.e. insertional mutagenesis. Transposable elements sometimes capture portions of DNA sequence adjacent to their original insertion site and, when they excise, carry it with them. They may thus convey novel regulatory and/or coding sequence to a new insertion site, and thereby endow the new host gene with novel properties. Transposable elements now constitute large fractions of all eukaryotic genomes sequenced so far (e.g. approximately 50% in humans).

Once genomes have acquired new genes, whether by duplication or from another species, numerous processes in addition to single-nucleotide mutation may modify them, sometimes far more dramatically than can single-gene mutation. As noted, retroposition, segmental duplication and transposable elements can each transfer sequence within the genome, creating combinations of sequences that were previously non-existent. Alternative splicing can likewise create novel combinations of gene products. Additionally, mechanisms such as exon shuffling, gene fission/fusion can create new genes(Long et al., 2003). In some rare cases, novel genes can even arise from nonsense or intergenic sequence(Long et al., 2003).

The foregoing mechanisms, although often individually distinct, can interact to yield evolutionary novelty. For example(Zhang et al., 2004), the Drosophila gene jingwei arose when a first gene, yellow emperor, duplicated to yield a second gene, yande, and a third gene, alcohol dehydrogenase, retroposed into the third intron of yande. Thereafter, at least 30 non-silent single-nucleotide mutations yielded the present-day gene. Unlike its ADH ancestor, which is specialized to detoxify and assimilate ethanol, the substrate specificity of jingweiis for long-chain primary alcohols, which are important in hormone and pheromone metabolism.

In another example (Walser et al.,2006), the P element, a transposable element of Drosophila, needs physical access to decondensed DNA to insert itself. In many genes, the chromatin is normally condensed and in nucleosomes,except when the gene is actually being expressed (i.e. chromatin is often an evolutionary capacitor). Perhaps due to their role as an emergency response to stress, the proximal promoters of the genes encoding heat-shock proteins are constitutively decondensed and nucleosome-free. This has made the heat-shock promoters distinctively susceptible to the insertion of P elements,which selection may then maintain or eliminate to manipulate the expression of their host genes.

Finally, because the sequence of transposable elements is often highly conserved, their insertion throughout the genome creates opportunities for homologous recombination among distant sites, as noted above. Such recombination, in turn, may invert, duplicate or otherwise scramble genes,multi-genic nucleotide sequences, or segments of chromosomes. Observing that one class of transposable element, a form of Alu element, is often present at the boundaries of segmental duplications in the human genome,Bailey and Eichler (Bailey et al.,2003; Liu et al.,2003) have hypothesized that the extensive segmental duplication in the human genome is due to the proliferation of one class of Aluelements in the human lineage (but not in close relatives), which then allowed ectopic recombination. Thus, to the extent that our species owes its capacities to gene duplication, these elements may have underlain the evolution of humankind.

The advent of molecular, genomic and bioinformatic techniques and their increasing applicability to diverse species has enormously enhanced experimental biologists' ability to understand `how animals work'(Schmidt-Nielsen, 1972). Adaptational biology will be incomplete, however, until the understanding of how adaptations came into being is equally advanced. This understanding may well come about through sustained interaction with modern evolutionary biologists, evolutionary genomicists and evolutionary systems biologists. One clear outcome of such interaction is that single-nucleotide mutation, often the mainstay of adaptational biologists' evolutionary thinking, will become viewed as only one of several mechanisms in evolution's toolkit. The other mechanisms may be far more powerful than single-nucleotide mutation in facilitating evolvability and, although they have not done so yet, be able to explain in detail the origin of the complex traits that fascinate adaptational biologists.

The author thanks the participants of the Journal of Experimental Biology symposium on `Post-Genomic Comparative Physiology' and two anonymous reviewers for insightful comments. This work was supported by NSF Grant 03-16627.