Constraints and plasticity in genome and molecular-phenome evolution

Abstract

Multiple constraints variously affect different parts of the genomes of diverse life forms. The selective pressures that shape the evolution of viral, archaeal, bacterial and eukaryotic genomes differ markedly, even among relatively closely related animal and bacterial lineages; by contrast, constraints affecting protein evolution seem to be more universal. The constraints that shape the evolution of genomes and phenomes are complemented by the plasticity and robustness of genome architecture, expression and regulation. Taken together, these findings are starting to reveal complex networks of evolutionary processes that must be integrated to attain a new synthesis of evolutionary biology.

Evolution is variously constrained on all levels of biological organization, from genome sequence to genome architecture, gene expression, molecular interactions and organismal phenotypes^1,2. Constraints on the rates and paths of evolution can be divided into genomic constraints, which are manifested at the level of the genome sequence and architecture, and phenomic constraints, which affect phenotypic characteristics. The vast amounts of diverse data recently generated by comparative genomics and systems biology provide previously inconceivable insights into the patterns and processes of genome and phenome evolution^2-5.

Comparative genomics has the potential to measure the strength of constraints on different classes of sites in genomes and to elucidate the biological nature of these constraints. Genome comparisons also help to address higher-level questions, including the degree to which constraints act on gene repertoires, genome architecture and the evolution rate itself. Similarly, the avalanche of systems biology data allows researchers to ask new, qualitative questions, such as how do constraints affect metabolic fluxes and the ‘molecular phenome’ (which includes gene-expression, regulatory and interaction networks)?

Here, we attempt a genome-wide and organism-wide assessment of the constraints that operate at different levels of biological organization and discuss the links between constraints and the robustness and plasticity of biological systems. In the interests of space, we omit important areas that are covered in recent reviews — for example, constraints on development⁶ or metabolic energy expenditure⁷ — and only touch upon issues such as the evolution of regulatory networks⁸.

We begin by discussing the constraints that affect sequence evolution and proceed to examine constraints on gene and genome architectures, genome size, gene number and gene repertoires. We then switch to considering constraints on molecular phenotypes, discuss universal models of protein evolution and conclude with a synopsis of the main factors that seem to constrain evolution at different levels.

We demonstrate the diversity of constraints on different classes of sites in genomes and contrast the constrained evolution of functionally important sites, particularly in proteins, with the plasticity of genome architecture and expression. Despite the variance of constraints in and across genomes, the emergence of these patterns might be explained by universal evolutionary patterns and simple stochastic models.

Constraints on sequence evolution

Constraints across a genome sequence

Constraints differ greatly across classes of genomic sequences and across taxa. Constraints on a particular class of sites can be measured only by comparison to another class of sites that are assumed to have evolved neutrally. The choice of an appropriate neutral model is a major problem in molecular evolution studies (BOX 1).

Such comparisons have revealed that, typically, the constraints on sites encoding protein sequences and those of structural RNAs (such as ribosomal RNAs and tRNAs) are stronger than those on non-coding sites, although the distributions of constraint strengths are broad and overlapping^9,10 (BOX 1). Most protein- coding genes evolve under purifying selection of widely varying strength, and few evolve under positive selection (BOX 1). Genes evolving under continuous, long-term positive selection encode specialized proteins for which rapid change is crucial for function, typically involving an ‘arms race’ between competing agencies, such as hosts and parasites; examples include bacterial surface proteins^11,12 and proteins involved in mammalian spermatogenesis, sperm competition and sperm-egg interaction^13,14. Evolution under positive selection is not unconstrained — constraints on the overall protein structure still apply¹⁵ — but evolution along the available trajectories proceeds rapidly.

The fact that most genes encoding proteins and structural RNAs evolve under purifying selection does not imply uniform constraints across sites. on the contrary, the evolutionary rates of codons (sites) in protein-coding genes — and by implication the strength of constraints on different sites — is well-described by a characteristic distribution in which a small fraction of sites are virtually unconstrained or subject to positive selection, and most sites are subject to constraints that vary within a broad range^16,17.

The assumption that synonymous sites in protein-coding genes evolve neutrally is useful for measuring selection at the protein level but is a rough approximation at best. The universal, significant positive correlation between Ka and Ks^18-21 (BOX 1) is compatible with the view that the evolution of synonymous sites is also constrained and that the forces shaping the evolution of non-synonymous and synonymous sites are related (see ‘Constraints on protein-coding genes’ below).

Constraints across taxa

The distributions of constraints across genomes are dramatically different in life forms with distinct genome architectures. In particular, there is a pronounced difference in constraint distribution between viruses and prokaryotes (bacteria and archaea), which have ‘wall-to-wall’ genomes that consist mostly of protein-coding and RNA-coding genes, and multicellular eukaryotes, in whose genomes the coding nucleotides are a minority^22,23 (FIG. 1). The constraints on compact genomes, particularly those of prokaryotes, are much stronger than the constraints on the genomes of multicellular eukaryotes (median Ka/Ks values for prokaryotes and multicellular organisms are typically 0.01-0.1 and 0.1-0.5, respectively). In viruses and prokaryotes, nearly all genomic sites are evolutionarily constrained, with the notable exception of pseudogenes, which are common in some parasitic bacteria, such as Rickettsia species or Mycobacterium leprae^24-26. Non- coding regions constitute only 10-15% of the genomes of most free-living prokaryotes, and a considerable fraction of these sequences encompasses regulatory elements that are substantially constrained in their evolution²⁷. The genomes of most viruses are even more compact, with almost all of the genome sequence taken up by protein-coding genes²³. The genome architecture of most unicellular eukaryotes resembles that of prokaryotes, although the fraction of non-coding sequences, of which a large portion is expected to evolve without constraints, is greater in these genomes than it is in prokaryotes.

Figure 1. Approximate distribution of evolutionary constraints across genomes with different architectures.

The fractions of different classes of sequences that are subject to constraints of varying strength are shown as rough approximations of the values that are typical of the respective class of genomes. The data are from REFS 2,23,29,58, as discussed in the main text.

By contrast, multicellular eukaryotes (plants and especially animals) have intron-rich genomes with long intergenic regions, and a large fraction of these non- coding sequences indeed seem to undergo unconstrained evolution (FIG. 1). The estimated fractions of constrained nucleotides in a genome differ substantially even between animals: in Drosophila melanogaster, ~70% of sites in the genome, including ~65% of the non-coding sites, seem to be subject to selection²⁸, whereas in mammals this fraction is estimated at only 5-6%²⁹ or even ~3%³⁰. Note, however, that the absolute numbers of sites subject to selection in these animal genomes of widely different size are quite close. By contrast, the fraction of constrained non-coding sites in Arabidopsis thaliana is estimated to be much lower than in D. melanogaster^31,32, although the two genomes are comparable in size and architecture. of course, the estimates of the fraction of constrained sites (that is, sites that are ‘visible’ to selection; see discussion below and BOX 1) are based on simple population-genetic models, and the assumptions of these models might be violated to different extents across evolving lineages, leading to illusory differences. Nevertheless, different evolutionary regimes would seem to be operating even among closely related species.

The selectome, the RNome and ‘junk’ DNA

The estimate of 3-6% for the fraction of constrained sites in mammalian genomes is remarkable from two opposite standpoints. First, it seems that most of the mammalian genome fits with the much-maligned definition of ‘junk’³³. of course, the functional recruitment of ‘junk’ sequences, such as mobile elements, is common³⁴ but at any given time most of the mammalian genome evolves without appreciable constraints. Second, as protein-coding sequences comprise only ~1.2% of the genome²⁹, the substantial majority of the sites under selection do not encode amino acids. In particular, the selective pressure on 5' and 3' uTRs of mammalian genomes is comparable to or even stronger than that on synonymous sites in coding regions^35,36. An even greater contribution to the mammalian ‘selectome’ (the total number of sites under selection³⁷) is the rapidly growing complement of non-coding RNA genes, the RNome³⁸. The RNome includes numerous regulatory microRNAs that are subject to a broad range of constraints^39,40 in addition to many long non-coding RNAs that might function in gene regulation and development and seem to be subject to constraints that are comparable to those on protein-coding genes⁴¹.

The known part of the RNome is the tip of the iceberg, especially considering that transcripts have been detected from nearly all sequences in mammalian genomes^42,43. Comparative-genomic analysis reveals numerous conserved sequences (including ‘ultraconserved elements’⁴⁴) in introns and intergenic regions of animal and plant genomes^45,46, but transcription into a specific functional RNA has been shown only for a few of these^47,48. Most of the transcription in organisms with complex genomes, such as mammals, is probably functionally irrelevant, but functional non-coding RNAs might still outnumber protein-coding genes. The extent of sequence conservation that is unrelated to specific functions of transcripts but rather caused by requirements of expression regulation, chromatin structure and other factors remains an open question.

The current understanding of the constraints on different types of sites across known genomes (FIG. 1) can be summarized as follows: sequences encoding structural RNAs and non-synonymous sites in protein-coding sequences are among the most strongly constrained; synonymous sites, sequences coding for regulatory RNAs, and non-coding regulatory sequences on average evolve under weaker selection but are not free of constraints; and characteristic distributions of constraints crucially depend on genome size and architecture (primarily gene density). Evolutionary regimes may differ widely even in closely related taxa, such as arthropods and vertebrates.

Constraints on gene and genome architectures

Constraints on gene architecture

An aspect of gene architecture that is common to all life forms but is particularly prominent in eukaryotes is the multidomain organization of proteins⁴⁹. Numerous proteins consist of multiple ‘evolutionary domains’, and the multidomain organization of some key proteins is conserved throughout the evolution of cellular life. Generally, however, domain rearrangements form an important resource of plasticity in all evolving lineages, particularly in the case of ‘promiscuous domains’^50,51.

A eukaryote-specific feature is the exon-intron organization of protein-coding genes. Intron positions are highly conserved over long evolutionary distances: up to 25-30% of intron positions are located in the exact same positions of orthologous genes between animals and plants^52,53. Some animal lineages, in particular vertebrates, evolve under almost complete intron stasis, with minimal intron loss and virtually no gain. By contrast, the evolution of other lineages, such as nematodes and many groups of unicellular eukaryotes, features extensive turnover of introns^54,55.

On the whole, the evolution of eukaryotic gene architecture is very diverse, with a highly dynamic evolutionary process in some lineages but much less change in others.

Constraints on genome architecture in prokaryotes

Genome architecture refers to the mapping of genetic elements onto the genome, including gene order, and the clustering and co-regulation of genes with related functions^2,23,54,56. Long-range gene order is surprisingly poorly conserved among sequenced archaeal and bacterial genomes, which is in contrast to the recurrence of operons in diverse prokaryotes⁵⁷. long-range gene orders in prokaryotes diverge roughly in proportion to sequence divergence in protein-coding genes; however, evolution of gene order can be so fast that, in many lineages, no long-range conservation is seen even when sequence divergence is very low⁵⁸. Beyond this general pattern, the rates of gene-order decay differ substantially among prokaryotic lineages⁵⁸. Gene order in many prokaryotes seems to be disrupted largely by inversions centred at the origin of replication, a process that does not seem to be strongly constrained by purifying selection and depends primarily on the activity of the relevant recombination machinery^59,58. The activity of mobile elements, in particular insertion sequences, is another important — and also weakly constrained — factor affecting genome rearrangement in prokaryotes⁶⁰.

In contrast to the lack of conservation of long-range gene order, prokaryotic operons combine evolutionary resilience and plasticity, forming overlapping gene arrays that are partially shared by evolutionarily distant organisms^61,62. Although some operons, notably several parts of the ‘superoperon’ that encodes ribosomal proteins, show a pronounced signal of long-term vertical evolution⁶³, the wide distribution of many operons among prokaryotes is attributable to horizontal gene transfer (HGT) under the selfish operon concept⁶⁴.

Despite the lack of conservation of the long-range gene order, some constraints affect the gross architecture of prokaryotic genomes: in particular, the preferential codirectionality of gene transcription with replication might arise from selection to minimize the frequency of collisions between RNA polymerase and replication forks⁶⁵.

Constraints on genome architecture in eukaryotes

Most eukaryotes have no operons, and existing operons are unrelated to prokaryotic operons and seem to have evolved de novo⁶⁶. In eukaryotes, the search for gene-order non-randomness — for example, the clustering of genes with connected functions or with similar expression levels and patterns — has led to mixed results^23,66,67. Evidence exists not only for advantages of clustering of functionally linked genes thanks to enhanced possibilities for co-regulation⁶⁷ but also for disadvantages of such clustering, conceivably due to transcriptional interference⁶⁸. With some striking exceptions, such as the strict order of animal Hox genes⁶⁹ or the clustering of enzymes in certain metabolic pathways in yeast⁷⁰, gene order in eukaryotes seems to be quasi-random²³. Generally, the evolution of gene order in eukaryotes seems to be dominated by random chromosomal rearrangements⁷¹, and no gene arrays are highly conserved between distantly related forms, such as different animal phyla, let alone between animals and fungi or plants²³.

Thus, the evolution of genome architecture seems to be shaped by the interplay of strong constraints that determine the conservation of operons, weak constraints on other forms of functional clustering and large-scale gene organization, and extensive dynamics of genome rearrangements, duplication and HGT. These dynamics both counteract weak constraints by disrupting gene associations and reinforce the effect of stronger constraints, as in the case of the horizontal spread of selfish operons.

Constraints on gene number and gene repertoires

The number of protein-coding genes in cellular life forms varies within a surprisingly narrow range compared with the genome size, especially considering the difference in biological complexity between prokaryotes and multicellular eukaryotes. Excluding the extremely reduced genomes of some intracellular parasitic bacteria and the huge, polyploid plant genomes, the number of encoded proteins varies from ~500 to ~35,000, less than two orders of magnitude²³. In unicellular organisms, especially in prokaryotes, the number of encoded proteins closely correlates with genome size (gene density is roughly constant, at around one gene per kilobase of DNA), whereas in multicellular organisms, especially animals, the two quantities are decoupled.

Constraints on the lower gene bound

What constrains the number of encoded proteins from below and above? The lower threshold intuitively corresponds to a minimal gene set for cellular life^72,73. Minimal gene sets derived from comparative-genomic and experimental approaches converge at 250-350 genes — or more precisely, orthologous gene lineages⁷⁴ — and seem to encode most essential cellular functions^72,73. An apparent paradox is that the conserved set of 250-350 orthologous genes can be derived only by comparing small sets of not-too-diverse genomes, as pioneered in a comparison of Haemophilus influenzae and Mycoplasma genitalium⁷⁵. By contrast, the core set of ubiquitously conserved genes is continuously shrinking with the addition of new sequenced genomes and seems to be limited to approximately 30 genes, all encoding proteins involved in translation and transcription^57,76. The solution to the paradox is non-orthologous gene displacement (NoGD): most of the essential cellular functions can be performed by members of more than one orthologous gene set, which in many cases are completely unrelated^72,77. The minimal genetic complement of a cell is not a unique minimal gene set but rather a unique set of indispensable functional niches that can be filled with diverse collections of genes.

Constraints on the upper gene bound and gene repertoire

The maximum number of genes in a cellular life form does not substantially increase despite the rapid growth of the collection of sequenced genomes. Thus, although an upper bound of genetic complexity seems to exist, the nature of this limit remains obscure. one attractive hypothesis is the ‘bureaucratic ceiling of complexity’ (BOX 2).

This hypothesis seems to be particularly plausible given the lack of large expansions in gene number in vertebrates, in which the number of genes and the genome size are decoupled. In these organisms, the principal constraint on the upper limit is probably imposed by the cost of regulation and expression rather than the cost of replication. It is not surprising then that vertebrates evolved elaborate means of increasing proteomic complexity — such as alternative splicing, alternative transcription⁷⁸ and regulatory RNAs — that do not involve inflating the number of protein-coding genes.

Gene duplication and constraints on evolution of paralogous gene families

The formation of paralogous gene families through gene duplication is the main route of innovation, especially in eukaryotes^79,80. The size distribution of paralogous families in each genome follows a power-law-like function⁸¹ that is well reproduced by a simple gene birth and death model conditioned on the equilibrium (constant size) of genomes during evolution^82,83. This process seems to define a fundamental constraint on gene demography that is coupled to the constraint on the upper bound of the total number of genes.

Gene loss is a major factor of evolution, in agreement with the findings on the small and shrinking cores of conserved genes, NoGD and extensive redundancy. Gene loss is extensive in all lineages, in particular in the evolution of animal taxa: there is a high degree of orthology between vertebrates and primitive animals, such as sea anemones and Trichoplax adhaerens, which is in contrast to the much more limited orthologous relationships between vertebrates and arthropods or nematodes^84,85. Individual genes show a broad distribution of gene-loss rates^86,87; moreover, it seems that the observed evolutionary and phenomic features of genes are compatible with a steady-state model of genome evolution under which the distribution of gene-loss and -gain rates remain effectively constant over extended evolutionary spans⁸⁸. This distribution could reflect another important constraint governing genome evolution.

Thus, the genetic complexity of genomes in all life forms (as reflected by the number of genes) is tightly constrained from the above (that is, constrained with respect to the maximum number of genes), whereas the genome size may be either coupled to the gene number and hence similarly constrained (as in prokaryotes and most unicellular eukaryotes) or decoupled and hence much less strictly constrained (as in multicellular eukaryotes, especially animals). The constraints on gene number and the functional distribution of genes remain to be explored in detail but are likely to be determined by fundamental ‘laws’ of cell functioning. The gene repertoire in all organisms depends on the process of evolution by gene duplication and loss that is well-described by birth and death models.

Constraints on protein-coding genes

Protein-coding genes are generally highly constrained, but the distribution of the rates of evolution among non-synonymous sites in orthologous genes in any pair of compared genomes spans three to four orders of magnitude and is much broader than the rate distribution for synonymous sites (BOX 1). Remarkably, the shapes of the rate distributions for orthologous proteins are highly similar for all studied cellular life forms, from bacteria to archaea to mammals⁸⁸ (FIG. 2). Another universal of genomic and phenomic evolution is the anticorrelation between the rate of evolution of a protein-coding gene and its expression level: highly expressed genes evolve slowly, a dependence that is observed in all model organisms for which expression data are available, although the magnitude of the negative correlation widely differs^86,89-91.

Figure 2. The universal distribution of evolutionary rates across orthologous gene sets.

The evolutionary rates for five pairs of closely related organisms from different branches of life were calculated as nucleotide distances for the complete sets of orthologous genes⁸⁸. The shapes of the rate distributions are very similar from bacteria to humans. The relative evolution rate for each gene was obtained by dividing its evolution rate by the median rate for the respective pair of organisms. ‘Model’ refers to estimated transition rates in 134 mutationally connected networks for simulated, robustly folding 18-mer protein-like molecules⁹⁶. Original model rates were normalized by their median value and scaled to a standard deviation of 0.25 to match the width of the distributions derived from biological data.

The existence of these universals suggests that the primary determinants of protein evolution have more to do with fundamental principles of protein structure and folding that are common across all life than with unique biological functions. It has been proposed that the principal selective factor underlying evolution of proteins is robustness to misfolding, as misfolded proteins can be toxic to the cell^21,90. This model is compatible with the preferential use of optimal codons (strong codon bias) in highly expressed and highly conserved protein-coding genes^92-94 and with the aforementioned positive correlation between Ka and Ks. under the misfolding hypothesis, the evolution of synonymous sites is constrained, at least in part, by the same factors as the evolution of proteins owing to the pressure for the preferential use of optimal codons in highly expressed proteins and in specific sites that are important for protein folding^90,95.

A recent analysis of protein evolution produced estimates of evolutionary rates under the assumption that misfolding is the only source of fitness cost⁹⁶. The results reproduce the universal distribution of protein evolutionary rates as well as the dependence between evolutionary rate and expression (FIG. 2), and suggest that the universal rate distribution indeed might be a consequence of fundamental physics of proteins (BOX 3).

Thus, the primary constraints on protein evolution might have more to do with the maintenance of the native folding and intermolecular interactions than with unique protein functions, a view that seems to be supported by a recent large-scale analysis of protein- family evolution¹⁵.

Constraints on molecular phenotypes

Owing to advances in systems biology, it is possible to assess evolutionary variance and constraints by comparing various features of the molecular phenotype — such as gene expression, protein abundance and architecture of interaction networks — among different organisms.

Molecular phenomic variables show a distinct structure of dependences among themselves and with evolutionary variables, such as the rates of sequence evolution and gene loss⁹⁷. The correlations between phenomic variables are typically positive (for example, highly expressed proteins also tend to interact with many other proteins and have many paralogues), whereas the correlations between phenomic and evolutionary variables are generally negative (for example, highly expressed genes on average evolve more slowly than those expressed at a low level). Most of these correlations are statistically significant but relatively weak, so caution is required (experimental biases should be investigated as potential causes^98-100), but the overall pattern of positive and negative correlations seems to be undeniable^97,101. Thus, constraints on the ranges of phenomic variables partly seem to constrain the evolution of gene sequences, gene repertoires and genome architectures, as shown by the model of protein evolution discussed above.

Constraints on gene expression and protein abundance

Comparative expression analysis of orthologous genes has suggested that gene expression in animals is not strongly constrained during evolution or at least has a major neutral component^102,103. However, subsequent analyses revealed signatures of selective constraints that affect gene expression^104-106,107. The expression level of orthologous proteins is strongly correlated even among distantly related animals: a correlation coefficient greater than 0.8 was observed for approximately 3,000 orthologous genes from Caenorhabditis elegans and D. melanogaster¹⁰⁸. Similar results have been recently reported for a broad range of model eukaryotes¹⁰⁹, which is in sharp contrast with the correlation coefficients of 0.2-0.4 that are seen in comparisons of other genomic and molecular phenomic variables⁹⁷.

Notably, protein abundance seems to be constrained to a substantially greater extent than gene expression, and the degree of constraint is even stronger than the rate of DNA-sequence evolution within the same set of orthologous genes^108,110.

Constraints on network architecture

Global architectures of protein interaction and gene co-expression networks seem to be universal across life forms, and the network node degree (the number of connections) has a characteristic power-law-like distribution¹¹¹. local network structures are much less constrained and differ even among closely related organisms^112,113. However, a comparison of gene co-expression networks from mutation accumulation lineages of C. elegans, in which selection is effectively removed¹⁰⁴, with those from the natural isolate suggests that the local wiring of the co-expression network is constrained by selection, whereas the global properties are not¹¹⁴. Thus, the similarity of global-network properties in widely different organisms might reflect ‘neutral’ rather than selective constraints — that is, these properties could have evolved through simple, stochastic, non-selective processes, as in birth-and-death models of genome and network evolution^83,115. This view is reinforced by the fact that certain evolving but non-biological networks, such as the internet, possess similar global properties¹¹¹.

The study of molecular-phenome evolution is still in its infancy, but the advances of evolutionary systems biology have already revealed unexpectedly strong constraints on some phenomic features.

Constraints on evolutionary trajectories

What would happen if the tape of evolution were rewound?

An intriguing question is the degree to which the course of evolution itself is constrained¹¹⁶. To what extent is the evolutionary process free to explore different trajectories between the given initial and end states? Direct studies of evolutionary trajectories in sequence space are very limited but have already shown that, although historical contingency is central to evolution¹¹⁷, exploration of the sequence space is strongly constrained so that only a minority of theoretically possible trajectories are accessible.

In theory, mutational trajectories in sequence space are fundamentally stochastic¹¹⁸. However, experimental evolution studies indicate that paths of adaptive evolution are constrained by interactions between mutations (epistasis and pleiotropy), although not to the point of becoming deterministic. Experimental evolution of bacterial antibiotic resistance resulting from 5 point mutations in the β-lactamase (bla1) gene showed that, of the 120 possible trajectories across the sequence space, 102 were inaccessible, and of the remaining 18, several had negligible probability of realization¹¹⁹. Even stronger constraints were identified in a subsequent study that explored a more complex fitness landscape by simultaneously evolving resistance to two antibiotics¹²⁰.

Systematic studies of bacterial evolution under controlled conditions reveal both parallel emergence of the same mutations under a particular selective pressure and the realization of multiple trajectories^116,121,122. For instance, the evolution of the same, rare phenotype — the ability to grow on citrate — proceeded along distinct trajectories in different Escherichia coli populations¹²³.

These results complement and reinforce previous observations of the convergent and parallel evolution of proteins that perform the same function. Such evolution is limited but demonstrable, and was first observed in the evolution of lysozymes in ungulates and langur monkeys^124-127.

The extent of constraints on an evolutionary trajectory and, conversely, the likelihood of parallel evolution crucially depend on the shape of the fitness landscape: the more rugged the landscape, the stronger the constraints. The shape of the landscape itself is determined by the nature, strength and interactions of the relevant selective factors; furthermore, the landscape evolves with time, making it into a fitness seascape¹²⁸.

Ne as the general gauge of evolutionary constraints

Under population genetics theory, the effectiveness of purifying selection is proportional to the effective population size (Ne) of a given organism, assuming a uniform mutation rate. only mutations for which |s| > 1/Ne, (in which s is the selection coefficient) can be fixed or efficiently eliminated during evolution². Conversely, mutations with |s| < 1/Ne are ‘invisible’ to selection. This simple dependence could be the primary determinant of the constraints that affect different aspects of genome and phenome evolution. Differences in Ne seem to underlie the qualitative difference in the genome architectures of unicellular and multicellular organisms^22,129. Substantial genome expansion seems to be attainable only in organisms with small populations and the attendant weak selection, as is the case in plants and animals. In these organisms, the deleterious effect of propagating nonfunctional sequences is often too small to allow the ‘detection’ and elimination of such sequences by purifying selection. Accordingly, evolutionary conservation does not automatically imply that the conserved feature is constrained by purifying selection owing to its functional importance but rather, somewhat paradoxically, might reflect weak purifying selection that is insufficient to eliminate non-adaptive ancestral features.

Gene architecture: case example

The evolution of exon- intron gene structure in eukaryotes is an excellent demonstration of this population-genetic paradigm. Most introns are functionless and weakly deleterious for an organism partly due to their energetic cost. However, a simple estimate based on typical mutation rates shows that the deleterious effect of introns is ‘visible’ to purifying selection only in populations with Ne of 10⁷ or greater. This is within the typical Ne range of unicellular eukaryotes, whereas multicellular eukaryotes have smaller populations^2,22,130. These differences have a dramatic effect on the evolution of genome architecture in eukaryotes. unlike the genomes of unicellular forms, which typically contain fewer than one intron per gene, plants and animals possess numerous introns⁵³. Furthermore, conservation of intron positions in orthologous genes of animals and plants (see above) seems to be due to the inefficient elimination of introns in organisms with small Ne and not to constraints on intron positions per se. Finally, introns in unicellular eukaryotes are short and have conserved, optimized splice signals^131,132. By contrast, introns in intron-rich genomes are often long and bounded by relatively weak splice signals¹³³, thus providing for the evolution of alternative splicing.

Its importance notwithstanding, Ne determines the course of evolution only on a broad scale. A comparative analysis of the Ka/Ks values among prokaryotic lineages did not detect a negative correlation between selective constraints and genome size⁵⁸, as implied by the straightforward population-genetic perspective¹³⁴. on the contrary, larger genomes tend to evolve under stronger constraints (even when only free-living microbes are analysed), implying that lifestyle could be a crucial determinant of genome evolution (favouring, in particular, gene acquisition through HGT in variable environments) independently of Ne⁵⁸. Thus, evolution driven solely by population-genetic factors could be an appropriate null hypothesis, but the actual evolutionary trajectories are determined — and constrained — by specific biological contexts (FIG. 3).

Figure 3. Genomic and phenomic constraints on different levels of biological organization.

The degree of constraint and plasticity experienced by genomic and phenomic properties across different levels of biological organization are shown. The arrows should be perceived as indicating the midpoints of the respective intervals, with at least the adjacent intervals overlapping. The relationships between some of the constraints on some classes of sites — for example, synonymous sites and disordered segments in proteins — are ‘educated guesses’ based on current data, and could change with further accumulation of the relevant data and advances in methods for quantifying selective pressures. The scales are rough approximations.

Robustness and plasticity of biological systems

The aspects of evolution that are complementary to constraints are the plasticity of genomic and phenomic characteristics, and the robustness of molecular phenotypes¹³⁵. As mentioned above, plasticity is pronounced at many levels. In many organisms, large-scale genome organization seems to be only weakly constrained, so gene order substantially differs even among closely related organisms, especially among prokaryotes^23,58. The gene repertoires of many organisms, especially prokaryotes, show plasticity that may even exceed the plasticity of genome architecture, as shown by rapid genome reduction in parasitic bacteria²⁶ and by acquisition of pathogenicity islands that may comprise over 30% of the genome in bacterial pathogens¹³⁶. The plasticity of genome organization and composition is paralleled by the evolutionary flexibility of regulatory networks, and complements the more strongly constrained evolution of individual genes^137,138.

Evolutionary plasticity and the strength of evolutionary constraints are tightly linked to the robustness of biological systems. Robustness seems to be an evolved property, as shown by the study of specialized buffering mechanisms, the impairment of which reveals hidden genetic variation and capacitates evolution^139,140. Recently, the concept of evolutionary capacitation has been extended to numerous genes with extremely diverse functions; stabilization seems to be a general property of interaction networks, so disruption of almost any highly connected node reduces the robustness of the system and leads to increased variation¹⁴¹. A comprehensive study of capacitor properties of yeast mutants revealed that disruption of any one of approximately 300 yeast genes (about 6% of the total) significantly decreased the robustness of yeast to environmental perturbations¹⁴².

Thus, robustness is likely to be a major, selectable mechanism that dampens evolutionary constraints — particularly those caused by interaction between mutations — and enhances plasticity. However, evidence has also been presented that certain forms of robustness, in particular the robustness of metabolic networks that are dependent on redundant reactions, could be byproducts of evolution driven by other factors, such as regulatory versatility¹⁴³.

Concluding remarks and outlook

The prevailing theme that has emerged from recent advances in evolutionary genomics and systems biology is the plurality of constraints that affect the evolution of different types of sequences in genomes, genome architectures, gene repertoires and molecular phenomes. In addition, it has emerged that there are major differences in evolutionary regimens among taxa. Beyond this diversity, comparative-genomic and molecular-phenomic analyses reveal universal patterns that could be compatible with relatively simple, general models of evolution. As discussed here, these models are starting to suggest fundamental causes underlying important aspects of evolution, such as the universal constraints on the evolution of proteins and gene repertoires (TABLE 1). It seems appropriate to expand the notion of constraints to include not only selective but also ‘neutral’ constraints that are determined by non-selective, stochastic properties of biological systems and are often amenable to modelling using techniques borrowed from statistical physics^144,145 (TABLE 1).

Table 1.

Main evolutionary constraints and universal dependences across levels of biological organization

Level of organization	Relevant constraints	Universal dependences/patterns	Putative underlying process/model
Sequence of protein-coding genes	Protein robustness to misfolding depends on translation rate*	• Approximately log-normal distribution of evolutionary rates of protein-coding genes • Anticorrelation between sequence evolution rate and expression level (translation rate) • Positive correlation among Ka, Ks and 3′ UTR evolution rate	Protein folding
Sequence of genes for non-protein-coding RNAs	RNA secondary structure and binding sites within loops	Differential conservation of ases in stem and loop structures	RNA folding
Non-coding sequences	• Regulatory site structure • Chromatin structure	• Weak constraints (on average) • Unconstrained evolution	• Biased mutational process • Sequence elimination by purifying selection depends on Ne
Gene architecture	• Compactness • Versatility	• Multidomain organization of proteins • Promiscuous domains	• Streamlining by purifying selection dependent on Ne • Recombination leading to domain rearrangement
Local gene order	Co-regulation of functionally linked genes (primarily in prokaryotes)	Closely spaced genes in partially conserved operons	‘Selfish’ operon spread
Long-range gene order	• Large-scale chromatin organization • Interaction between expression and replication	• Weak constraints • Unconstrained evolution	Random chromosome rearrangement, including origin-centred inversion in prokaryotes
Gene number (lower bound)	Minimal set of functions for sustaining a functional cell	Roughly constant size of minimal gene sets obtained with different approaches (250–300 genes)	Genome reduction; non-orthologous gene displacement
Gene number (upper bound) and gene repertoire	• Ratio of regulators to regulated genes • Lifestyle-dependent functional demands	Distinct scaling laws for different functional classes of genes	‘Toolbox’-like growth of metabolic networks
Paralogous gene families	Dependence of gene gain and loss rates on family size under the condition of genomic equilibrium	Power law-like distribution of the sizes of paralogous gene families	Birth and death process of gene evolution
Interaction and regulatory networks	Contingency of network evolution on the pre-existing network structure	Power law-like distribution of node degrees	Network evolution by preferential attachment and/or by gene duplication

^*The total number of molecules of a given protein made per unit time by translation of all cognate mRNA molecules in the cell. Ne, effective population size.

Evolutionary trajectories in sequence space seem to be strongly constrained, thus substantially limiting the ‘tinkering potential’ of evolution¹¹⁷. The evolutionary process thus seems to be a compromise “between design and bricolage”¹⁴⁶, the design aspect brought about by constraints and the bricolage stemming from the evolved robustness and the ensuing plasticity of evolving organisms.

Comparative genomics and systems approaches are transforming evolutionary biology into a much more complex but also more precise, quantitative field than it was in the twentieth century. of course, evolutionary biology is only at the beginning of the path from ‘stamp collection’ to physics. Many key quantitative relationships, including some of those discussed in this Review, remain open for reinterpretation in light of new data and theoretical models. Nevertheless, it is our belief that the transition has passed the point of no return. Next-generation sequencing, quantitative proteomics and other new methods, combined with more specific approaches of experimental evolution, should reveal the specific constraints that affect diverse aspects of genome and phenome evolution.

Robustness.

The ability to maintain a phenotype or function in the presence of internal or external perturbations.

Purifying selection.

(Also known as negative or stabilizing selection.) Mode of natural selection that eliminates deleterious mutations and preserves the status quo; in protein-coding genes, it is manifested as Ka/Ks << 1.

Box 1 Measuring selection.

Over 25 years ago, Motoo Kimura proposed that pseudogenes could be used as a neutral baseline for measuring selection¹. In general, this contention stands²⁴; furthermore, genomics revealed additional sources of (apparently) neutrally evolving sequences, such as introns and intergenic regions in animals^147,148. However, different parts of a genome differ in their mutation rate¹⁴⁹, so the neutral model can only be a reliable estimate of the strength of selection or constraints if it is derived from the same gene or genomic region in which selection is being measured. Several such measures have been developed^17,150,151.

The common gauge of selection pressure on protein-coding sequences follows from the redundancy and non-random structure of the genetic code, in which the same amino acid typically is encoded by codons that differ only in their third (less commonly first) positions. This measure, Ka/Ks (dN/dS), is the ratio of the per-site number or rate of non-synonymous substitutions to the number or rate of synonymous substitutions^20,152. The assumption underpinning the use of Ka/Ks as a measure of selection is that, unlike non-synonymous sites, synonymous sites evolve neutrally, allowing the use of synonymous sites as the baseline for measuring the constraints on protein evolution. As a crude approximation, this assumption holds: for most protein-coding genes from any organism, Ka/Ks << 1, which indicates that most proteins are subject to purifying selection. The figure shows the distributions of evolutionary rates for non-synonymous and synonymous sites of protein-coding genes in primates (a) and the Ka/Ks ratios for three diverse pairs of species (b)⁸⁸. The distribution of Ka is much broader than the distribution of Ks, indicating that the constraints affecting proteins are qualitatively different from and much more diverse than those affecting synonymous sites. For unconstrained, neutral evolution, Ka = Ks, as is the case for most pseudogenes. For a small subset of protein-coding genes, Ka/Ks > 1, which might indicate positive selection.

More accurate and powerful tests for detecting purifying and positive selection on different classes of sites are variations of the McDonald-Kreitman test. This test compares the patterns of substitutions for within-species variation (polymorphisms) with those for between-species divergence, under the assumption that the fraction of non-neutral polymorphisms is negligible¹⁷.

An independent approach for estimating the fraction of constrained sites in a genome is based on the deviations from the expected neutral distribution of insertions and deletions³⁰. It has to be kept in mind that population genetics itself is a field in flux, in large part owing to the advances of comparative genomics. The tests outlined here are based on simplified models of evolution, so more realistic and appropriately complex models may affect the estimates of selection strength in different classes of site¹⁵¹.

Non-synonymous substitutions.

Nucleotide substitutions in protein-coding genes that lead to amino acid changes in the encoded protein.

Synonymous substitutions.

Nucleotide substitutions in protein-coding genes that occur in synonymous positions of codons and accordingly do not lead to amino acid changes in the encoded protein.

Positive selection.

(Also known as directional or Darwinian selection.) Mode of natural selection that increases the frequency of initially rare beneficial alleles in a population; in protein-coding genes, this often leads to Ka > Ks.

Ultraconserved elements.

Sequences in animal genomes that have retained their identity throughout long evolutionary spans, such as the entire course of vertebrate evolution.

Evolutionary domains.

Distinct units of gene/protein evolution that form combinations with varying degrees of evolutionary stability. evolutionary domains may or may not correspond to structural domains (that is, an evolutionary domain could encompass one or more structural domains).

Promiscuous domain.

A protein domain that combines with diverse other domains in numerous proteins, providing malleable connections in interaction and regulatory networks and complexes.

Orthologues.

Genes that evolved from a single ancestral gene in the last common ancestor of the compared genomes (in contrast to paralogues).

Selfish operon concept.

A hypothesis according to which the presence of the same or similar operons in different prokaryotes is due more to the horizontal transfer of operons as distinct units than to selection for co-expression and co-regulation. When a transferred piece of DNA includes an entire operon consisting of genes encoding a complete pathway or functional system, the chances of fixation dramatically increase.

Minimal gene set for cellular life.

The minimal set of genes that is sufficient to maintain a functional cell.

Non-orthologous gene displacement.

The utilization of unrelated or distantly related (not orthologous) genes for the same function.

Box 2 The bureaucratic ceiling of genomic complexity.

The bureaucratic ceiling of genomic complexity is a hypothesis that is proposed to explain the upper gene bound. According to this theory, the number of regulators scales much steeper than linearly with the total number of genes; at a certain point the regulatory network inflation would become unsustainable, curbing the expansion of the gene complement.

Different functional classes of genes scale differently with the total number of genes in a genome. Some variation notwithstanding, in bacteria and archaea (prokaryotes) there seem to be three fundamental exponents that characterize these dependences: 0, 1 and 2 (REFS 57,153). The figure shows the differential scaling of four broad classes of genes with the total number of genes in prokaryotic genomes. Genes for proteins involved in information processing (translation, transcription and replication; yellow) scale with a 0 exponent: the number of these genes reaches a plateau even in the smallest genomes and effectively does not depend on overall genomic complexity. Metabolic enzymes and transport proteins (green) scale roughly proportionally to the total number of genes, whereas regulators (blue) and signal transduction system components (red) scale quadratically.

The characteristic exponents of the three broad functional classes of genes show remarkably little variation across prokaryotic lineages, suggesting that the differential evolutionary dynamics of genes with different functions reflect fundamental ‘laws’ of evolution of cellular organization¹⁵⁴ — that is, distinct, strong constraints on the functional composition of genomes. Eukaryotic genes show similar but less pronounced patterns of power-law gene scaling, with the exponent for the regulatory genes being substantially greater than 1 (although less than 2)¹⁵³.

The deep underlying causes of the superlinear scaling of the regulators are not understood. A simple toolbox model of evolution of prokaryotic metabolic networks seems to be compatible with the quadratic scaling of regulators¹⁵⁵. Regardless of the exact underlying mechanisms, the superlinear scaling of the regulators clearly could determine the upper limit of the growth of the gene number. At some point (which is not easy to identify precisely), the cost of adding extra regulation (‘inflating bureaucracy’) will inevitably become unsustainable, curbing the growth of genetic complexity.

The data in the figure are taken from REF. 57.

Toolbox model of evolution.

A model according to which enzymes for utilizing new metabolites, together with their dedicated regulators, are added (primarily by horizontal gene transfer) to a progressively versatile reaction network. Because of the growing complexity of the pre-existing network that provides enzymes for intermediate reactions, the ratio of regulators to regulated genes grows steadily.

Paralogous gene families.

Gene families that evolved by duplication.

Box 3 Misfolding-driven protein evolution.

In general, the rate of evolution of a protein is determined by the size of the (nearly) neutral sequence network^139,156. If the fitness of a particular sequence mostly depends on its robustness to misfolding and on expression level^90,96, the size of the nearly neutral network depends on the height and shape of the robustness peak occupied by this sequence and its neighbours in the sequence space. The figure shows a conceptual model of misfolding-driven protein evolution.

The cartoon schematically shows the effective robustness/fitness landscapes, which integrate the robustness of the native sequence and all of its mistranslated variants for two protein families (X and Y) at high and low expression levels. The high fitness/robustness area (green) reflects the size of the nearly neutral network in the sequence space.

Recent analysis of the relationships between protein abundances and evolution rates in orthologous proteins reveals a significant non-random association between intrinsic structural properties of proteins and their translation rate (as approximated by measured abundance)¹¹⁰. These results suggest that highly expressed proteins that were selected for robust folding occupy tall, steep peaks that have small areas of high fitness (that is, the proteins occupy a small nearly neutral network); as such, these proteins evolve slowly. By contrast, proteins with lower robustness occupy lower and wider peaks that have larger areas of high fitness at typical expression levels, and this larger nearly neutral network allows faster evolution (see figure).

The original hypothesis on misfolding-dominated evolution of protein-coding genes held that misfolding was largely induced by mistranslation of the coding sequence^90,157. Whether this is the case or whether stochastic misfolding of the native sequence is equally or even more common, mistranslation (somatic mutation), which is relatively frequent (10^-4-10^-5 per codon¹⁵⁸), is likely to be an important factor affecting protein evolution. Error-prone translation of a protein-coding gene produces a ‘cloud’ of neighbours in sequence space, resulting in smoothing of a potentially very rugged robustness landscape²¹. The fitness of an extremely robust sequence is decreased owing to the appearance of mistranslated variants; conversely, a mutation leading to a poorly folding protein could be less detrimental because of non-negligible production of robustly folding mistranslated proteins. This effect could open up evolutionary trajectories on the fitness landscape that are inaccessible under perfect fidelity of translation and might lead to the striking phenomenon of evolutionary anticipation¹⁵⁹.

Neutral sequence network.

A network of sequences connected by effectively single-step mutation distances (although not necessarily by single replacements), and in which there is a negligible fitness difference between neighbours.

Evolutionary anticipation.

(Also known as the look- ahead effect.) A scenario for the evolution of complex traits that require multiple mutations. In this scenario, the fixation of the final, beneficial mutation that leads to the emergence of the complex feature is enabled by a preceding random mutational walk over the neutral sequence network or by phenotypic mutations, such as mistranslation.

Experimental evolution.

The evolution of organisms with precisely defined genetic backgrounds and known evolutionary histories under controlled laboratory conditions.

Epistasis.

When non-allelic genes interact to produce a joint phenotype that differs from the one that would have been produced if the two genes had acted independently.

Pleiotropy.

Describes the multiple functions or mutation consequences of a single gene.

Fitness landscape.

A multidimensional surface defining the relationships between the fitness and the genotype spaces.

Fitness seascape.

A generalization of the concept of a fitness landscape, in which the dependence of fitness on sequence evolves over time.

Effective population size.

The size of an idealized panmictic population whose evolutionary behaviour is equivalent to that of the analysed population.

Pathogenicity islands.

Large clusters of genes in bacterial genomes that are typically transferred horizontally and contain pathogenicity determinants.