The vast majority of genes in every genome are selectively constrained, in that most nucleotide changes that alter the fitness of the organism are deleterious. How do we know this? Comparisons between genomes clearly demonstrate that coding sequences diverge at slower rates than non-coding regions, largely due to a deficit of mutations at positions where a base change would cause an amino-acid change. Gene duplication provides opportunities to explore this forbidden evolutionary space more widely by generating duplicates of a gene that can ‘wander’ more freely, on condition that between them they continue to supply the original function.

Susumu Ohno was the first to comprehensively elucidate the potential of gene duplication, in his book Evolution by Gene Duplication, published more than 30 years ago (Ohno 1970). The prescience of Ohno's book is highlighted by the fact that his book has almost certainly been cited more times in the past five years than in the first five years after its publication.

Whole genome sequences of closely related organisms have allowed us to identify changes in the gene complements of species over relatively short evolutionary distances. These comparisons typically reveal dramatic expansions and contractions of gene families that can be related to underlying biological differences. For example, humans and mice differ in their sensory reliance on sight and smell respectively; colour vision in humans has been significantly enhanced by the duplication of an Opsin gene that allows us to distinguish light at three different wavelengths, while mice can distinguish only two. By contrast, a much higher proportion of the large gene family of olfactory receptors have retained their functionality in mice, as compared to humans.

‘Retrotransposition’ describes the integration of reverse transcribed mature RNAs at random sites in a genome. The resultant duplicated genes (retrogenes) lack introns and have poly-A tails. Separated from their regulatory elements, these integrated sequences rarely give rise to expressed full-length coding sequences, although functional retrogenes have been identified in most genomes.

Tandem duplication of a genomic segment (segmental duplication) is one of the possible outcomes of ‘unequal crossing over’, which results from homologous recombination between paralogous sequences. These recombination events can also give rise to the deletion or inversion of intervening sequences. Recent evidence suggests that the explosion of segmental duplications in recent primate evolution has been caused in part by the rapid proliferation of Alu elements about 40 MYA. Alu elements are derived from the 7SL RNA gene and represent the most frequent dispersed repeat in the human genome, with the approximately 1 million copies of the 300-bp Alu element representing around 10% of the entire genome. The striking enrichment of Alu elements at the junctions between duplicated and single copy sequences implicates unequal crossing over between these repeats in the generation of segmental duplications (Bailey et al. 2003).

The observation of segmental duplication events with no evidence for homology-driven unequal crossing over suggests that segmental duplications can also arise through non-homologous mechanisms. A recent screen for spontaneous duplications in yeast suggests that replication-dependent chromosome breakages also play a significant role in generating tandem duplications, because duplication breakpoints are enriched at replication termination sites (Koszul et al. 2004).

Genome duplication events generate a duplicate for every gene in the genome, representing a huge opportunity for a step-change in organismal complexity. However, genome duplication presents significant problems for the faithful transmission of a genome from one generation to the next, and is consequently a rare event, at least in Metazoa. In principle, genome duplications should be easily identified through the coincident emergence within a phylogeny of many gene families. Unfortunately, this signal is complicated by subsequent piecemeal loss and gain of gene family members. Consequently, there is heated debate over possible ancient genome duplication events in early vertebrate evolution and more recently in teleost fish, both of which must have occurred hundreds of millions of years ago (McLysaght et al. 2002; Van de Peer et al. 2003).

Despite the slackened selective constraints, mutations can still destroy the incipient functionality of a duplicated gene: for example, by introducing a premature stop codon or a mutation that destroys the structure of a major protein domain. These degenerative mutations result in the creation of a pseudogene (nonfunctionalization). Over time, the likelihood of such a mutation being introduced increases. Recent studies suggest that there is a relatively narrow time window for evolutionary exploration before degradation becomes the most likely outcome, typically of the order of 4 million years (Lynch and Conery 2000).

During the relatively brief period of relaxed selection following gene duplication, a new, advantageous allele may arise as a result of one of the gene copies gaining a new function (neofunctionalization). This can be revealed by an accelerated rate of amino-acid change after duplication in one of the gene copies. This burst of selection is necessarily episodic—once a new function is attained by one of the duplicates, selective constraints on this gene are reasserted. These patterns of selection can be observed in real data: most recently duplicated gene pairs in the human genome have diverged at different rates from their ancestral amino-acid sequence (Zhang et al. 2003). A convincing instance of neofunctionalization is the evolution of antibacterial activity in the ECP gene in Old World Monkeys and hominoids after a burst of amino-acid changes following the tandem duplication of the progenitor gene EDN (a ribonuclease) some 30 MYA (Zhang et al. 1998). The divergence of duplicated genes over time can be also monitored in genome-wide functional studies. In both yeast and nematodes, the ability of a gene to buffer the loss of its duplicate declines over time as their functional overlap decreases.

Rather than one gene duplicate retaining the original function, while the other either degrades or evolves a new function, the original functions of the single-copy gene may be partitioned between the duplicates (subfunctionalization). Many genes perform a multiplicity of subtly distinct functions, and selective pressures have resulted in a compromise between optimal sequences for each role. Partitioning these functions between the duplicates may increase the fitness of the organism by removing the conflict between two or more functions. This outcome has become associated with a population genetic model known as the Duplication–Degeneration–Complementation (DDC) model, which focuses attention on the regulatory changes after duplication (Force et al. 1999). In this model, degenerative changes occur in regulatory sequences of both duplicates, such that these changes complement each other, and the union of the expression patterns of the two duplicates reconstitutes the expression pattern of the original (Figure 2).

A recent study by Dorus and colleagues (Dorus et al. 2003) investigated the retrotransposition (since the existence of a human–mouse common ancestor) of one of the two autosomal copies of the CDYL gene to Y chromosome (forming CDY). In the mouse, both Cdyl genes produce two distinct transcripts, one of which is expressed ubiquitously while the other is testis-specific. By contrast, in humans both CDYL genes produce a single ubiquitously expressed transcript, and CDY exhibits testis-specific expression. As CDY is a retrogene (see above) that has not been duplicated together with its ancestral regulatory sequences, it is clear that the DDC model is not the only route by which to achieve spatial partitioning of ancestral expression patterns.

Earlier, we saw that homologous recombination between paralogous sequences can result in rearrangements, including tandem duplications. Such recombination events need not cause rearrangements, but can also result in the nonreciprocal transfer of sequence from one paralogue to the other—a process known as gene conversion. Gene conversion homogenizes paralogous sequences, retarding their divergence, and consequently obscuring their antiquity. This leads to the observation of ‘concerted evolution’ whereby duplicates within a species can be highly similar and yet continue to diverge between species (Figure 3). Once gene duplicates have diverged sufficiently so that they differ in their functionality (or non-functionality), gene conversion events can become deleterious—for example, by introducing disrupting mutations from a pseudogene into its functional duplicate. A substantial proportion of disease alleles in Gaucher disease result from the introduction of mutations into the glucocerebrosidase gene from a tandemly repeated pseudogene (Tayebi et al. 2003). These kinds of recombinatorial interactions only occur between paralogues that are minimally diverged. Thus, while selective interactions and functional overlap between duplicates declines relatively slowly over evolutionary time, the potential for recombinatorial interactions between paralogues is relatively short-lived.

For some genes, duplication confers an immediate selective advantage by facilitating elevated expression, or as Ohno put it, ‘duplication for the sake of producing more of the same’. This has clearly been the case for histones and ribosomal RNA genes. In this scenario, gene conversion is of potential benefit in maintaining homogeneity between copies. Certainly both histone and rDNA genes are commonly found in arrays of duplicates: structures that facilitate array homogenization by both gene conversion and repeated unequal crossing over.