tity and therefore approximately imitates the natural process of homologous recombination (although the parents are often more diverged than naturally recombining proteins). A common variation on both of these techniques is to bias the creation of mutant proteins with the goal of increasing the fraction of improved mutants, for example, by targeting functionally important residues for mutagenesis (Park et al., 2005) or choosing recombination crossover points based on structural information (Voigt et al., 2002). Researchers also sometimes use techniques to specifically change several residues simultaneously with the goal of finding coupled beneficial mutations, although (as discussed below) the mutations discovered by such approaches often turn out to have been individually beneficial, and so presumably could have also been discovered separately with lower mutation rates.
The procedure used to identify improved mutants depends on the details of the particular protein and engineering goal. In some cases, the protein property of interest can be coupled to the survival of the host cell, thereby allowing for direct genetic selection of cells carrying improved mutants. Such selections can often be applied to libraries consisting of millions of different mutants. More frequently, it is not possible to design an effective selection, and mutants must be assayed directly in a high-throughput screen. In these cases, the researcher is typically able to examine libraries of a few thousand different mutants. Such screening is nonetheless sufficient to examine most of the possible individual mutations to a parent protein, because a 200-aa protein possesses only 19 × 200 = 3,800 unique single mutants (even fewer are accessible via single nucleotide changes).
The last step in the evolutionary algorithm is using the results from the screening/selection to choose the parents for the next generation. For reasons that are not entirely clear, directed evolution experiments rarely use schemes in which each mutant contributes to the next generation with a probability proportional to its measured fitness. (In contrast, fitness-proportionate selection is widely used in computational genetic algorithms.) Instead, researchers typically choose one or a few of the best mutants as parents for the next generation. Proteins undergoing directed evolution therefore experience a series of population bottlenecks in which most of the genetic variation is purged. At the same time, the adaptive-walk nature of these experiments provides little opportunity for deleterious or neutral mutations to spread (unless they hitchhike along with beneficial ones). These experiments therefore typically fail to fully recapitulate the evolutionary dynamics of either small populations (rapid genetic drift including the occasional fixation of deleterious mutations) or large populations (the maintenance of substantial levels of standing genetic variation). Directed evolution therefore probably sheds more light on the question of