Cell directed switching strategy in which cells bias their prog- sense that fewer mutations will prevent appropriate fold ny phenotypes based on the recent environment would ing. Interestingly, using in silico folding predictions, the ar preferable (Jablonka et al., 1995). This inheritance thors found that RNa sequences that are capable of fold- strategy, which is widely disbelieved(but experiencing ing into a given structure at a wide range of temperatures a recent resurgence), is now often referred to as"La are also less prone to change their structure as a conse marckism"and will be addressed at the end of this Re- quence of mutations. This case therefore provides an ex view. We first turn to the decrease of variability in certain ample of a mechanism for the evolution of robustness known as"congruent robustness, where genetic robust ness may occur as a side effect of selection on environ- Robustness and Canalization mental robustness(Ancel and Fontana, 2000) The phenomena described above are all examples of At a more global level, it has been suggested that organ- mechanisms that increase the rate of phenotypic change isms have evolved mechanisms to increase the genetic robustness of complex phenotypes(such as body plan) nany phenotypes have proven beneficial to cells over to protect vital phenotypes from genetic insults. This countless generations through many environments. Or- was first discussed in the seminal work of Waddington ganisms might therefore have evolved mechanisms to (Waddington, 1942, 1953 ), who noted the exceptional stabilize these traits against the random degradation of stability of organismal development in the face of environ- undirected mutation. Phenotypes stabilized in the face mental perturbations and genetic mutations. He sug of genetic mutation are known as genetically robust and gested that deep"canals"seemingly direct the develop should be separated from traits that are stable in a wide mental flow and called the process canalization. This inge of environmental regimes, which are environmen- idea has its echoes today in the systems-biology ap tally robust. The idea that a phenotype could be the result proach of mathematically modeling networks and asking of many genotypes(and hence stable to mutations that over what range of parameters a given behavior can be change one genotype to another) has been described as found (see Stelling et aL., 2004 for review). Recent studies canalization"(Waddington, 1942),buffering, " or"robust- have modeled complicated networks(such as the net ness. " For quantitative traits, robustness can be defined works controlling bacterial chemotaxis or those control- using the mutational variance Vm(see above) if Vm for ling segmentation in flies) and asked what fraction of phenotype P is lower in organism A than in organism B, parameters in the model will still support a given pheno- then organism a is more genetically robust than B type, with a common theme being that feedback loops There are a number of ways that individual genes may allow a desired behavior to exist through a large fraction robustly encoded. For example, several amino acids of"parameter space"(if it is imagined that mutation e encoded by multiple codons, and these codons may changes the parameters of the network, then the feed differ in the number of mutations that change the encoded back in question makes the network genetically robust amino acid. For example, CGA, CGC, CGG, CGT, AGA, Experimentally, a treatment that increases the pheno- and AGG all code for arginine Mutation of the third base typic variance of a trait in a genetically heterogeneous for any of the CGx codons will not change the amino population has generally been considered to have com- acid encoded. whereas mutation of the third base of AGA or AGG may change the protein sequence Encoding Waddington found that treating a population of Drosophila arginine with CGx thus reduces the mutational target size larvae with elevated temperatures increases variation in of the protein(assuming that arginine is essential for the several traits(Waddington, 1953). Moreover, the interindi protein's function) by about one nucleotide. One study vidual differences that appear after such a temperature discussed this property as"codon volatility"and sug- treatment are selectable, and, once selected for, the phe gested that genes under stabilizing selection are generally notypes can become fixed (stabilized)even in the absence encoded by low-volatility codons(Plotkin et aL., 2004). of heat stress. This suggests that the stress-induced in- However, it is currently unclear whether this enrichment crease in phenotypic variation in outbred lines is due to codon bias that is selected for some reason besides robustness In any case did not result in phenotypic differences prior to the treat- it is intuitive that decreasing a phenotype's mutational ment(McLaren, 1999). Hence, the elevated temperature target size will stabilize a phenotype against mutation. is argued to have compromised some as-yet-unknown Perhaps a more obvious and widespread mechanism to genetic-robustness mechanism, thereby revealing previ- establish robustness of certain traits is gene duplication ously hidden genetic variation where the second gene copy can provide a"backup"sys Recently, it has been proposed that the temperature m when one copy is mutated esponsive robustness factor in these particular exper example of robust encoding has been de- ments is the protein chaperone Hsp90. Several studies ber of different rna sequences are capable pharmacological油hpm of folding into a given secondary structure. Some RNA covers previously hidden selectable variation in multiple sequences are more genetically robust than others in the traits (Queitsch et al., 2002; Rutherford and Lindquist, Cell 128, 655-668, February 23, 2007 @2007 Elsevier Inc. 661directed switching strategy in which cells bias their prog￾eny phenotypes based on the recent environment would be preferable (Jablonka et al., 1995). This inheritance strategy, which is widely disbelieved (but experiencing a recent resurgence), is now often referred to as ‘‘La￾marckism’’ and will be addressed at the end of this Re￾view. We first turn to the decrease of variability in certain phenotypes. Robustness and Canalization The phenomena described above are all examples of mechanisms that increase the rate of phenotypic change beyond the rate due to random mutation. Conversely, many phenotypes have proven beneficial to cells over countless generations through many environments. Or￾ganisms might therefore have evolved mechanisms to stabilize these traits against the random degradation of undirected mutation. Phenotypes stabilized in the face of genetic mutation are known as genetically robust and should be separated from traits that are stable in a wide range of environmental regimes, which are environmen￾tally robust. The idea that a phenotype could be the result of many genotypes (and hence stable to mutations that change one genotype to another) has been described as ‘‘canalization’’ (Waddington, 1942), ‘‘buffering,’’ or ‘‘robust￾ness.’’ For quantitative traits, robustness can be defined using the mutational variance Vm (see above): if Vm for phenotype P is lower in organism A than in organism B, then organism A is more genetically robust than B. There are a number of ways that individual genes may be robustly encoded. For example, several amino acids are encoded by multiple codons, and these codons may differ in the number of mutations that change the encoded amino acid. For example, CGA, CGC, CGG, CGT, AGA, and AGG all code for arginine. Mutation of the third base for any of the CGX codons will not change the amino acid encoded, whereas mutation of the third base of AGA or AGG may change the protein sequence. Encoding arginine with CGX thus reduces the mutational target size of the protein (assuming that arginine is essential for the protein’s function) by about one nucleotide. One study discussed this property as ‘‘codon volatility’’ and sug￾gested that genes under stabilizing selection are generally encoded by low-volatility codons (Plotkin et al., 2004). However, it is currently unclear whether this enrichment reflects some correlated property of codon bias that is selected for some reason besides robustness. In any case, it is intuitive that decreasing a phenotype’s mutational target size will stabilize a phenotype against mutation. Perhaps a more obvious and widespread mechanism to establish robustness of certain traits is gene duplication, where the second gene copy can provide a ‘‘backup’’ sys￾tem when one copy is mutated. Another example of robust encoding has been de￾scribed for RNA secondary structures (Ancel and Fontana, 2000). A number of different RNA sequences are capable of folding into a given secondary structure. Some RNA sequences are more genetically robust than others in the sense that fewer mutations will prevent appropriate fold￾ing. Interestingly, using in silico folding predictions, the au￾thors found that RNA sequences that are capable of fold￾ing into a given structure at a wide range of temperatures are also less prone to change their structure as a conse￾quence of mutations. This case therefore provides an ex￾ample of a mechanism for the evolution of robustness known as ‘‘congruent robustness,’’ where genetic robust￾ness may occur as a side effect of selection on environ￾mental robustness (Ancel and Fontana, 2000). At a more global level, it has been suggested that organ￾isms have evolved mechanisms to increase the genetic robustness of complex phenotypes (such as body plan) to protect vital phenotypes from genetic insults. This was first discussed in the seminal work of Waddington (Waddington, 1942, 1953), who noted the exceptional stability of organismal development in the face of environ￾mental perturbations and genetic mutations. He sug￾gested that deep ‘‘canals’’ seemingly direct the develop￾mental flow and called the process canalization. This idea has its echoes today in the systems-biology ap￾proach of mathematically modeling networks and asking over what range of parameters a given behavior can be found (see Stelling et al., 2004 for review). Recent studies have modeled complicated networks (such as the net￾works controlling bacterial chemotaxis or those control￾ling segmentation in flies) and asked what fraction of parameters in the model will still support a given pheno￾type, with a common theme being that feedback loops allow a desired behavior to exist through a large fraction of ‘‘parameter space’’ (if it is imagined that mutation changes the parameters of the network, then the feed￾back in question makes the network genetically robust). Experimentally, a treatment that increases the pheno￾typic variance of a trait in a genetically heterogeneous population has generally been considered to have com￾promised a mechanism for robustness. For example, Waddington found that treating a population of Drosophila larvae with elevated temperatures increases variation in several traits (Waddington, 1953). Moreover, the interindi￾vidual differences that appear after such a temperature treatment are selectable, and, once selected for, the phe￾notypes can become fixed (stabilized) even in the absence of heat stress. This suggests that the stress-induced in￾crease in phenotypic variation in outbred lines is due to the uncovering of pre-existing genetic differences that did not result in phenotypic differences prior to the treat￾ment (McLaren, 1999). Hence, the elevated temperature is argued to have compromised some as-yet-unknown genetic-robustness mechanism, thereby revealing previ￾ously hidden genetic variation. Recently, it has been proposed that the temperature￾responsive robustness factor in these particular experi￾ments is the protein chaperone Hsp90. Several studies in a number of organisms have shown that genetic and pharmacological interference with Hsp90 function un￾covers previously hidden selectable variation in multiple traits (Queitsch et al., 2002; Rutherford and Lindquist, Cell 128, 655–668, February 23, 2007 ª2007 Elsevier Inc. 661