Cel normal host protein into the prion form. However, in some ble systems often exhibit hysteresis, which means that the es, prion forms appear to be transmitted from mother most likely to daughter cells, and the evolutionary conservation of influenced by the past history of the organism For exam- prion-forming domains suggests that this ability can be ple, classic studies on the E. coli Lac operon identified beneficiaL In yeast, translational readthrough of stop co- conditions under which the transcriptional response to in- dons is caused by the aggregated prion state (called termediate levels of lactose was both bistable and hyster- [)of the translation termination factor Sup35 (Uptain etic(Novick and Weiner, 1957). In other words, transcrip- and Lindquist, 2002). Sup35 can aggregate in a variety tion of the operon occurs at one of two levels, and, at of prion conformations, thus leading to a range of pheno- intermediate lactose concentrations, the level of expres- types characterized fror ak"to"strong " prion states ion is determined by the past history of the cell. This cel and complicating the characterization of switching rates. lular memory is stable for several generations after cells However, it is clear that prion states are more stable epi- are shifted to the intermediate inducer level, thus provid- genetic states than subtelomeric expression states. The ritable network S. cerevisiae [PSI] state, for example, is stable for ap- state. Similar short-term inheritance of a memory state in proximately 105 to 10 generations( Lund and Cox, 1981). ntermediate inducer concentrations is also present in In multicellular organisms, a great deal of recent effort other regulatory systems, such as an experimentally has focused on the role of transgenerational inheritance modified yeast GAL network (Acar et al., 2005), and is of RNA molecules. Most notably, microinjection of dou- predicted to be a common feature of complex cellular ble-stranded RNAs into Caenorhabditis elegans is suffi- networks with feedback loops ient to produce a loss-of-function phenotype in a sub- Why Switch Stochastically stantial fraction of F2 animals, and this effect persists for As is clear in the examples above, traits associated with up to 80 generations after the injection(Fire et al., 1998 contingency loci and epigenetic switching typically alter Vastenhouw et aL., 2006). A small number of molecules nate between a limited number of phenotypes or"states, were sufficient to initiate this heritable effect with thanks such as the on -and off sion states of subtelomeric to amplification of the interfering RNas by RNA-depen- genes, the radial and bilateral floral symmetries, or the var dent RNA polymerase. In mammals, epigenetic inheri- ious repeat numbers in cell-surface proteins. This raises tance of RNA molecules was recently described in which the question of why organisms have developed such com- expression of unusual Kit RNAs in the germline of mice re- plex switching mechanisms to reach the seemingly simple sulted in a phenotypic effect(on coat color) in the progeny goal of tuning genes on and off. Is it not more straightfor- of the affected mice such that two genetically identical ward to use plasticity, e.g., transcriptional regulation, to mice might differ phenotypically based on their parents switch between a handful of phenotypic states? The genotypes(Rassoulzadegan et al., 2006). This last exam- adaptive benefits conferred by stochastically fluctuating ple mirrors the phenomenon of paramutation in plants phenotypes have been the subject of a number of model- which was first discovered in maize in the 1950s by ing studies (Jablonka et al., 1995; Kussell and Leibler, R Brink (see also the Essay by v. Chandler, page 641 of 2005: Wolf et al., 2005). In general, these studies suggest this issue) that random, heritable phenotypic switches may be bene Other mechanisms of epigenetic inheritance are even ficial when the environment fluctuates randomly over more cryptic than the examples above. In bacteria, persis- timescales that are roughly matched to the phenotype tence to antibiotic treatment is characterized by a small switching rate. Several of these studies also explicitly subpopulation that grows slowly and is not killed by antibi- compare stochastic switching with plasticity by modeling otic treatment(Balaban et aL., 2004). These slowly growing (1)the costs associated with maintaining sensing machin- bacterial"persister cells"are a rare phenotypic subpopu ery and(2) the time delay between sensing and phe lation, and the majority of progeny of these persisters re- typic change. For example an environment that proves in- vert to the sensitive but rapidly growing phenotype(Bala- stantly lethal cannot be dealt with by plasticity. Together, ban et al., 2004). Actively growing E coli switch to slowly these results demonstrate that some environments and growing persisters at a frequency of approximately 10-6 sensor regimes exist for which stochastic switching is (per hour, not generation), whereas persisters generate ac- the optimal organismal bet-hedging"strategy. Environ- tively growing progeny at a frequency of 10 per hour. the mental regimes where conditions persist for at least sev- mechanism for this phenotypic switch is unknown, al- eral generations, but not tens of thousands of generations though mutants with changed switching frequencies re expected to select for stochastic phenotypic variation have been identified, and the genes affected may provide on timescales not typically accessible to point-mutational clues as to the substrate for this phenotypic switch processes. Interestingly, for many of the examples de We expect that more examples of epigenetic switches scribed above, it can indeed be argued that the switching are likely to be found. Generally, switching between two frequencies could match the variability of the selective semistable states("bistability, ) is a common property of pressure that is acting upon the respective phenotype (genetic) networks with positive feedback loops(Rao (see further) et aL, 2002). By definition, bistability allows two stable Of course, random switching comes at a cost: it results states to exist in the same environment. In addition, bista- in some maladapted individuals in every generation. A 660 Cell 128, 655-668, February 23, 2007 @2007 Elsevier Indnormal host protein into the prion form. However, in some cases, prion forms appear to be transmitted from mother to daughter cells, and the evolutionary conservation of prion-forming domains suggests that this ability can be beneficial. In yeast, translational readthrough of stop co￾dons is caused by the aggregated prion state (called [PSI+ ]) of the translation termination factor Sup35 (Uptain and Lindquist, 2002). Sup35 can aggregate in a variety of prion conformations, thus leading to a range of pheno￾types characterized from ‘‘weak’’ to ‘‘strong’’ prion states and complicating the characterization of switching rates. However, it is clear that prion states are more stable epi￾genetic states than subtelomeric expression states. The S. cerevisiae [PSI+ ] state, for example, is stable for ap￾proximately 105 to 107 generations (Lund and Cox, 1981). In multicellular organisms, a great deal of recent effort has focused on the role of transgenerational inheritance of RNA molecules. Most notably, microinjection of dou￾ble-stranded RNAs into Caenorhabditis elegans is suffi- cient to produce a loss-of-function phenotype in a sub￾stantial fraction of F2 animals, and this effect persists for up to 80 generations after the injection (Fire et al., 1998; Vastenhouw et al., 2006). A small number of molecules were sufficient to initiate this heritable effect with thanks to amplification of the interfering RNAs by RNA-depen￾dent RNA polymerase. In mammals, epigenetic inheri￾tance of RNA molecules was recently described in which expression of unusual Kit RNAs in the germline of mice re￾sulted in a phenotypic effect (on coat color) in the progeny of the affected mice such that two genetically identical mice might differ phenotypically based on their parents’ genotypes (Rassoulzadegan et al., 2006). This last exam￾ple mirrors the phenomenon of paramutation in plants, which was first discovered in maize in the 1950s by R. Brink (see also the Essay by V. Chandler, page 641 of this issue). Other mechanisms of epigenetic inheritance are even more cryptic than the examples above. In bacteria, persis￾tence to antibiotic treatment is characterized by a small subpopulation that grows slowly and is not killed by antibi￾otic treatment (Balaban et al., 2004). These slowly growing bacterial ‘‘persister cells’’ are a rare phenotypic subpopu￾lation, and the majority of progeny of these persisters re￾vert to the sensitive but rapidly growing phenotype (Bala￾ban et al., 2004). Actively growing E. coli switch to slowly growing persisters at a frequency of approximately 106 (per hour, not generation), whereas persisters generate ac￾tively growing progeny at a frequency of 101 per hour. The mechanism for this phenotypic switch is unknown, al￾though mutants with changed switching frequencies have been identified, and the genes affected may provide clues as to the substrate for this phenotypic switch. We expect that more examples of epigenetic switches are likely to be found. Generally, switching between two semistable states (‘‘bistability’’) is a common property of (genetic) networks with positive feedback loops (Rao et al., 2002). By definition, bistability allows two stable states to exist in the same environment. In addition, bista￾ble systems often exhibit hysteresis, which means that the most likely state of the system in a given environment is influenced by the past history of the organism. For exam￾ple, classic studies on the E. coli Lac operon identified conditions under which the transcriptional response to in￾termediate levels of lactose was both bistable and hyster￾etic (Novick and Weiner, 1957). In other words, transcrip￾tion of the operon occurs at one of two levels, and, at intermediate lactose concentrations, the level of expres￾sion is determined by the past history of the cell. This cel￾lular memory is stable for several generations after cells are shifted to the intermediate inducer level, thus provid￾ing an example of an epigenetically heritable network state. Similar short-term inheritance of a memory state in intermediate inducer concentrations is also present in other regulatory systems, such as an experimentally modified yeast GAL network (Acar et al., 2005), and is predicted to be a common feature of complex cellular networks with feedback loops. Why Switch Stochastically? As is clear in the examples above, traits associated with contingency loci and epigenetic switching typically alter￾nate between a limited number of phenotypes or ‘‘states,’’ such as the on- and off-expression states of subtelomeric genes, the radial and bilateral floral symmetries, or the var￾ious repeat numbers in cell-surface proteins. This raises the question of why organisms have developed such com￾plex switching mechanisms to reach the seemingly simple goal of turning genes on and off. Is it not more straightfor￾ward to use plasticity, e.g., transcriptional regulation, to switch between a handful of phenotypic states? The adaptive benefits conferred by stochastically fluctuating phenotypes have been the subject of a number of model￾ing studies (Jablonka et al., 1995; Kussell and Leibler, 2005; Wolf et al., 2005). In general, these studies suggest that random, heritable phenotypic switches may be bene- ficial when the environment fluctuates randomly over timescales that are roughly matched to the phenotypic switching rate. Several of these studies also explicitly compare stochastic switching with plasticity by modeling (1) the costs associated with maintaining sensing machin￾ery and (2) the time delay between sensing and pheno￾typic change. For example, an environment that proves in￾stantly lethal cannot be dealt with by plasticity. Together, these results demonstrate that some environments and sensor regimes exist for which stochastic switching is the optimal organismal ‘‘bet-hedging’’ strategy. Environ￾mental regimes where conditions persist for at least sev￾eral generations, but not tens of thousands of generations, are expected to select for stochastic phenotypic variation on timescales not typically accessible to point-mutational processes. Interestingly, for many of the examples de￾scribed above, it can indeed be argued that the switching frequencies could match the variability of the selective pressure that is acting upon the respective phenotype (see further). Of course, random switching comes at a cost: it results in some maladapted individuals in every generation. A 660 Cell 128, 655–668, February 23, 2007 ª2007 Elsevier Inc.