JSTOR

The Digital Microcosm

We use the Avida digital evolution system (Lenski et al. 2003; Chow et al. 2004; Ofria and Wilke 2004) as a platform for these experiments. This and similar computational systems have been used to study a number of fundamental problems in evolution (Lenski et al. 1999, 2003; Yedid and Bell 2001, 2002; Cooper and Ofria 2002; Chow et al. 2004; Misevic et al. 2006). Ofria and Wilke (2004) and appendix A give more details about the general operation of Avida.

The virtual environment in our experiments has depletable resources, meaning an organism’s access to a given resource is reduced as the resource is consumed by competitors (Cooper and Ofria 2002; Chow et al. 2004). While the experimenter must specify in advance how functions performed by the digital organisms are mapped to resources, the particular organisms that use those resources—as well as the manner in which they do so—may evolve freely. Biotic interactions are simple and facilitative: the digital organisms consume resources and generate by‐products that can themselves serve as resources for other individuals, permitting construction of cross‐feeding, codependent environments with trophic structure. Therefore, the disappearance of organisms producing certain resources can also result in the extinction of other organisms dependent on those resources. A schematic of the interactive network used is shown in figure 1. For example, the simple logic function NOT is linked to three higher‐level functions (AND, ORN, and OR) and must be performed three times—each time consuming a unit of the resource mapped to NOT—in order to produce one unit of each resource mapped to those higher‐level functions (a 3:1:1:1 conversion). Consumption on higher trophic levels works similarly. This arrangement emulates a directed‐energy‐transfer system where efficiency decreases between levels (Lindeman 1942) and produces a trophic pyramid where most of the total functionality of the community (and thus produced energy) is at the bottom, decreasing progressively at higher levels. There are no poisons, and “downward” movement of resources from higher levels to lower levels is not implemented. Usually, the digital organisms evolve the ability to perform several kinds of logic function per organism (Cooper and Ofria 2002; Chow et al. 2004), so a given organism may sometimes use its own by‐products as additional resources.

(24 KB)

Figure 1: Schematic of the cascading trophic interactions used in this study. Resources are associated with each of the logic functions shown. The reward value for performing the particular function is shown in parentheses next to the function name. A line connecting resources signifies that the lower‐level function consumes the incoming resource and produces a by‐product that is available for the higher‐level function. Only the resources associated with the lowest‐level functions NOT and NAND are provided exogenously. Example conversion factors are shown to the right of one of the inflowing resources and on the connection arrows; in this case, three units of the NOT resource are required to produce one unit each of the resources for AND, ORN, and OR. Similarly, three units of the AND resource are required to produce one unit each of the ANDN, NOR, and XOR resources.

Open New Window

We emphasize that this arrangement is not a true food web because true predation (direct consumption of one organism by another) is not implemented. Rather, the cross‐feeding relationships that arise among the digital organisms more closely resemble the coexisting bacterial ecotypes reported by Rozen et al. (2005), where a later‐evolving type develops the ability to consume metabolites produced by an earlier‐evolved type, thus enabling mutual coexistence. Further, the arrangement and stoichiometry shown in figure 1 are user configured, not inherent to the system. Alternate environment setups that either use direct 1:1:1:1 conversions or have all resources provided at equal inflow rates without any ecological interactions are unsatisfactory, in that they produce unrealistic, inverted trophic pyramid structures where the middle and higher levels have more total functionality.

Instantiating Extinction

We examine “press” and “pulse” extinctions. In geological terms, pulse extinctions happen with sufficient speed and power that no adaptive change occurs during the extinction episode, while press extinctions occur over a longer period that (theoretically) allows for an adaptive response in affected populations (Erwin 1998b). Pulse extinctions occur here by an instantaneous mass culling of organisms from the population, leaving a small number of survivors whose descendants may diversify to fill any vacated niches. In contrast, press extinctions are provoked by lowering the inflow rate of resources to near‐starvation levels for a protracted period of time (a press episode). This episode creates conditions favoring different adaptations in the community. More complicated geochemical cycle models have simulated extinction‐driving conditions through sudden drastic reductions or shutdowns of productivity (Grard et al. 2005; Rampino and Caldeira 2005). Unlike the case in pulse extinctions, there is no massive population cull and resultant loss of biomass. Organisms with more complex phenotypes are actively replaced by simpler ones as resources for higher trophic levels become unavailable, so the extinction magnitude is not standardized across replicates, as with pulse extinctions. It can produce, at a maximum, 99.972% population extinction by eliminating all but one preextinction lineage, which then becomes the most recent common ancestor for the postextinction population. While the press episodes may be very severe in the magnitude of the reduction, results in appendix E suggest that reductions above a near‐starvation threshold are inadequate for producing any meaningful extinction response, at least in the time frame used in these experiments. Ecological recovery is then later initiated by restoring resource inflows to preextinction levels. A press episode creates altered selective pressures that often lower diversity, at least over the short term. The closest biological analogue is food web collapse resulting from disruption of primary production, which has been implicated in several of the major extinctions in Earth history (Rhodes and Thayer 1991; Martin 1996; Hallam and Wignall 1997; Twitchett 2001; Benton and Twitchett 2003). Our model differs from previous ones that rely on direct, targeted removals at the primary‐producer level (Amaral and Meyer 1999; Solé et al. 2002); further discussion of differences between our model and these is provided in appendix A. We trigger a major alteration of the abiotic environment and let the ecosystem adjust and evolve entirely on its own. Our focus here is on the community‐level effects of the extinction and recovery processes, particularly the latter.

Measuring Recovery

Paleontologists and paleoecologists use a number of different definitions of recovery that focus on different aspects of the recovery process and employ different types of data (Erwin 2000). Most common are measures based on counts of the number of taxa (at some taxonomic level) and their stratigraphic spans. However, it is difficult to apply conventional taxonomic concepts to these asexual digital organisms, and there is no a priori scheme for dividing time into defined “stratigraphic” periods. We have examined the following aspects of the system in order to gauge recovery.

Total functional activity at each trophic level over the course of the experiment. We measure the total level of expression for each of the logic functions on a particular trophic level. For example, the expression levels for the functions NOT and NAND are summed together to yield the total amount of primary‐producer activity. Functional activity is a “common currency” for all biological activity in Avida, which must be measured in different ways and by different methods for living organisms (e.g., evolved oxygen can be used to measure activity rates in photosynthetic organisms but not in heterotrophs). This metric is most similar to using paleogeochemical data to infer historical ecosystem activity (usually productivity), but we measure it directly in real time. Because of genetic correlations, logic functions are not independent of each other in their occurrence and output level, so a given organism can often perform several different functions on different trophic levels. Thus, most Avida communities do not have many strict representatives at each trophic level. For example, many organisms in a population may have some level of primary‐producer activity in addition to other functions they perform. Similar complications occur in real ecosystems, such as omnivores that consume both plants and animals or facultative autotrophs among microbes. These factors contribute to variable resilience to the press episode (see “Results”) but do mean that very high loss of productivity occurs before top‐level function disappears, probably much higher than would be required to collapse real‐world ecosystems with lesser degrees of omnivory.

Number of phenotypes representing distinct functional combinations present in the community. Phenotypes in Avida are based on the set of logic functions a digital organism can perform. Logic function expression is measured in a binary manner, such that two genotypes that perform the same set of functions are scored as the same phenotype. Phenotypic groupings are often not monophyletic because organisms from multiple clades may independently evolve the same set of functions. However, many organisms with the same phenotype will often share a recent common ancestor, especially for sets comprising several or higher‐level functions. This metric may underestimate diversity because two genotypes that perform the same set of functions but at different output levels and deriving different relative benefit from each are combined into one group.

Experimental Platform and Environmental Configuration

Specifics of the experimental platform and environmental configuration are given in appendix A.

Experimental Methodology

We perform a total of 100 replicates for each treatment (except as indicated otherwise). Absolute time is measured in “updates,” an arbitrary unit of time where each individual in the population executes, on average, 30 instructions. The following types of treatments are performed.

Preextinction evolution. Each replicate runs for 100,000 updates of evolutionary time.

Press/recovery treatment. Each replicate runs for 100,000 updates exactly as above, including using the same initial random‐number seed. Resource inflows are then lowered by two orders of magnitude for 5,000 updates and then restored to preextinction levels for a subsequent 100,000 updates of evolution. This period—the press episode—is applied uniformly across experiments as absolute time, much as extinction‐driving crises act for a specific amount of geological time, irrespective of the generation times of the organisms affected. However, they are generally short compared with preceding background periods. Results for a permanently applied press and a graded restoration of resources are shown in appendix E.

Press‐only treatment. Each replicate runs for 100,000 updates exactly as above, including the same initial random‐number seed. Resource inflows are then lowered by two orders of magnitude for 5,000 updates and then restored to preextinction levels for 100,000 updates. Mutation is then turned off in concert with restoration of resource inflows in order to isolate organisms from the end of the press episode that can coexist stably without any further evolutionary change (Cooper and Ofria 2002). These stably coexisting organisms are used for supplementary experiments (see app. D).

Pulse extinctions. Each replicate runs for 100,000 updates, at which time an instantaneous mass cull of the population is performed, with no alteration of resource inflows. Survivors are picked at random only from the pool of viable organisms. This pulse extinction is followed by 100,000 updates of recovery. We perform culls of four individuals from the preextinction size of 3,600 organisms (99.9% extinction) because culls of fewer than this number often isolate survivors that are rendered sterile by mutation shortly after the pulse.

Uninterrupted evolution. Each replicate runs for 205,000 updates, with no extinction treatment. This treatment serves as a control to see how evolution would have progressed in the absence of any perturbation. Control results are presented in figure B3.

We also perform a few “large‐world” experiments that feature a total community size of and where 30 logic functions were rewarded (out of a possible set of 77). These supplementary experiments examine the effect of increased population size and environmental complexity of ecosystems on the general results obtained in the main experiments. The environmental configuration and typical results are presented in figures B4 and B5.

Although we compare and contrast the effects of press and pulse extinctions, they differ in multiple respects, such as the effect of stochastic extinctions versus selective extinctions, as well as in demographic factors, including differences in survivor population size and duration of the extinction episodes. In particular, the press extinctions do not permit a straightforward estimation of extinction magnitude, as do the pulse extinctions. Quantifying the magnitude of press extinctions requires examination of genealogical and phylogenetic details of the survivors, which lies beyond the scope of this study. We present evidence that the effects of selection during the press episode are an especially important factor influencing the dynamics of recovery.

Response of Generation Time, Effects of Density, and Survivor Composition

We also examine the behavior of the average generation time (the total number of instructions required for an organism to execute its entire code and thereby replicate itself) for the press and pulse extinctions and the correspondence between absolute time and generational time, which helps verify and characterize the adaptive response of organisms during the press episode. Additionally, we examine the effect of varying the density and specific composition of the survivor organisms on the recovery dynamics. Results for these supplementary analyses and experiments are presented in appendix D.

Analysis

Where our graphs show trajectories averaged over multiple replicates of a treatment, we usually present approximate upper and lower 95% confidence series around the average trajectory, except in cases where these would obscure the result. When these intervals exclude the point estimates for another treatment at a particular time, those treatments are then judged to be significantly different. Although the values at sequential time points are not independent in the series here, the differences are clear and compelling at many (if not most) time points.

Press Extinctions: Loss of Top‐Level Function with Productivity Decline

In 78 of 100 replicates, trophic level L3 functionality first evolved and disappeared during the press episode. However, the virtual ecosystems demonstrated varying degrees of resilience. L3 expression disappeared completely only after trophic level L0 expression declined to between 91.27% and 99.98% (average [2 SE]) of its immediate preextinction value, which required between 825 and 4,900 updates to occur (average [2 SE]). Thus, top consumption was extinguished only after a very large decline in productivity. Additional information on functional degradation during the press episode is provided in figure B1 and tables B1, B2.

Press Extinctions: Recovery of Functional Activity

We first address the dynamics of total functional activity at each trophic level before and after the press treatments. Clear differences in the recovery of total functional activity are evident between trophic levels, designated L0–L3 (fig. 2A). The two lowest levels, L0 and L1, require, on average, 28,000 and 32,000 updates, respectively, before recovery to preextinction output levels is attained. L2 experiences a much longer delay of about 85,000 updates, on average, before recovery to the immediate preextinction value. The curve for L3, which contains only the difficult function EQU, suggests that recovery for this level is highly heterogeneous between replicates. This heterogeneity is indicated by the shape of the curve, the extremely wide confidence intervals (see fig. B2 for clearer view), and the absence of full recovery, on average, even after 100,000 updates of restored resource inflow. The time required for EQU to reevolve ranges from only a few tens of updates to tens of thousands of updates among the communities where it did reevolve. These patterns persist even when only those replicates that successfully reevolved EQU are considered (fig. B2), suggesting that even when EQU does reevolve, there is often a failure to reattain preextinction levels of expression in the allotted time. Examining the individual replicates reveals considerable heterogeneity in the results. Figure 2B plots results from a community where the great majority of organisms were specialists on a single resource just before the extinction episode. This community fails to recover its preextinction output levels for EQU, although the function itself does eventually reevolve (but note the considerable gap in time before this happens). The recoveries for trophic levels L0 and L1, by contrast, are very rapid once resources are restored. Further, the total functional activity for these levels at the end of the experiment is nearly four times the immediate preextinction level. This outcome suggests that the performance of the functions on these trophic levels is qualitatively superior to that of the preextinction versions. L2 for this community displays intermediate behavior, with an initial rapid reevolution of functionality and an increase in output followed by a deceleration, so that about 45,000 updates are required before preextinction output levels recover fully. Other characteristic results, for both basic environments and more complex environments, are presented in figures B3 and B6.

(37 KB)

Figure 2: Plots of recovery of functional output following press extinctions for trophic levels L0–L3. The Y‐axis is total functional output (in thousands of executions), and the X‐axis is time (in thousands of updates). A, Average of all 100 replicates. Horizontal lines originating at immediate preextinction values indicate approximate times at which functional outputs, on average, reattained their preextinction values where they cross the recovery time series. B, Result from a single‐example replicate where most organisms were single‐resource specialists before the extinction. Trophic levels: L0, green curve; L1, blue curve; L2, purple curve; L3, red curve.

Open New Window

Also observed is a substantial difference in speed between the preextinction increase and the postextinction increase in expression of L2 and L3 functions. Further discussion is provided in appendix D.

Press Extinctions: Recovery of Phenotypic Diversity

Following the press extinctions, the full preextinction phenotypic diversity is not, on average, recovered (fig. C1). This outcome is due largely to the fact that the most complex and difficult functions, XOR and EQU, often not do not reevolve in the allotted time. A slight continuation of the upward trend is, however, apparent toward the end of the experiments, suggesting that full recovery might eventually occur, albeit over a very long period relative to the preextinction diversification. When individual replicates are examined, again considering only the populations that reevolve EQU, we see diverse outcomes ranging from little recovery even by the end of the experiment (fig. 3A), through both lengthy and short delays with full recovery (fig. 3B, 3C), to rapid and full recovery of phenotypic diversity (fig. 3D). In figure C6, we present results of allowing longer recovery times for replicates that did not recover fully in the main experiments.

(35 KB)

Figure 3: Phenotypic recovery versus time in four illustrative press extinction communities. The Y‐axis is number of phenotypes, and the X‐axis is time (in thousands of updates). All of these replicates reevolved L3 (EQU) functionality after the press, yet they show a diversity of recovery patterns, ranging from long‐term loss of phenotypic diversity (A), to long and short delays (B, C), to rapid recovery (D).

Open New Window

Pulse Extinctions: Recovery of Functional Activity

Recall that press extinctions involve a prolonged period of an altered environment and a less complex ecosystem, whereas pulse extinctions reflect an instantaneous loss of individuals and diversity but no sustained change in the environment. One might thus expect faster recovery from pulse extinctions than from press extinctions. As summarized in table 1, the recovery times for pulse extinctions tend to be much shorter than for the press extinctions at all trophic levels. On average, the recovery times for total functional activity on the simple L0 and L1 levels require fewer than 8,000 updates, whereas recovery from the press extinctions requires around 30,000 updates for full functional recovery on both these trophic levels (table 1; fig. 4). The recovery of total functional activity at L2 to previous levels required, on average, around 19,950 updates, as compared with more than 85,000 updates for the press extinctions. In contrast to the press extinctions, on the highest and most complex trophic level L3, total functional activity often does, on average, recover to its preextinction level of activity (fig. 4). In part, this faster recovery is because the corresponding EQU expression was often not completely extinguished by the pulse event. When all replicates are considered, recovery of EQU expression takes, on average, 37,050 updates. EQU activity dropped to zero following the pulse in 27 of 100 replicates, and these cases require, on average, 75,800 updates before recovery to preextinction levels (fig. C2).

Table 1: Summary of recovery of functional activity showing time required to reach preextinction level

Open New Window

(31 KB)

Figure 4: Plots of recovery of functional output following pulse extinctions for trophic levels L0–L3. Time series are the average of all 100 replicates. Horizontal lines originating at immediate preextinction values indicate approximate times at which functional outputs, on average, reattained their preextinction values where they cross the recovery time series. Axes and colors are as in figure 2.

Open New Window

Variability in the recovery response among postpulse communities is largely confined to L3, the highest trophic level (fig. C3). When the early stages of the recovery are examined more closely, with time suitably rescaled to make the beginning of the recovery 0 for both extinction treatments, the pulse communities undergo a considerably faster rate of increase at all trophic levels than do the press communities (fig. 5A, 5B). The faster recovery in the pulse communities occurs even though these communities have, on average, lower levels of total functional activity for the first few hundred updates immediately after the end of the perturbation. (Fig. 5A, 5B presents only trophic levels L0 and L3; results for L1 and L2 are presented in fig. C4.) The trends are significantly different from each other at all times, except around where the trend lines intersect.

(22 KB)

Figure 5: Comparison of both functional recovery (A, B) and phenotypic recovery (C) in press extinction communities versus pulse extinction communities over the first 3,000 updates of recovery. Time has been rescaled so that the beginning of the recovery period is 0 for both extinction treatments. Curves show averages of all 100 replicates. In all panels, the X‐axis is time (in thousands of updates). Solid lines indicate pulse replicates, and dashed lines indicate press replicates. Error series are omitted for clarity. Trends are significantly different at all points, except around where trendlines intersect. A, Trophic level L0. The Y‐axis is functional activity (in thousands of executions). B, Trophic level L3. The Y‐axis is functional activity (in thousands of executions). C, Recovery of phenotypic diversity. The Y‐axis is number of phenotypes.

Open New Window

Pulse Extinctions: Recovery of Phenotypic Diversity

Unlike communities recovering from the press extinctions, most pulse communities quickly and fully recover phenotypic diversity comparable to their preextinction diversity, although a few communities are still a bit depressed, again typically on the higher trophic levels (fig. C5). As described above, when time is rescaled to the beginning of the recovery, the pulse communities again show faster recovery of phenotypic diversity than do the press communities, even though the pulse communities begin recovery, on average, with fewer phenotypes than the end‐press communities (fig. 5C). Again, the trends are significantly different from each other at all times, except around where the trend lines intersect.

Functional Recovery

When recovery is gauged by total functional activity on a particular trophic level, we observe that communities recover much faster, on average, from random pulse extinctions than they do from press extinctions. The largest absolute differences are seen at the higher trophic levels, L2 and L3 (figs. 2, 4; table 1). When we consider the relative differences between press recoveries and pulse recoveries, the differences are rather similar across trophic levels (table 1). The lower levels with more easily evolved functions, L0 and L1, have 3.7‐fold and 4.7‐fold longer average recovery times, respectively, from press extinctions than from pulse extinctions, while the corresponding difference for L2 is about 4.7‐fold. We cannot express this same ratio for L3 because, in many populations, its total functional activity still had not recovered to preextinction levels by the end of the experiments.

There are a number of important differences between the pulse extinctions and the press extinctions. Randomly chosen survivors of pulse extinctions are, from an individual standpoint, no better or worse adapted than previously, except for possibly missing other community members on whom they might depend for resources. If more organisms survive, then more of the preextinction ecological fabric will be preserved, thus spurring more rapid recovery. By contrast, the press episode requires the passage of time in order for its effects to become evident, and it provokes an adaptive response in the organisms to the changed environment that affects their recovery ability (app. D). The low‐resource press episode tends to favor organisms with shorter generation times and simplified functionality. Survivors of random pulse extinctions, by contrast, retain their preextinction functionality and generation times. Random pulse extinctions produce great loss of biomass and diversity but often fail to eliminate key taxa or ecological traits and thus have minimal ecological impact (though some exceptions certainly exist in our experiments). Communities subjected to press extinctions, by contrast, experience near‐total ecosystem collapse during the selective extinction episode and are replaced by new communities and ecosystems that evolve during the recovery period (cf. McGhee et al. 2004). Our results therefore indicate that selection favoring organisms with short generation times and associated “deevolution” of their higher trophic functions are major factors delaying recovery from press extinctions. This finding is further reinforced by the slowed recovery observed in additional experiments using preextinction organisms chosen for having short generation times (app. D). Evidently, the press episode damages not only the ecological fabric of the community but also the genetic architectures and functional potential of the surviving digital organisms. As a consequence, recovery from the press extinctions at higher trophic levels is limited both by availability of sufficient resources and by the need to reevolve the functions for using those resources. The latter component typically includes not only the time needed for the lost functions to reevolve but also additional time for quantitative expression of those functions to recover to preextinction levels. Both of these factors can be limited by the genetic backgrounds of the digital organisms.

Diversity Recovery

Diversity‐through‐time studies must accommodate the phenomena of backward and forward “smearing” of recorded organisms, which influence estimated times of originations and extinctions in the fossil record. The studies of Kirchner and Weil (2000a, 2000b) and Kirchner (2002) used the uncorrected fossil database of Sepkoski (1992, 1997, 2002), while Lu et al. (2006) used a database that incorporates corrections for effects of rock volume and preservation potential (Foote 2003). These studies accounted for smearing effects in different ways, using different assumptions, and reached differing conclusions regarding the biological basis of lags in recovery. Kirchner and Weil (2000b) and Kirchner (2002) infer “niche construction” (Odling‐Smee et al. 2003) to explain their findings: the rate of recovery is limited by the efficacy of diversification mechanisms and the availability of opportunities following the ecological breakdown caused by the mass extinction. Lu et al. (2006) instead attribute the appearance of delayed recoveries to low preservation potential during the early stages of recovery periods, claiming that extinction‐causing crises also disrupt geological processes that lead to fossil formation. They conclude that diversification commences soon after extinction‐causing perturbations have subsided and that originations are usually rapid and even pulselike (in geological time). They suggest that better‐preserved geological sections from progressively later times following the extinction, which contain more fossil taxa, lead to inference of delayed episodes of diversification (and thus recovery). In their view, most recoveries from mass extinctions occur without prolonged delays (the early Triassic being a notable exception). Another analysis of marine fossil organism databases suggests that pulsed episodes are supported more strongly for extinction of existing genera than for origination of new genera; mass‐extinction episodes are more likely to be clustered in geological time than are episodes of origination, which may be more spread out through succeeding stages (Foote 2005). Important questions remain as to the best methods to correct for preservation biases because sequence‐stratigraphic architecture can exert strong effects on fossil occurrence times and influence the resulting interpretations of paleontological data (Holland 1995, 2000; Holland and Patzkowsky 2002; Kidwell and Holland 2002). In our study, we are able to investigate both ecosystem function and phenotypic diversity unobscured by preservation biases. We present a rough simulation of incomplete preservation applied to our data in appendix E.

The metric of phenotypic diversity we use reflects the number of distinct binary functional combinations present in the population at a given time. Evolution of new functions leads to expansion of ecological diversity by exploiting a previously unused resource and/or production of new resources. It also increases the number of realized functional combinations and the possible “niches” that may be accessible. Measuring diversity this way, we observe that phenotypic extinction is indeed most pronounced during the press episode, and we obtain a variety of possible outcomes for recovery. These range from very rapid recovery of phenotypic diversity to the community settling into a state of lower phenotypic diversity over the long term, even when all functional activities fully reevolve (fig. 3). Thus, rapid recovery following extinction represents only one of a number of possible outcomes in our system, though it may occur frequently in both digital worlds and biological worlds. Previous modeling studies (Sepkoski 1978, 1979, 1984) also feature rapid radiations as regular features of diversification following low‐diversity episodes. (Additional notes on rapid recovery are given in app. D.)

However, other outcomes for recovery of diversity (fig. 3A–3C) are certainly evident and support the view of Kirchner and Weil (2000a, 2000b) and Kirchner (2002) that recovery from extinction may depend on the effectiveness of mechanisms of diversification. A large quantity of resource may be available for exploitation but may go unused if no organism evolves the function necessary to access it. Conversely, an organism may evolve a potentially beneficial function from which it derives no benefit because the associated resource is either not present or too rare to be exploited effectively, and so the organism will fail to persist. Further, if the genetic and functional potential in a community decays sufficiently, as often occurs during prolonged press extinctions, it then becomes much more difficult to rebuild the highest‐level trophic functions. Thus, we see compelling evidence of the interaction between genetic factors, which are internal to the organisms, and the nature of the environmental perturbation leading to the extinction, which together influence the ecological and evolutionary dynamics of the recovery from a mass extinction in these digital communities.

Parallels and Contrasts with Paleontology: Extinction

Our results have clear and interesting parallels with data from the paleontological record where both extinction and recovery are concerned.

Extinction selectivity. Although often anecdotal, numerous studies have reported disproportionate extinction of larger‐sized and/or more morphologically complex forms (Russell 1977; Fischer 1981; Jablonski 1986, 1996; LaBarbera 1986). For example, larger planktonic foraminifera with complex architectures disappeared at both the Permian‐Triassic boundary (Stanley and Yang 1994; Ross and Ross 1995) and the Cretaceous‐Tertiary boundary (Norris 1991; Arnold et al. 1995), leaving smaller, morphologically simplified survivors. Sutural complexity of ammonoid shells was also “reset” following each major extinction (Saunders et al. 1999). Size reduction in particular has been termed the “Lilliput effect” (Urbanek 1993) and has been observed after the major extinctions (Twitchett 2001, 2006). While digital organisms lack morphology per se, they possess other phenotypic traits (such as genome size, generation time, and functional complexity) that permit observation and quantification of the evolution of simpler phenotypes. The dominant organisms observed at the end of press episodes are rapidly replicating and functionally simplified, broadly analogous to Lilliput organisms. In the Avida system, these kinds of organisms would normally be inferior competitors during background times, and so they may also be considered analogous to “disaster taxa” that proliferate opportunistically during (and in the immediate aftermath of) biotic crises (Schubert and Bottjer 1992; Looy et al. 1999), though not all real examples of disaster taxa are also Lilliput organisms. Further, real organisms have many other physiological traits that would influence their sensitivity to extinction‐driving conditions (Bambach et al. 2002), such as sensitivity to elevated CO₂ (Knoll et al. 2007). Most of these would not apply to the digital organisms, though future work should elucidate what factors help digital organisms either resist and/or rapidly adapt to press episodes of the type used here.

Removal of incumbent taxa. Mass extinctions are perhaps most notable for removal of dominant taxa in particular ecological roles, which are later filled by descendants of extinction survivors that diversify to assume those roles (Jablonski 2005). With respect to top consumption in our experiments (represented here by the function EQU), it is often the case that organisms that assume that role following the press extinctions are not descendants of the preextinction incumbent group(s). These often become completely extinct or leave degenerate survivors that do not reoccupy the ancestral niche (Yedid et al. 2008). The rapid recovery of functional activity evident in most of the pulse extinction experiments suggests that the survivors are members of preextinction incumbent groups; pulse‐type populations seeded with nonincumbents have a recovery response more similar to that of the press extinctions (app. D). With very few exceptions, the highly selective press episode leaves practically no surviving, phenotypically unaltered incumbents.

Parallels and Contrasts with Paleontology: Diversification and Recovery

We also find a number of parallels in our experiments for paleontological observations of diversification and recovery.

Delayed ecological recovery on higher trophic levels. Our results for delayed recovery of ecological activity at higher trophic levels are broadly analogous with the findings of a study using carbon isotopic flux data to infer recovery. Marine productivity recovered quickly after the Cretaceous‐Tertiary mass extinction, whereas many species on higher trophic levels (ranging from zooplankton to marine megafauna) became extinct (D’Hondt et al. 1998). However, the ecosystem as a whole took considerably longer to recover because many megafauna required replacement through adaptive radiation from terrestrial stocks. Our results for both pulse extinctions and press extinctions also display rapid recovery of productivity and longer delays on higher trophic levels. Because restoration of resources in our experiments is sudden, at least some of the delay in recovery can be explained by “readaptation” at the organismal and individual levels, effectiveness of diversification, and internal community dynamics. While these factors almost certainly operate for recoveries of real ecosystems, some delays in recovery (both ecological and taxonomic) may be caused by continued environmental perturbation after the main extinction (Pruss et al. 2006; Knoll et al. 2007; Sahney and Benton 2008). In our digital system, neither ecological nor phenotypic recovery occurs if the press episode is made permanent, and recovery may be delayed further if restoration is graded rather than sudden (app. E).

Missing phenotypes and incomplete ecosystems. Related to the previous point, a component of delayed recoveries is certain niches going unexploited for considerable lengths of time. Among terrestrial tetrapods, following the end‐Permian extinction, certain ecological guilds (including small insectivores, small piscivores, large herbivores, and top predators) had failed to reappear after approximately 15 million years of recovery, even though most ecosystem functions had been restored by this point (Benton et al. 2004). A similar gap is reported for the earlier end‐Guadalupian extinction (Sahney and Benton 2008). Our results demonstrating long delay in reevolution of function on higher trophic levels (particularly the difficult functions XOR and EQU) and the resulting incomplete communities are roughly comparable with such paleontological findings.

Relative durations of recovery. In our experiments, the pulse extinctions generally had much shorter recovery time than did the press extinctions for both functional expression (figs. 2, 4; table 1) and phenotypic diversity (fig. 5C). The pulse extinction recoveries tend to resemble extinction episodes such as the end‐Ordovician, end‐Triassic, and Cenomanian‐Turonian, where major biotic elements and guilds were less affected and communities reassembled relatively rapidly. The more extended recoveries seen in press extinctions, on the other hand, resemble those of the more ecologically disruptive end‐Permian and end‐Cretaceous extinctions, which produced more extensive restructuring of biotic communities (Erwin 1998b; McGhee et al. 2004). The end‐Cretaceous extinction is often cited as the canonical example of a pulse extinction (e.g., D’Hondt et al. 1998), and the end‐Permian extinction has been interpreted by some authors as an abrupt, single‐phase event (Jin et al. 2000; Rampino et al. 2000; Twitchett et al. 2001; but see Yin et al. 2007 for a contrary view). Here, we emphasize the severity of the disruption and selectivity during the extinction episode that later led to delayed recovery.

“Exuberant” diversification episodes. During early preextinction evolution, phenotypic diversity “overshoots” and then declines to an equilibrium value (fig. 3). This phenomenon occurs occasionally in postextinction recovery, such as in figure 3B. Some paleontologists may compare this “exploratory” phase (fig. 3) with the many bizarre organisms of the Cambrian explosion that left no descendants in later geological eras (Gould 1989). Perhaps a more modest comparison is with the early Triassic recovery of terrestrial tetrapods, where many new genera appear rapidly (in geological time) following the end‐Permian extinction, though mature communities and ecosystems emerge only later (Botha and Smith 2006; Roopnarine et al. 2007). This is an interesting avenue for future work in general because some authors report more subdued diversification in the early Triassic, compared with earlier diversification episodes in the early Cambrian and Ordovician (Erwin et al. 1987; Foote 1996, 1999)

We did not investigate this phenomenon deeply, but it may indicate that the default seed ancestor and its early descendants have an evolutionary flexibility that is not retained by most postextinction organisms. This flexibility may be expressed by (1) early, rapid diversification into a large number of phenotypic groupings, many of which are present in small numbers but do not survive over the long term, and (2) relative ease in evolving complex functions due to lack of inhibiting epistatic effects that accumulate over time. These results could be tested further, using alternative ancestors (evolved base replicators or randomly assembled replicators).

New “stable states.” More intriguing, perhaps, is the community from figure 3A, which failed to recover to preextinction levels of phenotypic diversity despite having evolved all functions. The failure of at least one community to recover preextinction phenotypic diversity even with reevolution of all functions may indicate that some communities may settle into alternative “stable states” following a mass extinction. Such a shift toward a new taxonomic and ecological balance seems to have occurred for marine communities as a result of the end‐Permian extinction, where communities dominated by sessile suspension feeders were replaced by more mobile and infaunal taxa in the Triassic (Wagner et al. 2006). Metazoan‐based reefs were largely wiped out by the end‐Permian extinction, and the succeeding microbial reefs remained dominant until the mid‐Triassic (Pruss and Bottjer 2005). The highly destructive press extinctions in particular seem more likely to provoke exactly the sort of community restructuring that would result in alternative stable states, and this is certainly one of the more interesting avenues for future work with digital organisms.

Concluding Remarks

While digital organisms and the communities they form lack many intricacies of real‐world analogues, our results bear a number of similarities to what is known from previous mass extinctions and recoveries. We believe this demonstrates that digital evolution experiments can yield insight into problems in paleobiology and paleoecology that would not be amenable to other experimental approaches, and we have suggested several promising avenues for future research. One must be cautious in extending such results to the organic world, but a key implication of this study is that full recovery from a biotic crisis may be hampered by adaptation of the survivors to the transiently impoverished conditions. It is becoming accepted that we are in the midst of another mass extinction, this one anthropogenically driven (McKinney 1997; Myers and Knoll 2001; Bottjer and Gall 2005). Both examining past extinction/recovery episodes and modeling extinction and recovery as general phenomena can shed light on what may yet occur. If so, our results have bearing on potential future trends in ecology and biological diversity. Human activities constitute a continued environmental stress that removes certain taxa yet promotes survival of others. If, in the near future, the world comes to be dominated by taxa that not only survive in the conditions created by human activity but also are well‐adapted to and thrive in them, then recovery of the biosphere to a state comparable with even preindustrial conditions may occur on a timescale measured in many thousands of human generations.