Direct and correlated responses to selection on functions OR and EQU in Ancestor1
Figure 2 shows the results of experiments to examine the evolutionary association between
OR and EQU in Ancestor1. At the start of the experiment, the ancestor could perform
each of these functions only once. Each point represents the average number of times
a given function was performed by 100 independently evolved organisms, as a function
of the environment in which they evolved. Starting on the left side of the figure,
the data show that the performance of OR was higher in the two environments where
it was rewarded, +OR and +OR/-EQU, than in the two environments where it was either
not rewarded or punished (comparing upper left to lower left). This result indicates
that OR responded positively to direct selection. Of the environments in which it
was rewarded, the performance of OR was higher when EQU was punished than when it
was not (mean?=?146.07 in +OR/-EQU environment versus mean?=?117.90 in +OR environment),
although this difference was marginally non-significant (Mann–Whitney U?=?5763, n?=?200,
P?=?0.062). Focusing on the lower left part of the figure, the performance of OR also
increased above the ancestral level of 1 as a correlated response to selection on
EQU (mean OR?=?21.1 in +EQU environment). However, when OR was punished, it was invariably
lost (mean OR?=?0 in +EQU/-OR environment). This result indicates that, despite the
correlated response of OR to selection for EQU, the association between these functions
could be broken when selection acted on them in opposing directions.
Figure 2. Average performance of functions OR and EQU by organisms evolved from Ancestor1. Ancestor1
performed each function only once per reproductive cycle. The y-axis shows the number
of times a given logic function is output per reproductive cycle. Each point represents
the mean value, based on 100 replicate populations in each environment, based on the
most common genotype at the end of the experiment. The environment is shown next to
each point, and lines connect measurements made on the same set of populations. Error
bars indicate one standard error. All 100 populations lost OR in the +EQU/-OR environment,
and all 100 populations lost EQU in both the +OR/-EQU and +OR environments.
The response of EQU to these same environments is shown on the right in Figure 2. Its performance, like that of OR, evolved to higher levels in the environments where
it was directly selected (i.e., the +EQU and +EQU/-OR environments). The performance
of EQU did not differ significantly depending on whether OR was being punished or
not (Mann–Whitney U?=?4386, n?=?200, P?=?0.128), although EQU tended to be performed more often when OR was not punished.
EQU, again like OR, was always lost when its performance was punished (mean EQU?=?0
in the +OR/-EQU environment). However, EQU was also invariably lost when only OR was
rewarded. These results thus reveal an asymmetry in the correlated responses: selection
for EQU (+EQU environment) resulted in a correlated increase in the performance of
OR, but selection for OR (+OR environment) led to the complete loss of EQU.
This asymmetry is not surprising in light of the genotype-phenotype map for Ancestor1
(Figure 1A). This map shows that all the genomic instructions that, when deleted, knock out
the OR function also knock out the EQU function. However, the reverse is not true:
some instructions can be deleted that knock out EQU without affecting the performance
of OR. Put another way, the instructions encoding OR are a subset of those encoding
EQU. The observed asymmetry in the correlated responses thus reflects the underlying
asymmetry in the mapping between genotype and phenotype for these two traits.
Direct and correlated responses to selection on functions OR and EQU in Ancestor2
Qualitatively similar patterns arose when we examined the correlation between OR and
EQU in populations evolved from Ancestor2 (Figure 3A). This ancestor also performed OR and EQU only once. Selection for EQU (+EQU environment)
led to a correlated increase in OR (lower left), but selection for OR (+OR environment)
led to the complete loss of EQU (lower right). Once again, the direct response to
selection was stronger than the correlated response to selection, with the performance
of OR higher in the +OR and +OR/-EQU treatments than in the +EQU or +EQU/-OR environments
(comparing upper left to lower left). Similarly, the performance of EQU was higher
in the +EQU and +EQU/-OR treatments than in the +OR and +OR/-EQU environments (comparing
upper right to lower right). In both cases, punishing performance of the alternative
function had little effect on the evolution of the rewarded function; thus, the performance
of OR did not differ appreciably between the +OR and +OR/-EQU treatments (upper left),
and the performance of EQU was indistinguishable in the +EQU/-OR and +EQU treatments
(upper right).
Figure 3. Average performance of functions OR, AND, and EQU by organisms evolved from Ancestor2.
Ancestor2 performed each function only once per reproductive cycle. (A) Evolved performance of OR and EQU. (B) Evolved performance of AND and EQU. See the legend to Figure 2 for description of the points and error bars. All 100 populations lost OR in the
+EQU/-OR environment, and all 100 populations lost EQU in both the +OR/-EQU and +OR
environments. Also, all 100 populations lost EQU in both the +AND/-EQU and the +AND
environments. However, only 47 populations lost AND in the +EQU/-AND environment.
Direct and correlated responses to selection on functions AND and EQU in Ancestor2
At first glance, the pattern was similar when we examined the functions AND and EQU
in populations that evolved from Ancestor2 (Figure 3B). Selecting for the performance of EQU resulted in a correlated increase in the
performance of AND from its ancestral level of 1 (lower left), but selecting for AND
caused EQU to be invariably lost (lower right). This result was also expected from
the genotype-phenotype map for this ancestor; all genomic instructions that, when
deleted, knock out AND also knock out EQU, but the reverse is not true (Figure 1B). In other words, the genome instructions encoding the AND function are a subset
of those encoding EQU.
However, the outcomes of these experiments differed from the preceding ones in two
important respects. First, the performance of AND did not invariably decline to zero
when it was selected against (Figure 3B, lower left). In fact, only 47 of 100 populations that evolved in the +EQU/-AND
environment lost the ability to perform that function (Table 3; far right column). Second, EQU evolved to higher levels when AND was punished than
when it was not (average performance of EQU?=?342.8 in +EQU/-AND environment, compared
to 261.2 in +EQU environment; Figure 3B, upper right), and this difference was significant (Mann–Whitney U?=?5914.5, n?=?200,
P?=?0.025).
Table 3. Effect of punishment on loss of other functions
One possible explanation is that the higher performance of EQU that evolved in the
+EQU/-AND environment (relative to the +EQU environment) did not translate into higher
overall fitness. For example, it may have caused a correlated increase in replication
rate, such that the two outcomes—although different—represented equally good evolutionary
outcomes. To see if these organisms did indeed have equivalent fitness, we took the
ones evolved in the +EQU/-AND environment and transplanted them to the +EQU environment,
where we assayed their fitness. Surprisingly, we found that they were also significantly
more fit in that environment than the organisms that had evolved there (Mann–Whitney
U?=?6220, n?=?200, P?=?0.003). A similar pattern, where a population reaches higher fitness in one environment
through evolution in some alternative environment, has been observed in laboratory-evolved
populations of bacteria 41],42] – a phenomenon dubbed “roundabout selection†42],43]. Figure 4 demonstrates the occurrence of roundabout selection in exemplary populations. Early
in the runs (5,000 updates, at left), organisms evolving in the +EQU/-AND environment
had much lower fitness in the +EQU environment than those that were evolving in the
+EQU environment. However, by the end of the experiment (100,000 updates, at right),
the same +EQU/-AND evolved population had achieved higher fitness in the +EQU environment
than the organisms that evolved in that environment.
Figure 4. Snapshot of phenotypic variation present in two populations at two time points. Each
point represents an individual organism; the height of the point shows its fitness
in the +EQU environment, while the other two axes show the number of times it performs
the EQU and AND functions. Blue: organisms from a population evolving in the +EQU
environment. Red: organisms from a population evolving in the +EQU/-AND environment.
Left: organisms that evolved in the +EQU/-AND environment (red) had lower fitness
after 5,000 updates, when they were measured in the +EQU environment, compared to
individuals that evolved in the +EQU environment (blue). Right: By the end of the
experiment (100,000 updates), organisms that evolved in the +EQU/-AND environment
(red) had attained higher fitness in the +EQU environment than those that evolved
in the +EQU environment (blue).
The finding that the +EQU/-AND populations, which started from the same ancestor but
evolved in a different environment, had significantly higher fitness in the +EQU environment
than the populations that evolved there strongly suggests the existence of multiple
adaptive peaks. More precisely, it demonstrates that higher fitness is possible in
the +EQU environment, and thus something must have prevented the +EQU-evolved populations
from reaching that higher fitness. Nevertheless, although both EQU performance and
fitness differed on average between the populations evolved under the two treatments,
there was substantial variation among populations within each treatment. Thus, the
results do not mean that populations in the +EQU environment always failed to reach
the higher fitness peak, but only that they did not reach it as often as the populations
that evolved in the +EQU/-AND environment.
It is unclear what prevents the +EQU-evolved populations from reaching higher fitness.
We see two related possibilities. First, the punishment for performing the AND function
in the +EQU/-AND environment might alter the adaptive landscape in such a way that
what was previously an adaptive valley becomes flat or even uphill from a now-sunken
starting point (Figure 5A). In this case, the +EQU/-AND populations could have moved into areas of genotypic
space that would have been less accessible in the +EQU environment owing to the intervening
valley. When examined back in the +EQU environment, these +EQU/-AND populations would
occupy a higher peak (Figure 5A). This process, whereby selection in a fluctuating environment permits populations
to attain a higher fitness than would otherwise be possible, was described by Wright
as “mass selection under changing conditions†44],45].
Figure 5. Two hypotheses about how ruggedness of the adaptive landscape might yield the observed
results. Schematic shows two hypotheses that could explain the counterintuitive result
in which populations that evolved in the +EQU/-AND environment reached higher fitness
in the +EQU environment than did those that evolved in the +EQU environment. (A) Populations evolved to the lower fitness peak in the +EQU environment (upper left)
because an adaptive valley prevented them from finding the higher peak. However, that
valley did not exist in the +EQU/-AND environment (lower left), allowing selection
to drive the populations up the single peak. When organisms that evolved in the +EQU/-AND
environment were assayed for their fitness in the +EQU environment, they were on the
higher peak. (B) The ancestor was between two adaptive peaks, either of which could be reached from
that point. However, evolution in the +EQU environment (upper right) predisposed populations
toward the lower peak because its initial ascent was steeper or mutations were more
likely to move the populations into its domain of attraction. By contrast, evolution
in the +EQU/-AND environment (lower right) predisposed populations to evolve toward
the higher peak.
A second possibility is that the ancestor sits at some distance from both peaks, and
that either peak can potentially be reached from that starting point without crossing
a valley (Figure 5B). However, in the +EQU environment, selection preferentially moves populations toward
the lower peak, either because the initial ascent is steeper or because there are
more paths leading to this peak than to the other. For example, progress toward the
higher peak might require traversing a single narrow ridge, whereas there are many
paths that lead to the lower peak. The key distinction between this hypothesis and
the previous one is that there need not be an intervening adaptive valley that prevents
the populations from finding the higher peak. Rather, the evolutionary trajectory
might depend on the likelihood of populations stumbling upon the rare genetic variants
that permit them to travel along the narrow ridge to the higher peak. Moreover, selection
against AND in the +EQU/-AND environment would favor the loss of AND, perhaps increasing
the likelihood of discovering the trajectory to this higher peak in the +EQU/-AND
environment.
These hypotheses are not mutually exclusive; the correct explanation for the difference
in the evolutionary trajectories might involve a mix of these and other processes.
Nevertheless, support for the first hypothesis would involve showing that the trajectories
of populations in the +EQU/-AND environment involved genotypic intermediates that
would have been deleterious had they arisen in the +EQU environment, but which were
neutral or even beneficial in the environment where they arose. Evidence for the second
hypothesis would require finding that evolution in the +EQU/-AND environment often
involved the substitution of the same few mutations in the replicate populations,
which would support the idea that there was a paucity of paths that lead to the higher
peak. Below, we present evidence that bears on these two possibilities.
Fitness effects of mutations that resulted in the loss of AND
The only difference between the +EQU and +EQU/-AND environments was that performing
AND reduced fitness in the latter. Thus, a good way to find mutations with differential
fitness effects in the two environments would be to identify the mutations that caused
the loss of AND in the +EQU/-AND environment. Forty-seven of the 100 populations that
evolved in this environment lost AND (Table 3). This fact suggests that the adaptive-valley hypothesis is unlikely to explain fully
the results, unless evolution in the +EQU/-AND environment reduced, but did not eliminate,
the valley. Otherwise, we would expect the loss of AND in the +EQU/-AND environment
to have occurred more often than it did or even invariably, as we saw in the five
other cases we analyzed (Table 3, far right column). In any case, we determined the fitness effects of the mutational
steps that caused the loss of AND in these 47 populations. In three populations, organisms
along the line of descent lost AND, regained it, and then lost it again. For these
three cases, we count only the first loss. In one other population, the genotype that
first lost the AND function was not assigned a fitness by the test CPU (see Methods),
so we excluded this case from our analysis. In the remaining 46 populations, the mutations
that caused the loss of AND were universally beneficial in the +EQU/-AND environment
where they arose. By contrast, 32 of these 46 mutational steps were deleterious when
assayed in the +EQU environment, a further nine were neutral, and only five were beneficial.
(The results are similar if we include instead the second, final loss of AND for the
three populations that lost AND twice; the only difference is that one beneficial
step is shifted into the neutral category.) Taken together, these data indicate that
the AND function was often retained in the +EQU environment, at least in part, because
its loss was usually caused by deleterious mutations. Changing the environment by
imposing selection against AND thus opened up certain evolutionary paths that were
not otherwise adaptive. This result provides clear support for our first hypothesis—that
changing the environment altered the adaptive landscape in such a way that it allowed
populations to evolve into regions that were otherwise inaccessible, enabling them
to approach a peak of higher fitness.
The number of paths leading to the loss of AND
Our second hypothesis considered the relative likelihood of reaching one peak versus
another owing to limitations on the production of relevant variation. To address it,
we again focused on the mutations that caused the loss of the AND function. We examined
the line of descent in the 46 populations that lost AND when it was selected against.
In each case, we identified the particular genotype on this line that first lost the
function. Figure 6 shows the alignment of the genome sequences of these genotypes; the mutations that
distinguish each of these genotypes from its immediate parent are highlighted.
Figure 6. Genome alignment showing mutations associated with loss of the AND function. Aligned
genome sequences of genotypes on the lines of descent in 46 populations, showing the
mutations that were associated with the loss of the ability to perform the AND function.
The immediate parent of each genotype could perform AND; thus, one or more of the
mutations (highlighted in yellow) caused the loss of that ability. Each letter denotes
a particular instruction; a red letter indicates a point mutation, whereas green letters
and blue asterisks are insertion and deletion mutations, respectively. In three cases
(bold run ID), the AND function was lost and regained multiple times, and only the
first loss is shown. Several sequences have been trimmed owing to their length, and
they show only the relevant portions of the genomes.
Several patterns are immediately apparent. First, either one or two of the same four
sites changed in 44 of the 46 replicate populations. In many instances, replicate
populations even converged on the exact same substitution. Second, and even more surprisingly,
in 27 (59%) cases, the loss of AND was associated with a double mutation. Given the
fact that the genomic mutation rate was 0.1 in these experiments, there should be,
on average, 0.1 mutations per genome per generation. Because a genotype on the line
of descent necessarily differs from its parent by at least one mutation, we calculated
the probability that a particular genotype would differ from its parent genotype by
two or more mutations. From the Poisson distribution, only about 4.9% of the genotypes
with at least one mutation are expected to have two or more mutations. Thus, the mutations
that caused the loss of the AND function involved double mutations almost 12 times
more often than expected by chance. In fact, two pairs of sites were involved in 15
of the 27 double mutations (first 15 rows in Figure 6). These results argue strongly that the number of ways of generating this beneficial
phenotype – which required eliminating the AND function while retaining the EQU function
and all other aspects of organismal performance – were very limited. Importantly,
the genotype-phenotype map (Figure 1) shows that there were many ways to knock out the AND function; therefore, it must
have been the pleiotropic effects of most such mutations on EQU or other fitness components
that placed such severe constraints on the mutations that could be substituted in
the +EQU/-AND environment.
Finally, the overrepresentation of double mutations among those genotypes that lost
the AND function is interesting its own right, because it implies that the component
mutations were not beneficial when they arose individually in the +EQU/-AND environment.
It supports the presence of an adaptive valley in a relevant part of the genotypic
space in the +EQU/-AND environment, albeit a narrow valley that could be traversed
by double mutations. This result may also explain why only some of the populations
that evolved in this environment lost the AND function, despite selection for that
loss. Although punishing the performance of AND caused its loss far more often than
when it was not punished (Table 3), the mutations necessary to produce this loss while also retaining other critical
functions were evidently so infrequent that many populations failed to lose it. Thus,
these results demonstrate that constraints on the types of variation available in
these populations limited their ability to reach the alternative, higher adaptive
peak.