Decision letter: Host-parasite coevolution promotes innovation through deformations in fitness landscapes

Abstract

Article Figures and data Abstract Editor's evaluation Introduction Results Discussion Materials and methods Appendix 1 Appendix 2 Data availability References Decision letter Author response Article and author information Metrics Abstract During the struggle for survival, populations occasionally evolve new functions that give them access to untapped ecological opportunities. Theory suggests that coevolution between species can promote the evolution of such innovations by deforming fitness landscapes in ways that open new adaptive pathways. We directly tested this idea by using high-throughput gene editing-phenotyping technology (MAGE-Seq) to measure the fitness landscape of a virus, bacteriophage λ, as it coevolved with its host, the bacterium Escherichia coli. An analysis of the empirical fitness landscape revealed mutation-by-mutation-by-host-genotype interactions that demonstrate coevolution modified the contours of λ’s landscape. Computer simulations of λ’s evolution on a static versus shifting fitness landscape showed that the changes in contours increased λ’s chances of evolving the ability to use a new host receptor. By coupling sequencing and pairwise competition experiments, we demonstrated that the first mutation λ evolved en route to the innovation would only evolve in the presence of the ancestral host, whereas later steps in λ’s evolution required the shift to a resistant host. When time-shift replays of the coevolution experiment were run where host evolution was artificially accelerated, λ did not innovate to use the new receptor. This study provides direct evidence for the role of coevolution in driving evolutionary novelty and provides a quantitative framework for predicting evolution in coevolving ecological communities. Editor's evaluation This study uses the parlance and framing of the fitness landscape to articulate a co-evolution story between host and parasite. It utilizes a tractable system, bacteriophage λ and E. coli, to ask questions that unite different pillars of evolutionary theory – evolutionary genetics (via the fitness landscape analogy), co-evolution, and host-parasite interactions. The findings will be relevant to a number of audiences, and will likely spawn downstream studies that further interrogate the molecular specifics that underlie host-parasite co-evolution. https://doi.org/10.7554/eLife.76162.sa0 Decision letter Reviews on Sciety eLife's review process Introduction A starting point for understanding how populations evolve is to assume that they exist in an unchanging world where they can adapt toward optimality (Orr, 2005; Pigliucci and Müller, 2010). However, even in static environments, populations never reach optimality because their circumstances continuously change as neighboring species coevolve with them (Valen, 1973). This more dynamic view of the evolutionary process opens the potential for unbounded evolution and creates new opportunities for evolutionary innovation (Doebeli, 2011; Thompson, 2005; Nahum et al., 2017; Thompson and Cunningham, 2002; Zaman et al., 2014). Darwin recognized this potential in the final pages of On the Origin of Species, where he wrote that, ‘It is interesting to contemplate an entangled bank’ of organisms evolving with one another to produce such a variety of forms and functions (Darwin, 1859). But he also realized the empirical challenges created by the richness of species interactions within ecological communities in his further description of ‘these elaborately constructed forms, … dependent on each other in so complex a manner…’ (Darwin, 1859). The complexity arises because an organism’s fitness is a function of its interactions with other species, and the strength and form of these interactions can continuously change as they coevolve. Furthermore, the coevolving traits of organisms are encoded within genomes by mutations that might interact with one another, a pervasive phenomenon called epistasis (Weinreich et al., 2006). This means that interactions at all levels must be considered; from mutation-by-mutation within a species (classical epistasis), to mutation-by-mutation between species (interspecific epistasis), and higher order phenomena such as the combination of classic and interspecific epistasis, where the within genome mutation-by-mutation interactions depend on the genotype of an interacting species. Many advances have been made over recent decades that enable us to tackle this combinatorial problem. Efficient genetic engineering methods permit the construction of genetic libraries with combinatorial sets of mutations that can be used to measure epistasis (Kosuri and Church, 2014; Fowler and Fields, 2014). Also available are convenient approaches to measure Darwinian fitness of the mutant libraries (Weinreich et al., 2006; Palmer et al., 2015; Khan et al., 2011; Chou et al., 2011). Coupling these two technologies allows the creation of extensive genotype-to-fitness maps, or fitness landscapes (Wright, 1932), that provide information important for predicting adaption (de Visser and Krug, 2014; de Visser et al., 2018; Lee et al., 2018). However, these maps alone are often not sufficient to predict evolution because their topographies can depend on abiotic environmental conditions (Ogbunugafor et al., 2016; Lindsey et al., 2013; Steinberg and Ostermeier, 2016; Flynn et al., 2013) and biotic interactions (Cervera et al., 2016; Fragata et al., 2019). Here, we take two significant steps forward in fitness landscape research. First, we build on the observation that landscape structures depend on species interactions by studying the interdependence of two species’ landscapes and how they shift during coevolution. Second, we test whether these shifts facilitate the evolution of a key innovation, where the species evolves a new function that unlocks new ecological opportunities. As a model system of coevolution, we studied the host-parasite interaction between bacteriophage λ and its host, Escherichia coli, because of the extensive background research completed on their coevolution and the availability of well-developed molecular tools (Meyer et al., 2012; Maddamsetti et al., 2018). When λ and E. coli are cocultured in the laboratory, one quarter of the λ populations evolve to use a new receptor (Meyer et al., 2012). λ’s native receptor is E. coli’s outer-membrane protein LamB, but through mutations in its host-recognition gene J, λ evolves to use a second receptor protein, OmpF. While only four mutations are necessary for OmpF+ function (Maddamsetti et al., 2018), more J mutations typically evolve along the way (Meyer et al., 2012). λ gains this new function after E. coli evolves resistance through malT mutations (Meyer et al., 2012) that cause reduced LamB expression (Boos and Böhm, 2000). Thus, it was hypothesized that the evolution of resistance in E. coli deformed λ’s fitness landscape in ways that promoted λ’s innovation (Thompson, 2012). In line with this, it was previously shown that four out of six λ genotypes randomly chosen on the path to evolve OmpF+ had higher relative fitness when cultured with resistant malT– cells rather than ancestral cells (Burmeister et al., 2016), suggesting that the host’s coevolution would promote key steps in λ’s evolution. However, one out of six genotypes had higher fitness in the presence of the ancestral host, and the last λ’s fitness was neutral to the host’s genotype. Given the conflicting pattern and small sample size, it could not be concluded whether coevolution was responsible for the innovation. Here, we build on this study with high-throughput technologies capable of measuring the fitness of hundreds of λ genotypes. The technology also produces combinatorial-mutation libraries of λ genotypes that can be used to quantify epistasis. This allows us to establish the contours of λ’s adaptive landscape rather than simply studying isolated genotypes within the space. Given the efficiency of this method, we can now measure λ’s landscape repeatedly in different host contexts in order to test whether host-induced deformations that naturally arise during coevolution promote OmpF+ evolution. Although it has been shown that antagonistic coevolution can hasten molecular evolution of phages (Paterson et al., 2010) and lead them to broader host ranges (Hall et al., 2011), coevolution’s role in unlocking unexplored regions of the fitness landscape has not been directly tested. Results λ’s fitness landscape at different stages of coevolution To construct λ’s fitness landscape, we focused on 10 J mutations that were a subset of mutations λ repeatedly evolved on its path to use OmpF (Meyer et al., 2012; Supplementary file 1a). Together, they form a 10-dimensional genotype space with a total of 1024 (210) unique variants of different combinations of the mutations including, the wild type (WT) allele configuration. Using Multiplexed Automated Genome Engineering (MAGE) (Wang et al., 2009), a technique that uses repeated cycles of homologous recombination in the λ-red system to produce combinatorial genomic diversity, we were successful at engineering a library of 671 genotypes out of the possible 1024 (see Materials and methods). To measure the fitness of each genotype in this library, we competed the full library en masse and monitored their frequency changes using next-generation sequencing (Figure 1—figure supplement 1; Kelsic et al., 2016; Russ et al., 2020). The fitness of each genotype was then calculated by comparing its change in frequency relative to the non-engineered ancestor. Fitness was measured in four replicate competitions for both the ancestral host and malT– host (see Materials and methods). To reduce the effect of sequencing errors and to overcome other methodological pitfalls, we modified the MAGE protocol by introducing neutral watermark mutations in the library construction and developed a high-throughput competition assay that yielded reproducible results (see Materials and methods, Figure 1—figure supplement 2, Appendix 1). Overall, we were able to measure the fitness of 580 λ genotypes cocultured with ancestral E. coli and 131 genotypes with malT– (Figure 1—figure supplement 3). The reduced number of genotypes compared to the initial library was due to a combination of factors. The randomness of the MAGE editing process caused some genotypes to be rarely constructed and to have low frequencies in the initial library. If these genotypes did not possess disproportionately high fitness, then their frequency would fall below the limit of detection during the competition, removing them from the analysis. This effect was more pronounced in the malT– landscape, where fitness differences were even more extreme. Visual inspection of the two fitness landscapes reveals host-dependent structures; the landscape with the ancestral host has a standard diminishing-returns pattern (Khan et al., 2011; Chou et al., 2011; Kryazhimskiy et al., 2014; MacLean et al., 2010; Guerrero et al., 2019), while the landscape with the malT— host has an atypical sigmoidal shape that plateaus at a higher fitness than the first (Figure 1a and b). The non-linear relationship between mutation number and fitness suggests the presence of epistasis (mutation-by-mutation interactions), the differences in the magnitude of fitness effects between landscapes suggests mutation-by-host interactions, and different shapes suggest host-dependent epistasis (mutation-by-mutation-by-host interactions). To determine how much variation in fitness is explained by these interactions, we performed multiple linear regression analyses (see Materials and methods). We found pervasive epistasis in both landscapes (Figure 1c). For the ancestral landscape, 58.66% of the variation was explained by the direct effects of mutations and 24.69% by pairwise interactions (Radj2=0.8172, F55,439 = 39.97, p<0.0001). Similarly, 48.35% of the variance in the malT— landscape was explained by the direct mutation effects and 27.61% by the interaction terms (Radj2=0.7072, F55,252 = 14.48, p<0.0001). To test for mutation-by-mutation-by-host interactions, we regressed another linear model that includes host as a predictor variable. In this model, we found significant mutation-by-host interactions (Figure 1d), and sizeable effects of the host-dependent epistasis (21 mutation-by-mutation-by-host interaction terms were significant out of 45 and 12.62% of the total variance in the data were attributable to these terms, Figure 1d). This three-way interaction term measures the extent to which the landscape structure is transformed by host evolution and suggests that λ’s evolutionary trajectory could depend on its host’s genotype. Figure 1 with 4 supplements see all Download asset Open asset Empirical fitness landscapes of λ when infecting the (a) ancestral host and (b) malT– host, and their statistical analyses in (c) and (d). Each node in (a) and (b) represents a unique genotype and two nodes are connected by edges if the corresponding genotypes are separated by one mutation. The node at zero mutations is ancestral λ. Selection rate (per 4 hr competition experiment) is the difference of Malthusian growth rates of a given genotype i to ancestral λ over 4 hr, calculated as ln⁡(λi, 4⁄λi,0)-ln⁡(λanc, 4⁄λanc,0), where λi, t denotes the density of the given genotype at time t. (c) Statistical analysis of direct and interactive effects of mutations in both the landscapes. Colored cells represent statistically significant terms determined by multiple regression analysis after correction for multiple hypothesis testing (see Materials and methods). The diagonal elements of the matrix represent single mutation effects and all the off-diagonal terms represent pairwise epistatic interactions. See Supplementary file 1i for identity of mutations corresponding to different Gi . (d) Statistical test of whether the two landscapes varied in topology. The additional variable, E, represents environment (host) to indicate mutation-by-host effects in the lower-left matrix and mutation-by-mutation-by-host (G×G×E) in the upper-right matrix. Light colored cells indicate terms present in the final AIC-optimized model out of the full-factorial model (F76,726 = 37.45, p<0.0001), and dark colored cells indicate statistically significant terms after controlling for rate of false positives (see Materials and methods). The role of shifting landscapes in promoting λ’s innovation To test whether changes in the structure of λ’s landscape opened trajectories to OmpF exploitation, we simulated λ’s evolution on the landscapes using a modified Wright-Fisher model (see Materials and methods, Figure 2—figure supplement 1). Before running the simulations, we imputed the missing λ genotypes’ fitness values to complete the landscapes. We did this by successively choosing missing genotypes at random and assigning them the average fitness of their nearest neighbors. The simulations were run based on conservative estimates of the number of generations and population sizes from previously published (Meyer et al., 2012) and this study’s coevolution experiments (960 generations; ~6.3×109 λ particles, Figure 4—figure supplement 1; Meyer et al., 2012) and λ’s intrinsic mutation rate (7.7×10–8 base–1 replication–1) (Drake, 1991). We predicted that λ would be more likely to evolve OmpF function (three specific mutations plus one additional Maddamsetti et al., 2018) in simulations that accounted for coevolution by shifting the population from one landscape to the next. We ran trials where λ evolved on only one landscape at a time to establish a baseline for the frequency of OmpF+ evolution without coevolution. Next, we ran nine shifting landscape scenarios where we varied how many generations λ evolved on the ancestral host landscape before switching to the malT– landscape. As anticipated, the switching protocol increased the frequency of OmpF+ evolution in all nine treatments above the single host simulations, but only seven out of nine treatments were found to be significantly higher (Figure 2; ANOVA: F-ratio=6.14, d.f.=99, p<0.0001, Supplementary file 1b). This result was robust to changes in population size and total number of generations, and when controlling for both, different number of genotypes measured in the two landscapes, and noise created by imputing missing data points (Figure 2—figure supplement 2, Appendix 2). Figure 2 with 2 supplements see all Download asset Open asset Simulation results of the frequency of OmpF-use evolution observed when fitness landscapes were shifted at different frequencies. Each bar represents an average of 300 simulation runs. Error bars indicate 95% confidence intervals. OmpF evolution is favored when λ evolves on shifting landscapes. The only two shifting landscape treatments that are not significantly higher than simulations on the constant malT— landscape are the 0.2 and 0.4 treatments (Supplementary file 1b). Reconstructing coevolution in an experimental population The simulation results suggest that the shifting landscape encourages λ’s evolution to gain the required mutations for OmpF function. In particular, the simulations show that the first steps along the path to the innovation are more likely if λ first adapts to the ancestral bacterium, meanwhile the final steps are more likely to occur if the host coevolves resistance. To verify this result with laboratory experiments, we analyzed the path λ took to OmpF+ in a single population cryopreserved from the previous coevolution study (population ‘D7’ in Meyer et al., 2012; Table S1). Population ‘D7’ was chosen because λ evolved relatively few mutations in this population. We believed this choice was conservative and constitutes a strong test of our hypothesis since fewer λ mutations would provide fewer opportunities to detect host-induced contingency. We sampled λ strains from different timepoints of population ‘D7’ and sequenced their J gene (Figure 3a, Supplementary file 1c and Supplementary file 1d). Next, we ran pairwise competition experiments between λ genotypes at different stages of evolution on the two hosts. We found that the first mutation on the line of descent to OmpF+ required ancestral E. coli to evolve, while the second mutation required malT– E. coli (Figure 3b and c). In addition, the eventual OmpF+ genotype with five J mutations only outcompeted the genotype with two mutations when provided with malT– hosts (Figure 3—figure supplement 1). These findings show that the path λ took in population D7 required it to sequentially adapt to both host types and that λ’s fitness landscape changed during coevolution in a way that ultimately facilitated evolutionary innovation. Figure 3 with 1 supplement see all Download asset Open asset J evolution (a) and evidence of the interdependency between λ and E. coli fitness during their coevolution (b-d). (a) Phylogenetic reconstruction and relative abundance of λ genotypes isolated through time from a previously coevolved community (Meyer et al., 2012). Each letter and star indicate a non-synonymous mutation in J (see Supplementary file 1c for labels’ corresponding mutations). A genotype’s relative abundance on a given day is denoted by the fraction of the total height of the y-axis that it occupies (e.g. on day 9, frequency of ABC is 0.2 and A**C is 0.8; see Supplementary file 1d). The lineage WT-A-ABC-ABCDE eventually evolves OmpF function and fixes in the population; resistance in E. coli through malT— rises to high frequencies between days 5 and 8 (Meyer et al., 2012). (b & c) Selection rates (per 24 hr) of phage genotypes on the two hosts. Each bar represents the mean of three experimental replicates. While mutation A is favored over wildtype (WT) λ in the presence of the ancestral host and not malT—, AB only outcompetes A in the presence of malT— and not the ancestral host. One tailed t-tests to test if the mean selection rate is significantly greater (or less) than zero: A vs WT with ancestor host- t=98.76, d.f. =2,p<0.0001; A vs WT with malT— - t=−4.99, d.f. =2,p=0.0190; AB vs A with ancestor- t=3.4, d.f. =2,p=0.0383; AB vs A with malT— - t=−8.88, d.f. =2,p=0.0062. (d) Selection rate (per 4 hr) of malT– E. coli relative to its ancestor in the presence of λ from different stages of coevolution. Each competition was replicated three times. Lowercase letters denote significance via Tukey’s honest significance test, see Supplementary file 1g for pairwise p-values (ANOVA: F-ratio=111.22, d.f.=11, p<0.0001). One tailed t-tests were also used to test if the selection rate of malT— was greater than zero WT: t=2.44, d.f. =2,p=0.676; A: t=5.12, d.f. =2,p=0.0181; ABC: t=26.59, d.f. =2,p=0.0007; A**C: t=71.67, d.f. =2,p<0.0001. This shows that malT— is unlikely to evolve in the presence of WT λ but it becomes progressively more likely as λ gains mutations. Asterisks over all the competitions indicate significance level corresponding to the p-values. Error bars in all bar graphs represent one sample SD. At this point, we shifted the focus of our study to testing whether E. coli’s resistance evolution was also impacted by λ’s evolution. Before reconstructing J evolution, it was believed that E. coli evolved resistance first and then J mutations evolved in response (Meyer et al., 2012, Figure 5). However, J evolved within a day, while malT— mutations were previously shown to fix between 5 and 8 days in the population (Meyer et al., 2012). The timing suggests that λ improved infectivity and then applied pressure on E. coli to evolve resistance. To test whether λ evolution promoted host resistance evolution, we ran competition experiments between ancestral and malT– hosts in the presence of phages isolated from four different time points. We found that malT– was not significantly more fit than the ancestral E. coli in the presence of the ancestral λ, but it was more fit in the presence of the evolved λs (Figure 3d). This provides another example of interspecific epistasis, but this time the parasite’s genotype alters the host’s landscape. This result combined with the others suggests that there is an intricate coevolutionary feedback at play between λ and E. coli: λ evolves J mutations that better exploit E. coli, which in turn applies pressure on E. coli to evolve resistance. Once resistance evolves, new adaptive pathways become available to λ that encourage the innovation. For the computer simulations, we arbitrarily chose timepoints to switch from one host to the other; however, in reality, the dynamics of the switch are dictated by the host-parasite coevolution. Evolutionary replay experiments To further test the role of coevolutionary processes at driving λ’s innovation, we ran replays of the coevolution experiment (Figure 4—figure supplement 2). We initiated 12 populations with malT– host that already possessed resistance, and 12 populations with ancestral host where λ and E. coli would coevolve normally. The former treatment should hinder the evolution of OmpF function because it denies λ the opportunity to evolve first with ancestral E. coli. In line with our expectations, 0 of 12 replicates evolved OmpF use in the malT– treatment, meanwhile 3 out of 12 evolved the innovation in the ancestral treatment (Figure 4a and b). By sequencing J alleles of the resulting λ genotypes, we found that fewer mutations evolved with malT– despite evolving for the same length of time. This suggests that λ’s evolution was stymied by starting with the resistant host (Supplementary file 1e), and by disrupting the coevolutionary process we interfered with λ’s ability to innovate. Figure 4 with 2 supplements see all Download asset Open asset Evolutionary replay experiments reveal that λ’s evolution to use OmpF depends on host coevolution. (a) Wildtype (WT) λ with ancestral host, (b) WT λ with malT–, (c) λ one mutation removed from evolving OmpF function with malT—, and (d) identical setup as (c), but with λ two mutations removed (see Supplementary file 1f for identity of mutations). (e) The bar graph provides the frequency of OmpF+ evolution compared to the frequency observed by Meyer et al., 2012. Given λ’s established one in four rates of OmpF+ with ancestral host, the probability of observing no OmpF+ evolution in 12 replicate populations is ~0.03. Thus, no positives for OmpF evolution in (b) shows that λ’s evolution to OmpF function is significantly hindered, when coevolution is initiated with malT— host. However, malT— does not impede OmpF+ evolution, when the coevolution is initiated with already evolved λs (p-values for Fisher’s exact test: between (b) and (c) p=0.0261; between (b) and (d) p=0.0122). Notably, some λ populations went extinct which is common for these experiments and was previously shown to be caused by the evolution of resistance mutations in E. coli’s ManXYZ protein complex (Meyer et al., 2012). Lastly, we tested whether malT– would still promote the evolution of OmpF use if the replay experiments were initiated with λ genotypes positioned further along the path to gaining OmpF function. We initiated two more replays: one with a λ strain that was just a single mutation away from becoming OmpF+, and another that was two mutations away (see Supplementary file 1f for J alleles). 16 of 24 λ populations evolved OmpF+ showing that whether malT– promotes OmpF+ evolution depends on where the λ genotypes are located in the fitness landscape (Figure 4c and d). While the goal of these experiments was to test a specific hypothesis about λ’s innovation, the results have broader implications for coevolutionary dynamics and the repeatability of evolution. We showed that if the host outpaces the parasite (Figure 4b; λWT and malT– E. coli), then the parasite is unable to innovate. This is in line with a previous study showing that if the host evolves even higher levels of resistance than provided by the malT mutation, λ loses the ability to innovate (Meyer et al., 2012). When specific pairs of host and parasite genotypes are combined that are at complementary steps in the coevolution, then the innovation becomes possible and even highly likely (Figure 4a, c and d; λWT and ancestral E. coli, λ–1 and malT– E. coli, and λ–2 and malT– E. coli). This suggests that the genotype of host and parasite that first encounter each other in nature, as well as the timing of their coevolution, play an important role in determining the dynamics and endpoints of their evolution. Discussion This study provides multiple direct tests of coevolution’s role in driving innovation, as well as revealing a more complicated model of λ and E. coli coevolution than previously published (Figure 5). λ was thought to evolve OmpF use as a direct response to malT– resistance (Meyer et al., 2012); however, here we learned that key steps were missing from that model. λ is initially poor at infecting its host, so it evolves mutations in J that enhance its infectivity. The new λ genotypes apply pressure on E. coli to evolve resistance. When host-resistance increases in the community, λ’s fitness landscape is deformed in a way that promotes J evolution toward OmpF use, but only if it had already acquired some J mutations. Remarkably, the timing and coordination of each of these interdependent steps is facilitated by the reciprocity ingrained in host-parasite coevolution. Altogether, we were able to provide direct experimental evidence that fluctuating landscapes, also known as fitness seascapes (Merrell, 1994), can promote evolutionary innovations. Figure 5 Download asset Open asset Previous and new version of λ-E. coli coevolutionary model. In the original model, E. coli evolves resistance by repressing λ’s receptor and then λ evolves mutations that allow it to use a new receptor. The updated model describes a more involved dynamic where λ evolves mutations that improve its ability to infect E. coli; with increased pressure to avoid infection, E. coli responds by evolving resistance, and then λ evolves the remaining mutations that leads to the ability to infect using the new receptor. Our studies show that the fitness of a parasite depends on complex genetic interactions within its own genome and with the genomes of interacting hosts. These interdependencies result in highly contingent evolution, where λ is unlikely to evolve an innovation unless it participates in a particular sequence of coevolutionary steps with its host. Despite the stochasticity that is expected to arise in systems with substantial historical contingency (Gould, 1989; Blount et al., 2018), λ’s evolution to use a new receptor is repeatable because the sequence is coordinated by coevolutionary feedback. While coevolution may yield tangled banks of interactions, we demonstrate how high-throughput technologies can be used to untangle them and to predict evolution. The ability to successfully predict evolution in any system represents a significant step forward, but it is particularly notable in conditions that incorporate species interactions. Our approach was data-intensive and relied on technologies that are not currently available for more natural and complex ecological systems; however, many efforts are underway to develop these technologies (Bergelson et al., 2021). This study shows that these efforts are worthwhile because even though the resulting data may appear to be an uninterpretable morass, computational analyses can be leveraged to penetrate the information to aid learning and prediction. Ideas and speculation This work was completed during the 2020–22 SARS-CoV-2 pandemic, raising the question of whether this research provides insight into strategies to prevent future pandemics. While it is difficult to extrapolate, there are