Challenges and potential solutions for studying the genetic and phenotypic architecture of adaptation in microbes

Abstract

All organisms are defined by the makeup of their DNA. Over billions of years, the structure and information contained in that DNA, often referred to as genetic architecture, have been honed by a multitude of evolutionary processes. Mutations that cause genetic elements to change in a way that results in beneficial phenotypic change are more likely to survive and propagate through the population in a process known as adaptation. Recent work reveals that the genetic targets of adaptation are varied and can change with genetic background. Further, seemingly similar adaptive mutations, even within the same gene, can have diverse and unpredictable effects on phenotype. These challenges represent major obstacles in predicting adaptation and evolution. In this review, we cover these concepts in detail and identify three emerging synergistic solutions: higher-throughput evolution experiments combined with updated genotype-phenotype mapping strategies and physiological models. Our review largely focuses on recent literature in yeast, and the field seems to be on the cusp of a new era in the study and predictability of evolution.

Evolution, as a dynamic process, has proven to be hard to predict. One way populations evolve is through adaptation, or the acquisition of traits that provide an increase in fitness in a given environment. Adaptation is an inherently competitive process, as adaptive mutations in asexually reproducing microbes must survive challenges such as clonal interference, genetic drift, deleterious hitchhiking, etc. to overcome their peers. Gaining a better understanding of the genetic and phenotypic changes that allow organisms to adapt to changing environments, and the broader effects of these changes, is critical to improving the predictability of evolution. In the simplest situation, the possible adaptive genetic targets in a given context are limited or all share similar characteristics, e.g. they all fall into the same handful of genes or molecular pathways that affect similar phenotypes, and the routes to adaptation are easy to predict. This is sometimes true, for example, in scenarios with incredibly strong selective pressures such as adaptation to an essential gene knockout, to nearly fatal antibiotic drug concentrations, etc. [1-3]. However, this has not borne out to be the case in the majority of situations, those in which selection is non-lethal, and the development of reliable tools or rules to predict adaptation have eluded us.

Here, we discuss challenges and recent advances in understanding the genetic architecture of adaptation, many of which also apply more generally to understanding genotype-phenotype mapping [4-6]. Many advances in this field were made possible due to the use of higher-throughput technologies that were developed in or have been applied to the model organism budding yeast, S. cerevisiae. These higher-throughput studies have revealed three problems (Figure 1) that impede evolutionary prediction: 1) the genetic targets of adaptation are often more complex than previously realized, 2) the genetic targets of adaptation change with genetic background, and 3) seemingly similar adaptive mutants have diverse and unpredictable effects on phenotype. We discuss these three problems in the first half of this review, then we suggest possible solutions that may yet allow evolutionary predictions in the future.

Figure 1.

The problems associated with predicting adaptation. A. problem 1: the genetic targets of adaptation can be numerous and varied. Traditional methods of isolating adaptive mutants from experimental evolutions were often too low-throughput to catch lower frequency adaptive lineages, leading to the hypothesis that selection only targeted a handful of genes in a given context. B. problem 2: the genetic basis of adaptation changes with genetic background. Mutations rarely affect phenotype in a way that is independent of the existing genetic background (dark blue yeasts). Instead, often epistasis leads to unpredictable phenotypes when a new mutation interacts with its genetic context (multicolored yeasts). C. problem 3: seemingly similar adaptive mutants have diverse and unpredictable effects on phenotype. Mutants evolved in a given condition, the “home” environment, have a similar phenotype: increased fitness in the home environment. One might expect them to have an equally uniform response to a new, “non-home”, environment. However, the pleiotropic effects of mutations often make the response to novel conditions discordant and unpredictable.

Problem 1: The genetic targets of adaptation can be numerous and varied.

Some studies suggest that only a handful of genes represent potential targets of adaptation to a given stressor. In cases where this is true, it yields simpler predictions about how adaptation will proceed. For example, it is widely accepted in Saccharomyces cerevisiae and Plasmodium falciparum that resistance to the drug pyrimethamine is acquired by ordered, sequential fixation of mutations in DHFR or pfdhfr, respectively [7-9]. And indeed, P. falciparum clinical samples follow these predictions as four point mutations (N51I, C59R, S108N and I164L) in pfdhfr tend to result in failed treatment [10-12]. Unfortunately, this one drug-one gene model is the exception rather than the rule. One study investigating the genetic basis of azole resistance evolved six yeast populations in a clinically relevant concentration of fluconazole and identified four genes, CDR1, CDR2, MDR1 and ERG11 as possible targets of adaptation [13]. Further complicating matters, within a single gene, ERG11, not all mutations observed in the clinic provide resistance [14]. Furthermore, more recent studies have identified additional genes that can be involved in azole resistance, such as ERG3, TAC1, MRR1, UPC2 and PDR3, some of which provide resistance to different levels of drug than others [15,16]. It is becoming increasingly clear that the diversity of adaptive mutations is more varied than previously thought [15].

Larger and larger experiments are beginning to really highlight this problem. A recent study using pooled populations of genetically barcoded yeast enabled ~500,000 evolutionary replicates to be performed simultaneously [17]. The researchers were able to detect ~25,000 unique adaptive lineages that had improved in their ability to survive glucose limitation; this is orders of magnitude more than any previous study. While many of the lineages carried the same or similar adaptive mutations, the lineages carrying smaller effect beneficial mutations would not have been detected in studies that utilized fewer replicates and/or had less power to distinguish low frequency adaptive mutations from sequencing errors (Figure 1A). These lower frequency adaptive mutants are likely to be important, given follow up studies showing that the most adaptive lineages often possess two mutations, including one from this low frequency category [18]. Additional experimental designs using barcodes or other high replicate approaches continue to improve our ability to collect a fuller spectrum of mutations that are adaptive in a given conditions [19-22]. For example, recent work evolved S. cerevisiae to 80 different chemical compounds and identified 1405 mutations in 137 genes that provided resistance to at least one compound [20]. While several of these mutations were in known targets of adaptation, the most frequently hit genes were transcription factors, many of which had not previously been associated with drug resistance. While the focus of this review is on yeast as a model organism, we note that extensive experimental evolution data in bacteria also support the diversity of adaptive solutions [23-25]. These examples demonstrate there are many mutations available to solve an evolutionary challenge, making predictability in any given case that much harder.

Problem 2: The genetic basis of adaptation changes with genetic background.

Adding further complications, new adaptive mutations can interact with existing genetic variation in unpredictable ways (i.e., epistasis, Figure 1B) [26-28]. For example, several studies have shown that if an organism is already relatively fit, subsequent beneficial mutations will have a diminishing impact. This observation may reflect a global constraint, as fitness cannot increase linearly indefinitely– there are ceilings and floors on fitness (i.e., diminishing returns epistasis) [29-31]. However, recent studies suggest diminishing returns epistasis can also arise from idiosyncratic interactions among mutations to a small number of genes [32,33]. Other studies of epistasis have shown more generally that the impact of a mutation can differ across genetically diverse strains [26,34,35]. For example, Jerison et al. [34] evolved 230 yeast offspring that each differ by approximately 25,000 base pairs, finding that fitness effects of adaptive mutations were different in different offspring. A related complication is that the effects of subsequent mutations can depend on those that emerged in a previous round of adaptation, meaning that the outcome of adaptation becomes increasingly unpredictable with each successive mutation that fixes in the population [9,36-38]. Similarly, multiple mutations might arise in the same lineage in a short time period (relative to the generation time of the organism). Non-beneficial mutations that appear concurrent with adaptive genotypes can rise in frequency in a population in a phenomenon called hitchhiking. The stochastic appearance of these hitchhiker mutations reduces the reproducibility of evolution and thus its predictability [39-41].

Problem 3: Seemingly similar adaptive mutants have diverse and unpredictable effects on phenotype.

In addition to the unpredictable interactions between adaptive mutations and their genetic backgrounds, the phenotypic effects of adaptive mutations can also be surprisingly difficult to predict [42-44]. This lack of predictability at the phenotypic level exists despite several studies showing that the mutations that help an organism survive a particular stress tend to fall into genes with similar functions [18,45], despite suggestion that the phenotypic basis of adaptation should be less complex than the genetic basis [42,46], and despite strategies in evolutionary medicine that aim to exploit predictability at the phenotypic level [47,48]. But why do mutations that are all similarly adaptive in one “home” environment behave differently in other “non-home” environments (Figure 1C)?

Several recent studies find that mutations are often pleiotropic in that they do not necessarily affect only a single trait [42-44,49-51]. Pleiotropy can negatively impact predictions about the fitness of mutations in novel environments, if the suite of traits affected by each mutant differs. For example, Kinsler et al. investigated the number of phenotypes individual adaptive mutations can affect by measuring the fitnesses of hundreds of yeast strains adapted to a single environment in a range of non-home environments. They found that very similar seeming adaptive mutations to negative regulators of the same pathway, or even within the same gene, can affect different sets of phenotypes and thus behave dissimilarly in non-home environments [42]. This idea was reaffirmed by the experimental evolutions of Bakerlee et al. which demonstrated that lineages evolved in one condition had divergent fitness trajectories in new environments [43]. Surprisingly, some lineages are more fit in non-home environments than their home environments, which may suggest evolving in the home environment isn’t necessarily the only or even the fastest way to adapt to that condition [42,43,51]. These studies highlight that the effects of adaptive mutations in non-home environments can be difficult to predict.

Why trying to predict adaptation is a worthwhile endeavor, despite the 3 aforementioned problems:

The many factors that confound predictions of adaptation may cast doubt on the usefulness and practicality of the endeavor, but we believe the benefits of such predictions have the potential to be wide-reaching. A better understanding of how small effect mutations, interactions with genetic background, and pleiotropy all affect adaptation will shed light on the evolution of complex traits, especially since polygenic models of adaptation and complex trait architecture are becoming increasingly prevalent [52-54]. The applications of these predictions will extend beyond evolutionary biology to fields as diverse as medicine, agriculture, and conservation, for example, potentially allowing us to predict how pathogens will adapt to a drug or how organisms will be affected by climate change [44,55-58].

Strategies to make evolutionary predictions that take the complex genetic and phenotypic architecture of adaptation into account:

We propose three potential, but not exclusive, solutions to the above problems when trying to predict adaptation. The first involves leveraging recent high-throughput technologies to collect richer data.This will inform us about how the effects of adaptation change with contexts like genetic background or environment. The second involves building more complex and accurate mathematical models of the genotype-phenotype map that take phenomena like epistasis and pleiotropy into account [59]. Finally, the third involves drawing insights from cell biology that can form the basis of physiological models that explain and predict the impacts of mutation on phenotypes and fitness [60-62].

Strategy 1: Leveraging recent technologies to collect richer data:

A key difficulty in predicting adaptation remains that we have not surveyed the possible adaptive mutations deeply enough, and that we do not understand whether all of these adaptive solutions are equivalent (e.g., do all mutations behave similarly in all genetic backgrounds or environments). For example, while low replicate evolution experiments suggest mutations that resist one drug often induce ubiquitous sensitivity to the same second drug, higher replicate studies uncover multiple resistance mechanisms with different tradeoffs [63]. Thus, higher replicate evolution experiments are essential for comprehensively understanding the genetic architecture of adaptation. Formerly, performing very high replicate evolution experiments, and then surveying the behavior of diverse adative mutants in new environments and/or genetic backgrounds, presented a labor intensive and time-consuming challenge. Happily, recent technologies, many of which were developed in budding yeast, allow this sort of rich experimental design. For example, the use of genomically integrated DNA barcodes has vastly expanded the number of replicate lineages that can be evolved in the laboratory [17,37,43,64-67], and has also hastened surveys of adaptive mutant fitnesses in diverse novel (e.g., non-home) environments [42,43,68]. Such surveys of how adaptive mutations impact fitness or phenotype can also be accomplished via high-throughput single-celled methods, such as microscopy [50] or emerging ultra-high-throughput single-cell RNA sequencing techniques [69-73]. Yeast also leads the way in terms of surveying the impacts of adaptive mutations across diverse genetic backgrounds, mainly because genetically diverse strains can be easily mated to generate thousands of unique genetic recombinants [26,34,74]. In sum, many technologies are emerging in yeast that allow high-throughput genome engineering [20,75] or recombinant strain construction [26,34,74], high-replicate evolution [17], and high throughput phenotyping and fitness measurement of adaptive mutants [42,43,68,69]. These technologies are opening doors to long-standing questions about the repeatability of evolution, the predictability of adaptation, and the architecture of the genotype-phenotype map.

Strategy 2: Building more complex models of the genotype-phenotype map:

In order to create predictive models of adaptation, it is necessary to develop novel mathematical frameworks that can analyze the humongous datasets generated by new methodologies described under Strategy 1. One approach involves examining the extent to which the phenotypic effects of different mutants are correlated in order to predict which mutants will behave similarly in novel contexts, rather than assuming that mutations that evolved in the same home environment will always have similar behavior in non-home environments (Figure 1C) [63]. In the past, genotype-phenotype maps were usually represented by simple bipartite gene-trait maps (Figure 2A.i), which lacked the complexity to model the degree to which pleiotropic adaptive mutations affect overlapping or completely dissimilar groups of traits. Recently, several new models that allow pleiotropic genotype-phenotype maps and leverage the complex correlations among multiple phenotypes have been proposed (Figure 2A.ii) [42,44,50,76]. For example, models using a framework of correlated evolution of traits can be built to model the emergence of cross resistance or cross sensitivity to multiple drugs [44,47,76]. In addition, correlative modeling techniques such as singular-value decomposition, principal component analysis, or machine learning can be applied to datasets that measure the fitness of different adaptive mutants across many environments in order to construct fitness predictions [42,77,78]. A different framework for interpreting the massive amounts of data pertaining to adaptive mutants and their effects may involve meta-analysis comparing different approaches. For example, comparing different statistics used for studying the level of convergence among adaptive mutations helps better understand the mapping among mutations, phenotypes, and fitness during adaptation [46]. As more technologies emerge to yield more data about the genetic architecture of adaptation, more modeling approaches must follow.

Figure 2.

Illustrations of concepts described in Strategies 2 and 3. A. genotype-to-phenotype maps. i. a simple bipartite model in which a single phenotype predictably correlates with genotype. ii. a more realistic, complex model in which genotype affects many correlated phenotypes through direct and indirect actions. B. the network effects of mutations. Network models could help explain the pleiotropic and epistatic interactions of mutations. Detailing how genes interact can give us a basis for predicting mutational effects. For example, when a hub gene is hit with a mutation (blue), the effects of that mutation may extend to all of the genes with which it interacts.

Strategy 3: Drawing insights from cell biology that can form the basis of physiological models:

Building molecular and biochemical models informed by physiology also helps us understand the genetic architecture of adaptation and predict the effects of adaptive mutants in new genetic backgrounds and environments. For example, if a cell gains an adaptation to a high drug condition by overexpressing an efflux pump, we might be able to predict that the cell will also have higher fitness in conditions with other environmental toxins. Since the mechanisms of adaptation often extend beyond a single pump, we need more complicated models incorporating more knowledge of the cell. Several types of modeling approaches have been considered. Here we will discuss two: network modeling and growth law theory.

Interaction networks summarize the connections among different elements in the cell (e.g., transcriptional networks, protein-protein interactions, etc.), and may provide insights about the effects of adaptive mutations. For example, consider a transcription factor that regulates stress related genes. If it gains a mutation that affects its DNA binding dynamics to favor stronger expression, we might expect similar fitness gains under diverse stressful environments. Previous studies have suggested that adaptive mutations within the same network may improve fitness via similar phenotypic changes [18]. This intuition was supported by work showing that the beneficial effects of mutations in similar functional units were often not additive when combined, presumably because their effects are redundant [45,79]. But more recent work has revealed that even similar seeming adaptive mutations can have dissimilar fitness effects in non-home environments [42,43]. Thus, more complex network models are emerging that interrogate network structure to predict the epistatic and pleiotropic effects of mutation [59,80-82]. For example, the number of interacting partners (i.e., hubness) for a component, and the nature of these interactions (i.e., whether they are activating or repressing) (Figure 2B), may inform the amount of pleiotropy [79,83] and the amount and type of epistasis [84].

The network modeling mentioned above tends to require an exhaustive survey of genetic or metabolomic components per cell type. Opposingly, there are also attempts to use phenological models with less molecular-level details to model the physiology of microbes. One potential approach considers the “growth laws”, which model the rates of microbial growth as being set by one simple parameter: the percentage of proteome allocation to ribosomes [62,85-88]. For example, Scott et al. used a growth law model to predict how E. coli changes their growth rates when there is a ribosome-inhibiting antibiotic in the medium or ribosome-disrupting mutations in the genome [62]. Similarly, You et al. predicted the growth rates of E. coli under the disruption of signaling pathways critical for the starvation condition [88]. And Kavčič et al. predicted the growth-rate change of E. coli by the interactions between ribosome-inhibiting antibiotics and expression changes of translation-regulating genes [89]. These examples highlight the potential usage of growth-law models to understand and predict interactions among adaptive mutations (posing a solution to Problem 2) and the environment-specific effects of adaptive mutations (posing a solution to Problem 3).

Conclusion:

In conclusion, evolution is hard to predict. There are many routes to adaptation, and mutations can interact with existing genetic context and/or have unforeseeable effects on many phenotypes. However, we believe the expansion of high-throughput techniques, large data sets, and new mathematical and representative models that integrate these data can begin to explore the parameter space of adaptation, and generate predictive tools that will enable us to better forecast evolutionary trajectories in the future.