Editor's evaluation: Artificial selection methods from evolutionary computing show promise for directed evolution of microbes

Abstract

Article Figures and data Abstract Editor's evaluation eLife digest Introduction Methods Results and discussion Data availability References Decision letter Author response Article and author information Metrics Abstract Directed microbial evolution harnesses evolutionary processes in the laboratory to construct microorganisms with enhanced or novel functional traits. Attempting to direct evolutionary processes for applied goals is fundamental to evolutionary computation, which harnesses the principles of Darwinian evolution as a general-purpose search engine for solutions to challenging computational problems. Despite their overlapping approaches, artificial selection methods from evolutionary computing are not commonly applied to living systems in the laboratory. In this work, we ask whether parent selection algorithms—procedures for choosing promising progenitors—from evolutionary computation might be useful for directing the evolution of microbial populations when selecting for multiple functional traits. To do so, we introduce an agent-based model of directed microbial evolution, which we used to evaluate how well three selection algorithms from evolutionary computing (tournament selection, lexicase selection, and non-dominated elite selection) performed relative to methods commonly used in the laboratory (elite and top 10% selection). We found that multiobjective selection techniques from evolutionary computing (lexicase and non-dominated elite) generally outperformed the commonly used directed evolution approaches when selecting for multiple traits of interest. Our results motivate ongoing work transferring these multiobjective selection procedures into the laboratory and a continued evaluation of more sophisticated artificial selection methods. Editor's evaluation The study offers a valuable contribution to the field. While the fields of artificial life and experimental evolution in microbes have been connected for many years, there have been few studies to meaningfully demonstrate how work in evolutionary computation can meaningfully inform the design and execution of microbial experiments. This study represents a truly innovative approach and may fuel further studies at the intersection between computational evolution and experimental evolution. https://doi.org/10.7554/eLife.79665.sa0 Decision letter Reviews on Sciety eLife's review process eLife digest Humans have long known how to co-opt evolutionary processes for their own benefit. Carefully choosing which individuals to breed so that beneficial traits would take hold, they have domesticated dogs, wheat, cows and many other species to fulfil their needs. Biologists have recently refined these ‘artificial selection’ approaches to focus on microorganisms. The hope is to obtain microbes equipped with desirable features, such as the ability to degrade plastic or to produce valuable molecules. However, existing ways of using artificial selection on microbes are limited and sometimes not effective. Computer scientists have also harnessed evolutionary principles for their own purposes, developing highly effective artificial selection protocols that are used to find solutions to challenging computational problems. Yet because of limited communication between the two fields, sophisticated selection protocols honed over decades in evolutionary computing have yet to be evaluated for use in biological populations. In their work, Lalejini et al. compared popular artificial selection protocols developed for either evolutionary computing or work with microorganisms. Two computing selection methods showed promise for improving directed evolution in the laboratory. Crucially, these selection protocols differed from conventionally used methods by selecting for both diversity and performance, rather than performance alone. These promising approaches are now being tested in the laboratory, with potentially far-reaching benefits for medical, biotech, and agricultural applications. While evolutionary computing owes its origins to our understanding of biological processes, it has much to offer in return to help us harness those same mechanisms. The results by Lalejini et al. help to bridge the gap between computational and biological communities who could both benefit from increased collaboration. Introduction Directed evolution harnesses laboratory artificial selection to generate biomolecules or organisms with desirable functional traits (Arnold, 1998; Sánchez et al., 2021). The scale and specificity of artificial selection has been revolutionized by a deeper understanding of evolutionary and molecular biology in combination with technological innovations in sequencing, data processing, laboratory techniques, and culturing devices. These advances have cultivated growing interest in directing the evolution of whole microbial communities with functions that can be harnessed in medical, biotech, and agricultural domains (Sánchez et al., 2021). Attempting to direct evolutionary processes for applied goals has not been limited to biological systems. The field of evolutionary computing harnesses the principles of Darwinian evolution as a general-purpose search engine to find solutions to challenging computational and engineering problems (Fogel, 2000). As in evolutionary computing, directed evolution in the laboratory begins with a library—or population—of variants (e.g., communities, genomes, or molecules). Variants are scored based on a phenotypic trait (or set of traits) of interest, and the variants with the ‘best’ traits are chosen to produce the next generation. Such approaches to picking progenitors are known as elitist selection algorithms in evolutionary computing (Baeck et al., 1997). Notably, evolutionary computing research has shown that these elitist approaches to artificial selection can be suboptimal in complex search spaces. On their own, elitist selection schemes fail to maintain diversity, which can lead to populations becoming trapped on suboptimal regions of the search space because of a loss of variation for selection to act on (Lehman and Stanley, 2011a; Hernandez et al., 2022b). Elitist selection schemes also inherently lack mechanisms to balance selection across multiple objectives. These observations suggest that other approaches to selection may improve directed microbial evolution outcomes. Fortunately, artificial selection methods (i.e., parent selection algorithms or selection schemes) are intensely studied in evolutionary computing, and many in silico selection techniques have been developed that improve the quality and diversity of evolved solutions (e.g., Spector, 2012; Mouret and Clune, 2015; Hornby, 2006; Goldberg and Richardson, 1987; Goings et al., 2012; Lehman and Stanley, 2011b). Given their success, we expect that artificial selection methods developed for evolutionary computing will improve the efficacy of directed microbial evolution in the laboratory, especially when simultaneously selecting for more than one trait (a common goal in evolutionary computation). Such techniques may also be useful in the laboratory to simultaneously select for multiple functions of interest, different physical and growth characteristics, robustness to perturbations, or the ability to grow in a range of environments. Directed microbial evolution, however, differs from evolutionary computing in ways that may inhibit our ability to predict which techniques are most appropriate for the laboratory. For example, candidate solutions (i.e., individuals) in evolutionary computing are evaluated one by one, resulting in high-resolution genotypic and phenotypic information that can be used for selecting parents, which are then copied, recombined, and mutated to produce offspring. In directed microbial evolution, individual-level evaluation is generally intractable at scales required for de novo evolution; as such, evaluation often occurs at the population level, and the highest performing populations are partitioned (instead of copied) to create ‘offspring’ populations. Moreover, when traits of interest do not benefit individuals’ reproductive success, population-level artificial selection may conflict with individual-level selection, which increases the difficulty of steering evolution. Here, we ask whether artificial selection techniques developed for evolutionary computing might be useful for directing the evolution of microbial populations when selecting for multiple traits of interest. We examine selection both for enhancing multiple traits in a single microbial strain and for producing a set of diverse strains that each specialize in different traits. To do so, we developed an agent-based model of directed evolution wherein we evolve populations of self-replicating computer programs performing computational tasks that contribute either to the phenotype of the individual or the phenotype of the population. Using our model, we evaluated how well three selection techniques from evolutionary computing (tournament, lexicase, and non-dominated elite selection) performed in a setting that mimics directed evolution on functions measurable at the population level. Tournament selection chooses progenitors by selecting the most performant candidates in each of a series of randomly formed ‘tournaments.’ Both lexicase and non-dominated elite selection focus on propagating a diverse set of candidates that balance multiple objectives in different ways. These selection techniques are described in detail in the ‘Methods’ section. Overall, we found that multiobjective selection techniques (lexicase and non-dominated elite selection) generally outperformed the selection schemes commonly applied to directed microbial evolution (elite and top 10%). In particular, our findings suggest that lexicase selection is a good candidate technique to translate into the laboratory, especially when aiming to evolve a diverse set of specialist microbial populations. Additionally, we found that population-level artificial selection can improve directed evolution outcomes even when traits of interest are directly selected (i.e., the traits are correlated with individual-level reproductive success). These findings lay the foundation for strengthened communication between the evolutionary computing and directed evolution communities. The evolution of biological organisms (both natural and artificial) inspired the origination of evolutionary computation, and insights from evolutionary biology are regularly applied to evolutionary computing. As evolutionary computation has immense potential as a system for studying how to control laboratory evolution, these communities are positioned to form a virtuous cycle where insights from evolutionary computing are then applied back to directing the evolution of biological organisms. With this work, we seek to strengthen this feedback loop. Directed evolution Humans have harnessed evolution for millennia, applying artificial selection (knowingly and unknowingly) to domesticate a variety of animals, plants, and microorganisms (Hill and Caballero, 1992; Cobb et al., 2013; Driscoll et al., 2009; Libkind et al., 2011). More recently, a deeper understanding of evolution, genetics, and molecular biology in combination with technological advances has extended the use of artificial selection beyond domestication and conventional selective breeding. For example, artificial selection has been applied to biomolecules (Beaudry and Joyce, 1992; Chen and Arnold, 1993; Esvelt et al., 2011), genetic circuits (Yokobayashi et al., 2002), microorganisms (Ratcliff et al., 2012), viruses (Burrowes et al., 2019; Maheshri et al., 2006), and whole microbial communities (Goodnight, 1990; Swenson et al., 2000; Sánchez et al., 2021). In this work, we focus on directed microbial evolution. One approach to artificial selection is to configure organisms’ environment such that desirable traits are linked to growth or survival (referred to as ‘selection-based methods’; Wang et al., 2021). In some sense, these selection-based methods passively harness artificial selection as individuals with novel or enhanced functions of interest will tend to outcompete other conspecifics without requiring intervention beyond initial environmental manipulations. In combination with continuous culture devices, this approach to directing evolution can achieve high-throughput microbial directed evolution, ‘automatically’ evaluating many variants without manual analysis (Wang et al., 2021; Toprak et al., 2012; DeBenedictis et al., 2021). For example, to study mechanisms of antibiotic resistance, researchers have employed morbidostats that continuously monitor the growth of evolving microbial populations and dynamically adjust antibiotic concentrations to maintain constant selection on further resistance (Toprak et al., 2012). However, linking desirable traits to organism survival can be challenging, requiring substantial knowledge about the organisms and the functions of interest. Similar to conventional evolutionary algorithms, ‘screening-based methods’ of directed evolution assess each variant individually and choose the most promising to propagate (Wang et al., 2021). Overall, screening-based methods are more versatile than selection-based methods because traits that are desirable can be directly discerned. However, screening requires more manual intervention and thus limits throughput. In addition to their generality, screening-based methods also allow practitioners to more easily balance the relative importance of multiple objectives. For example, plant breeders might simultaneously balance screening for yield, seed size, drought tolerance, etc. (Cooper et al., 2014; Bruce et al., 2019). In this work, we investigate screening-based methods of directed microbial evolution as many insights and techniques from evolutionary computation are directly applicable. When directing microbial evolution, screening is applied at the population (or community) level (Xie and Shou, 2021; Sánchez et al., 2021). During each cycle of directed microbial evolution, newly founded populations grow over a maturation period in which members of each population reproduce, mutate, and evolve. Next, populations are assessed, and promising populations are chosen as ‘parental populations’ that will be partitioned into the next generation of ‘offspring populations.’ Screening-based artificial selection methods are analogous to parent selection algorithms or selection schemes in evolutionary computing. Evolutionary computing research has shown that the most effective selection scheme depends on a range of factors, including the number of objectives (e.g., single- versus multiobjective), the form and complexity of the search space (e.g., smooth versus rugged), and the practitioner’s goal (e.g., generating a single solution versus a suite of different solutions). Conventionally, however, screening-based methods of directing microbial evolution choose the overall ‘best’-performing populations to propagate (e.g., the single best population or the top 10%; Xie et al., 2019). To the best of our knowledge, the more sophisticated methods of choosing progenitors from evolutionary computing have not been applied to directed evolution of microbes. However, artificial selection techniques from evolutionary computing have been applied in a range of other biological applications. For example, multiobjective evolutionary algorithms have been applied to DNA sequence design (Shin et al., 2005; Chaves-González, 2015); however, these applications are treated as computational optimization problems. A range of selection schemes from evolutionary computing have also been proposed for both biomolecule engineering (Currin et al., 2015; Handl et al., 2007) and agricultural selective breeding (especially for scenarios where genetic data can be exploited) (Ramasubramanian and Beavis, 2021). For example, using an NK landscape model, O’Hagan et al. evaluated the potential of elite selection, tournament selection, fitness sharing, and two rule-based learning selection schemes for selective breeding applications (O’Hagan et al., 2012). Inspired by genetic algorithms, island model approaches (Tanese, 1989) have been proposed for improving plant and animal breeding programs (Ramasubramanian and Beavis, 2021; Yabe et al., 2016), and Akdemir et al., 2019 applied multiobjective selection algorithms like non-dominated selection to plant and animal breeding. In each of these applications, however, artificial selection acted as screens on individuals and not whole populations; therefore, our work focuses on screening at the population level in order to test the applicability of evolutionary computing selection algorithms as general-purpose screening methods for directed microbial evolution. Methods Conducting directed evolution experiments in the laboratory can be slow and labor intensive, making it difficult to evaluate and tune new approaches to artificial selection in vitro. We could draw directly from evolutionary computing results when transferring techniques into the laboratory, but the extent to which these results would predict the efficacy (or appropriate parameterization) of a given algorithm in a laboratory setting is unclear. To fill this gap, we developed an agent-based model of directed evolution of microbes for evaluating which techniques from evolutionary computing might be most applicable in the laboratory. Using our model of laboratory directed evolution, we investigated whether selection schemes from evolutionary computing might be useful for directed evolution of microbes. Specifically, we compared two selection schemes used in directed evolution (elite and top 10% selection) with three other methods used in evolutionary computing (tournament, lexicase, and non-dominated elite selection). Additionally, we ran two controls that ignored population-level performance. We conducted three independent experiments. First, we evaluated the relative performance of parent selection algorithms in a conventional evolutionary computing context, which established baseline expectations for subsequent experiments using our model of laboratory directed evolution. Next, we compared parent selection algorithms using our model of laboratory directed evolution in two contexts. In the first context, we did not link population-level functions (Table 1) to organism survival to evaluate how well each parent selection algorithm performs as a screening-based method of artificial selection. In the second context, we tested whether any of the selection schemes still improve overall directed evolution outcomes even when organism survival is aligned with population-level functions. Table 1 Computational functions that conferred individual-level or population-level benefits. The particular functions were chosen to be used in our model based on those used in the Avida system (Bryson et al., 2021). In all experiments, we included two versions of ECHO (each for different input values), resulting in 22 possible functions that organisms could perform. In general, functions that confer population-level benefits are more complex (i.e., require more instructions to perform) than functions designated to confer individual-level benefits. Function# InputsBenefitECHO1IndividualNAND2IndividualNOT1PopulationORNOT2PopulationAND2PopulationOR2PopulationANDNOT2PopulationNOR2PopulationXOR2PopulationEQU2Population2A1IndividualA21PopulationA31PopulationA+B2PopulationA×B2PopulationA−B2PopulationA2+B22PopulationA3+B32PopulationA2−B22PopulationA3−B32PopulationA+B22Population Digital directed evolution Figure 1 overviews our model of laboratory directed microbial evolution. Our model contains a set of populations (i.e., a ‘metapopulation’). Each population comprises digital organisms (self-replicating computer programs) that compete for space in a well-mixed virtual environment. Both the digital organisms and their virtual environment are inspired by those of the Avida Digital Evolution Platform (Ofria et al., 2009), which is a well-established study system for in silico evolution experiments (e.g., Lenski et al., 1999; Lenski et al., 2003; Zaman et al., 2014; Lalejini et al., 2021) and is a closer analog to microbial evolution than conventional evolutionary computing systems. However, we note that our model’s implementation is fully independent of Avida, as the Avida software platform does not allow us to model laboratory setups of directed microbial evolution (as described in the previous section). Figure 1 Download asset Open asset Overview of our model of directed microbial evolution. In (a), we found each of N populations with a single digital organism. In this figure, the metapopulation comprises three populations. Next (b), each population undergoes a maturation period during which digital organisms compete for space, reproduce, mutate, and evolve. After maturation, (c) we evaluate each population based on one or more population-level characteristics, and we select populations (repeat selections allowed) to partition into N ‘offspring’ populations. In this figure, we show populations being evaluated on three objectives (o1, o2, and o3). In this work, population-level objectives include the ability to compute different mathematical expressions (see Table 1). We see this as analogous to a microbial population’s ability to produce different biomolecules or to metabolize different resources. After evaluation, populations are chosen algorithmically using one of the selection protocols described in ‘Methods.’ In our model, we seed each population with a digital organism (explained in more detail below) capable only of self-replication (Figure 1a). After initialization, directed evolution proceeds in cycles. During a cycle, we allow all populations to evolve for a fixed number of time steps (i.e., a ‘maturation period’; Figure 1b). During a population’s maturation period, digital organisms execute the computer code in their genomes, which encodes the organism’s ability to self-replicate and perform computational tasks using inputs from its environment. When an organism reproduces, its offspring is subject to mutation, which may affect its phenotype. Therefore, each population in the metapopulation independently evolves during the maturation period. After the maturation period, we evaluate each population’s performance on a set of objectives and apply an artificial selection protocol to algorithmically choose performant populations to propagate (Figure 1c). In this work, we simulate a serial batch culture protocol. To create an offspring population (Figure 1d), we use a random sample of digital organisms from the chosen parental population (here we used 1% of the maximum population size). We chose this sample size based on preliminary experiments, wherein we found that smaller sample sizes performed better than larger sizes (see supplemental material, Lalejini et al., 2022). Digital organisms Each digital organism contains a sequence of program instructions (its genome) and a set of virtual hardware components used to interpret and express those instructions. The virtual hardware and genetic representation used in this work extends that of Dolson et al., 2019; Hernandez et al., 2022a. The virtual hardware includes the following components: an instruction pointer indicating the position in the genome currently being executed, 16 registers for performing computations, 16 memory stacks, input and output buffers, ‘scopes’ that facilitate modular code execution, and machinery to facilitate self-copying. For brevity, we refer readers to supplemental material for a more detailed description of these virtual hardware components (Lalejini et al., 2022). Digital organisms express their genomes sequentially unless the execution of one instruction changes which instruction should be executed next (e.g., ‘if’ instructions). The instruction set is Turing complete and syntactically robust such that any ordering of instructions is valid (though not necessarily useful). The instruction set includes operators for basic math, flow control (e.g., conditional logic and looping), designating and triggering code modules, input, output, and self-replication. Each instruction contains three arguments, which may modify the effect of the instruction, often specifying memory locations or fixed values. We further document the instruction set in our supplemental material. Digital organisms reproduce asexually by executing copy instructions to replicate their genome one instruction at a time and then finally issuing a divide command. However, copying is subject to errors, including single-instruction and single-argument substitution mutations. Each time an organism executes a copy, there is a 1% chance that a random instruction is copied instead, introducing a mutation (and a 0.5% chance to incorrectly copy each instruction argument). Mutations can change the offspring’s phenotype, including its replication efficiency and computational task repertoire. Genomes were fixed at a length of 100 instructions. When an organism replicates, its offspring is placed in a random position within the same population, replacing any previous occupant. We limited the maximum population size to 1000 organisms. Because space is a limiting resource, organisms that replicate quickly have a selective advantage within populations. During evolution, organism replication can be improved in two ways: by improving computational efficiency or by increasing the rate of genome execution (‘metabolic rate’). An organism’s metabolic rate determines the average number of instructions an organism is able to execute in a single time step. Digital organisms can improve their metabolic rate by evolving the ability to perform designated functions (referred to as individual-level functions), including some Boolean logic functions and simple mathematical expressions (Table 1). Performing a function requires the coordinated execution of multiple genetically encoded instructions, including ones that interact with the environment, store intermediate computations, and output the results. For example, the A+B function requires an organism to execute the input instruction twice to load two numeric inputs into its memory registers, execute an add instruction to sum those two inputs and store the result, and then execute an output instruction to output the result. When an organism produces output, we check to see whether the output completes any of the designated functions (given previous inputs it received); if so, the organism’s metabolic rate is adjusted accordingly. Organisms are assigned a random set of numeric inputs at birth that determine the set of values accessible via the input instruction. We guarantee that the set of inputs received by an organism result in a unique output for each designated function. Organisms benefit from performing each function only once, preventing multiple rewards for repeating a single-function result. In this work, we configured each function that confers an individual-level benefit to double an organism’s metabolic rate, which doubles the rate the organism can copy itself. For a more in-depth overview of digital organisms in a biological context, see Wilke and Adami, 2002. Population-level evaluation In addition to individual-level functions, organisms can perform 18 different population-level functions (Table 1). Unless stated otherwise, performing a population-level function does not improve an organism’s metabolic rate. Instead, population-level functions are used for population-level evaluation and selection, just as we might screen for the production of different by-products in laboratory populations. We assigned each population a score for each population-level function based on the number of organisms that performed that function during the population’s maturation period. The use of these scores for selecting progenitors varied by selection scheme (as described in the ‘Methods section). While population-level functions benefit a population’s chance to propagate, they do not benefit an individual organism’s immediate reproductive success: time spent computing population-level functions is time not spent on performing individual-level functions or self-replicating. Such conflicts between group-level and individual-level fitness are well-established in evolving systems (Simon et al., 2013; Waibel et al., 2009) and are indeed a problem recognized for screening-based methods of artificial selection that must be applied at the population level (Escalante et al., 2015; Brenner et al., 2008). Selection schemes Elite and top 10% selection Elite and top 10% selection are special cases of truncation selection (Mühlenbein and Schlierkamp-Voosen, 1993) or (μ,λ) evolutionary strategies (Bäck et al., 1991) wherein candidates are ranked and the most performant are chosen as progenitors. We implement these selection methods as they are often used in laboratory directed evolution (Xie et al., 2019; Xie and Shou, 2021). Here, both elite and top 10% selection rank populations according to their aggregate performance on all population-level functions. Elite selection chooses the single best-performing population to generate the next metapopulation, and top 10% chooses the best 10% (rounded up to the nearest whole number) as parental populations. Tournament selection Tournament selection is one of the most common parent selection methods in evolutionary computing. To select a parental population, T populations are randomly chosen (with replacement) from the metapopulation to form a tournament (T=4 in this work). The population with the highest aggregate performance on all population-level functions wins the tournament