As described by Thomas Ray, an early pioneer of digital evolution research, “[t]he object of an [artificial life] instantiation is to introduce the natural form and process of life into an artificial medium” (Ray, 1993). Digital evolution (as we define it in this review) does not attempt to simulate biological organisms and their evolution; instead, digital evolution systems instantiate evolutionary processes in a digital medium, allowing experimenters to directly observe evolution in populations of digital organisms (Pennock, 2007).

To accomplish this instantiation, digital evolution systems are specified at the individual level (as opposed to the population level) and include the requisites for evolution by natural selection: reproduction with inheritance, competition, variation, and time. Digital evolution systems define the genomes of organisms in many different ways, but organisms are always able to reproduce (often only asexually) and pass their genomes on to their offspring. Digital organisms most often compete for limited space in the environment (e.g., with offspring replacing random organisms) or for other limited resources (e.g., CPU cycles or energy). Variation is often introduced through mutations, but it can also arise during sexual reproduction (e.g., cross-over).

Genomes in digital evolution systems vary greatly in complexity and form, ranging from relatively short computer programs (using a specialized computer language), to sequences of 0's and 1's, to single floating point numbers (Hindré et al., 2012). However, there must be a practically unlimited number of possible genomes. That is, populations must be capable of containing many different mutants at one time, and the space of possible genomes should be too large for an evolving population to enumerate during an experiment. One of the strengths of digital evolution is its ability to support and compare a variety of selection-mutation regimes, including a “weak selection, strong mutation” regime (Adami et al., 2016) where mutants arise and co-exist in the population, leading to complex ecological dynamics.

Digital evolution systems include complex genotype-to-phenotype mappings, such that a single genotype can express different phenotypes depending on chance and the endogenous and exogenous factors affecting an individual organism at a given point in time. Because there are many possible genotypes, typically one phenotype can also be achieved by many different genotypes. In systems that support symbioses, symbionts can additionally affect the phenotype of a digital organism in complex ways.

Finally, here we focus on systems in which organism fitness is implicit. That is, an organism's fitness is determined by its ability to survive and reproduce in its computational environment (as influenced by chance events, its phenotype, life history, environmental conditions, interactions with other organisms, etc.). Because an organism's fitness is dependent on its phenotype and there is a non-trivial mapping between genotype and phenotype, there is also a non-trivial relationship between an organism's genotype and its fitness.

There is a spectrum between explicit and implicit fitness, with difficult-to-categorize systems in the middle. However, this distinction excludes many evolutionary algorithms where an individual's fitness is explicitly assessed using external criteria based on its genome (e.g., a fitness or objective function) and an artificial selection algorithm determines which individuals serve as parents to contribute their genetic material to the next generation.

The EVOLVE software system is one of the earliest examples of digital evolution (Conrad and Pattee, 1970; Conrad and Strizich, 1985; Rizki and Conrad, 1985; Brewster and Conrad, 1998). EVOLVE included environmental factors such as light intensity and temperature, as well as biologically-inspired organism genomes comprising up to 40 genes, each containing up to 200 nucleic acid bases. The organisms collected resources, which they then spent to avoid starvation and reproduce. O'Callaghan and Conrad (1992) introduced symbiosis in EVOLVE III by allowing organisms to gather resources from other organisms. The main symbiotic interaction they observed was parasitic, resulting in increased population size when symbiotic interactions were enabled because of the novel niche created.

Taylor and colleagues developed the RAM digital evolution system, demonstrating its use as a model system for studying lek formation among sage grouse, mosquito control under different environmental conditions, and predator-prey dynamics (Taylor et al., 1987). In subsequent work, Taylor and colleagues used the RAM system to study the ecological dynamics of the Hydra-Chlorella system (Taylor et al., 1989). In this system, 100 Hydra host organisms were represented with individual LISP computer programs and memory. Several algal organisms (represented with their own computer programs) were associated with each host. While the system would have been capable of evolutionary dynamics (there was reproduction with the possibility of mutation and competition for space), the experiments were not run long enough to demonstrate evolutionary timescales and the populations were very small due to computational limitations.

The inspiration for many modern digital evolution systems that are still actively supported and in use can be traced directly back to the 1984 computer game “Core War” (Dewdney, 1984). In Core War, human competitors used a simplified assembly language called Redcode to write “gladiatorial” computer programs that competed for space in the simulated core memory of a computer. To win a bout of Core War, a program had to shut down all of the processes associated with its competitor programs. Many of the most successful programs engaged in self-replication. Such replicator programs repeatedly created copies of themselves, each of which also repeatedly copied themselves ad infinitum. Thus, if one copy were to be destroyed by an adversary, other copies would still persist and continue replicating. Replicators could grow exponentially in memory, rapidly outcompeting other programs. To our knowledge, none of these replicator strategies engaged in any form of symbiosis. Despite having populations of self-replicating programs and competition, evolution did not occur in Core War because replicators always created perfect copies of themselves.

Inspired by Core War, Rasmussen and colleagues created Core World, in which they introduced the possibility for random mutations when a program copied itself (Rasmussen et al., 1989, 1990). The command used by replicator programs to copy themselves was imperfect, sometimes writing a random instruction instead of copying the intended instruction. With these mutations, Core World succeeded in facilitating the evolution of populations of computer programs. The Core World system, however, proved to be ill-suited for studying evolution. Programs written in Redcode were unlikely to survive mutations, and as such, quickly accumulated deleterious mutations that often drove the populations to extinction.

Ecologist Thomas Ray's Tierra system (Ray, 1991) innovated on the design of Core World and facilitated some of the first successful evolution experiments with self-replicating computer programs. Genomes in Tierra were more evolvable than those in Core World, because mutated daughter programs were less likely to be broken. Unlike Core World, Tierra protected “living” programs from being overwritten by their competitors, requiring programs to explicitly request a protected block of memory into which they could copy themselves. When the population grew to the environment's carrying capacity, Tierra removed the oldest programs from the population to make room for new programs to be born.

In initial experiments using Tierra, Ray founded populations with an ancestral program capable only of self-replication (Ray, 1991). Competition for space dominated these early studies, resulting in a strong selection pressure for organisms to increase their replication rate. Ray observed organisms with shorter genomes, which could be copied more quickly, evolve and outcompete organisms with longer genomes, which took longer to reproduce themselves.

In Tierra, Ray unexpectedly observed the evolution of obligate parasites—programs that co-opted the copy machinery of their competitors to copy themselves. While Ray labeled these programs parasites, they did not directly harm their hosts. Instead, they competed for the same limited resource (space) as the “host” organisms, indirectly harming their host's fitness by preventing the host from being able to reproduce if the environment lacked sufficient space for the host's offspring. In this way, they could be thought of as brood parasites. In most experiments, Ray observed an evolutionary arms race between hosts and parasites. Would-be host programs evolved mechanisms for resisting parasites, and in turn, the parasites evolved to penetrate those defensive mechanisms.

Ray subsequently observed what he termed “hyper-parasites,” which parasitized the original indirect parasites. A hyper-parasite directly stole resources (CPU cycles) from its host by tricking the host into replicating the parasite's genome when the host attempted to replicate its own genome. These hyper-parasites eventually evolved to become obligate cooperators with other hyper-parasites, relying on their neighbors' copy machinery to successfully reproduce. Finally, cheaters emerged that did not contain the copy machinery, instead taking advantage of neighbors' copy machinery, and replicating more quickly as a result.

Mutualism was never observed to evolve de novo in Tierra. To test if it was possible in his system, Ray created an artificial mutualism by hand-coding two species that each contained only part of the necessary code to copy themselves and relied on one another. Such mutualists were able to persist, demonstrating that mutualism could exist in Tierra (Ray, 1992).

The richness of observed evolutionary dynamics in Tierra was initially surprising, given the simplicity of Tierra's environment. Research in digital evolution at this point in time, however, was mostly observational, mirroring the historical stage of biological research when it relied most heavily on careful observation due to the difficulty of experimental manipulation. The next stage in digital evolution research would come with Tierra's intellectual successor.

The Avida Digital Evolution Platform expanded on the design of Tierra, adding cross-feeding behavior and true parasitism, as well as a suite of sophisticated data tracking tools and the ability for researchers to configure complex environments (Adami and Brown, 1994; Cooper and Ofria, 2002; Ofria and Wilke, 2004; Ofria et al., 2009; Zaman et al., 2011). With this support, digital evolution progressed beyond observational studies of the natural history of virtual organisms to controlled experimental studies.

In Avida, self-replicating “digital organisms” compete for space on a lattice of grid spaces (Ofria et al., 2009). When an organism reproduces, its offspring is placed in a nearby space (or in a random space if the population is well-mixed), replacing any occupants of that space. As in Tierra, improvements to the speed of self-replication are advantageous during competition for space in the environment, and organisms in Avida can increase their replication rates by evolving greater replication efficiency (e.g., using a more compact encoding). Additionally, Avida introduced the concept of resources that can be “metabolized” by a digital organism to accelerate the rate at which it expresses its genome (i.e., its “metabolic rate”). Resources in Avida are associated with completing designated tasks, such as computing Boolean logic functions on inputs from the environment. Avida gives experimenters fine-grained control over how resources are configured, including their abundance (Cooper and Ofria, 2002), spatial distribution (Dolson et al., 2017), and metabolic effects (Canino-Koning et al., 2016, 2019).

Symbiosis was made possible in Avida when Zaman et al. (2011) added support for parasites. These parasites are obligate endosymbionts that steal a user-configured amount of resources (in the form of CPU cycles) from their hosts. Parasites can horizontally transmit their offspring to other hosts by copying their genome and then attempting to inject their offspring into another host. The injection is only successful, however, if the parasite performed one of the same tasks as the would-be host, similar to how bacteriophage target bacteria's outer-membrane proteins to enter a host (Kutter and Sulakvelidze, 2004).

Recently, Luo and colleagues introduced the possibility of cross-feeding into Avida (Luo et al., 2019). Organisms can create waste byproducts that can be further metabolized by other organisms. These byproducts are not spatially limited, however, and instead are accessible to all organisms in the world. This simplification therefore precludes the evolution of symbiosis via cross-feeding, because individual organisms cannot engage in close and long-term interactions with others in their vicinity.

Digital evolution research has long featured studies of diversity, including symbiosis' role in its origins and maintenance (Ray, 1992). For example, Ray (1992) evolved two populations of digital organisms, introducing a parasite to one population but not the other to determine how host-parasite coevolution influenced diversity. After the evolutionary run, Ray observed that the population without parasites contained eight “species,” whereas the population with parasites contained sixteen species. However, Ray did not further investigate why the addition of parasites doubled the final diversity sustained in the population.

Complex traits are found everywhere in nature, from the complex organs such as eyes that first concerned Darwin, to the many biochemical pathways that allow each of us to function (Goldsmith, 1990; Szostak et al., 2001). However, many questions remain surrounding the processes by which this complexity could have arisen (Simpson and Simpson, 1949; Dawkins, 1997; Szostak et al., 2001; Shanahan, 2004; Lynch, 2007; Gould, 2011). In particular, there is no inherent link between complexity and fitness: if an organism is needlessly complex, its fitness will likely suffer. The question is then: why do some niches favor organisms with more complex traits?

Methodological challenges in biological systems have severely limited experimental testing of hypotheses for how complexity could have arisen. Observational studies of historical data are unable to determine causation, and experimental evolution studies require extraordinary time and resources due to the timescales necessary for complexity to evolve (Dawkins and Krebs, 1979; Hazen et al., 2007). There also is not firm consensus on an objective way to measure the complexity of biological organismal traits, leading to rhetorical arguments that can further obscure and delay empirical research (Adami et al., 2000; Adami, 2002).

Digital evolution systems are able to overcome this difficulty by limiting possible behaviors to those that are easily quantified as more or less complex, such as the exact size of the genome needing to be replicated (Liard et al., 2020) or the number of basic operations necessary to complete a logic task (Lenski et al., 2003). Digital evolution systems are also able to evolve hundreds of thousands of generations of organisms in days or weeks while keeping a perfect genealogical record.

The most basic functionality necessary for the evolution of complexity is the ability to replicate. In the RNA World Hypothesis, however, this replication was not self-replication, but instead replication of other RNA strings (Gilbert, 1986). This form of replication raises the issue of parasitic or cheater strings, as discussed by Colizzi and Hogeweg (2016), that are copied frequently, but do not in turn replicate other strings and thus lead to the whole system breaking down. However, Colizzi and Hogeweg (2016) used a digital evolution system with RNA-inspired replicators to demonstrate that, in the presence of spatial structure, these parasitic strings may actually increase the overall replication efficiency of a population of replicators.

The digital evolution system designed by Colizzi and Hogeweg (2016) consists of a population of replicators, which cannot replicate themselves, but instead must form “complexes” with a neighboring replicator to facilitate the replication of one of the strings. Which string is actually copied depends on characteristics of the replicators involved, meaning that parasitic or cheater string types can evolve that are copied more often than other types. Note that, like Ray's original “parasites” discussed in section 5.4, these parasites are more similar to brood parasites, since they do not directly harm the host and instead their offspring compete with the host's offspring for resources and space.

With this system, Colizzi and Hogeweg (2016) experimentally determined the effect of parasites and spatial structure on the evolution of replication and complexity. They found that, in the presence of spatial structure, the presence of co-evolving parasites actually increased the replication efficiency and complexity of the “host” strings. This result was due to the parasites dying out in small patches, leading to pressure on the surviving strings to replicate quickly to take advantage of these patches.

Hickinbotham et al. (2021) further investigated the evolution and effects of parasites in a digital evolution RNA world using an automata chemistry system, Stringmol (StringMol, 2021). They found that the presence of co-evolving parasites led hosts to evolve increasingly complex strategies to differentiate themselves from parasites as well as defend against parasites.

In Stringmol, replicators are a string of computer programming instructions that are not capable of replicating themselves (StringMol, 2021). Instead, they must bind to another string and copy that string, which may be a copy of themselves. Like Tierra (Ray, 1991), Stringmol strongly favors the evolution smaller organisms (those with fewer instructions to copy) as well as replicators that are more frequently the string copied instead of the string doing the copying, which Hickinbotham and colleagues termed parasites. However, again, these “parasites” do not directly harm the host and instead their offspring compete with the host's offspring, a process reminiscent of brood parasites in the natural world. In addition, replicators randomly engage with other replicators nearby instead of engaging in a lifelong symbiotic interaction with one other individual (Hickinbotham et al., 2021). Because organisms do not move, there is high likelihood of repeated interactions between two specific organisms. Thus, the interactions in this system are on the edge of parasitic symbiosis.

Hickinbotham and colleagues found that the evolution of parasites could have several different effects on the evolution of complexity of the host replicators (Hickinbotham et al., 2021). On one extreme, the evolution of parasites frequently caused the population to go extinct (12 out of 20 populations). For the remaining eight, the population typically increased in size due to an evolved increase in replicating efficiency. In addition, host replicators evolved several strategies for combating parasites:

In addition to the effect of parasites on host evolution, Hickinbotham and colleagues also observed interesting dynamics in parasite evolution. Instead of separate parasite lineages as is often assumed, new parasites in Stringmol continually arose from mutated replicators (Hickinbotham et al., 2021). They suggest that novel parasites may flourish because the features they inherit from their immediate replicator ancestor may allow them to circumvent that replicator's defense system. When the replicator subsequently evolves effective defenses against these parasites, the parasites die out and are replaced by newly parasitic offspring of the replicator. They point out that if these observed dynamics of parasite evolution hold in the prebiotic RNA world, that could lead to significantly different outcomes than expected from co-evolving independent host and parasite species (Hickinbotham et al., 2021).

While the previous articles focused on ectosymbiotic parasites' effect in a prebiotic system, Zaman et al. (2014) focused on the effects of endosymbiotic parasites on the behavioral complexity of self-replicating hosts. They hypothesized that co-evolving symbionts could influence the evolution of host complexity in two ways. First, based on the Red Queen Hypothesis, they hypothesized that the presence of co-evolving parasites could directly select for increased complexity. The Red Queen Hypothesis posits that co-evolving species can be in an arms race that forces them both to continually evolve to compete with each other (Dawkins and Krebs, 1979).

However, there is no inherent reason that escalation must occur (Thompson, 2005a,b), leading to Zaman and colleagues' second hypothesis: “genetic memory” can lead to the evolution of more complex strategies (Zaman et al., 2014). Zaman and colleagues point out that an arms race could be cyclical, such that the species involved cycle between two or more strategies of equal complexity. This cycle is prevented if one or both species retain a “genetic memory” of strategies that counter all previously encountered partners, forcing their opponents to continually evolve novel strategies to compete. If there are a limited number of unique simple strategies in a symbiotic system, they will eventually be exhausted. Once this occurs, the populations must evolve more complex strategies if they are to continue adapting.

Using the same parasites in the Avida digital evolution system as discussed in the previous section, Zaman and colleagues leveraged the clear definition of complexity in the Avida system to determine how parasites affect hosts' evolution to perform complex tasks (Zaman et al., 2014). In this system, hosts were under pressure to escape their parasites by changing their resource-gathering behavior. Hosts could do this by simply cycling between several simple behaviors if their parasites lost the ability to infect extinct hosts (i.e., lacked “genetic memory”).

Zaman and colleagues evolved populations both with and without parasites and recorded the complexity of the logic tasks the organisms evolved. Note that in these experiments, as opposed to the default configuration, there was no direct benefit to becoming more complex. Instead, a benefit for increased complexity could arise as a byproduct of co-evolving with parasites. They found that host populations with co-evolving parasites evolved significantly more complex behavior than hosts evolving without parasites. This increase in complexity could have been due to the step-wise escalation (i.e., arms race) between hosts and parasites, or it could have emerged because parasites were exerting some additional pressure on the hosts not to cycle back to previously evolved phenotypes.

To determine the cause of the increased host complexity, Zaman and colleagues experimentally manipulated the Avida system to eliminate any advantage to parasites in retaining a broad host range (i.e., their “genetic memory”) and reran their experiment. If the hosts had evolved increased complexity due only to the arms race with parasites, then they should have reached the same level of complexity. However, hosts evolved under this experimental design evolved significantly less complexity than hosts evolved with normal co-evolving parasites. In fact, these hosts were indistinguishable from host populations evolved without parasites. These results indicate that it was the diversity of co-evolving hosts and parasites that caused the parasites to retain a broad host range (a “genetic memory”), creating selection for increasingly complex hosts. Zaman and colleagues termed this the Population-Genetic Memory Hypothesis (Zaman et al., 2014).

Co-evolution with parasites could also affect the evolution of complexity via evolution of increased evolvability to escape those parasites. Zaman et al. (2014) were able to conduct an evolvability analysis by mapping every phenotypic change to every possible point mutation for the most prevalent host genotype from all populations of hosts with and without parasites at the end of the evolution experiments. They found that function-switching mutations were >10-fold more common in hosts that evolved with parasites than in hosts that evolved without parasites. This result indicates that co-evolving with parasites indirectly selected for more evolvable genomes, because hosts were then more likely to avoid infection.

Overall, these findings indicate that the presence of parasites can cause an increase in the complexity and evolvability of hosts. They also demonstrate the value of using digital evolution in examining these questions, because it enables increased control of possible complex behaviors and analysis of mutational effects.

However, previous work in digital evolution has not investigated the effect of mutualistic symbionts on the evolution of complexity, nor how endosymbiosis affects the evolution of the symbionts. These and other open questions are discussed in section 7.

The research discussed in the previous two sections has focused on the effect of a strictly parasitic or mutualistic symbiosis on various aspects of population dynamics. However, symbionts can evolve along a spectrum between parasitism and mutualism. Another important question is then: what influences that evolution? Two hypothesized influences are the vertical transmission rate and spatial structure. Vertical transmission (when a symbiont's offspring are able to infect a host's offspring) is hypothesized to align the reproductive interests of symbiont and host, because the symbiont's fitness can benefit from increases in the host's fitness (Shapiro and Turner, 2014). However, experimentally manipulating vertical transmission rate over evolutionary timescales can be difficult or impossible in many biological systems. Spatial structure (when offspring remain spatially close to their parents instead of dispersing away) has generally been found to increase the selection for cooperation, because it enables cooperators to stay near related organisms who are likely also cooperators, making cooperation both more beneficial and more resistant to cheating organisms (Yamamura et al., 2004). Part of the effect of spatial structure on cooperation, however, relies on kin selection (where a benefit to related organisms increases a focal organism's indirect fitness) and may not apply directly to cooperation between organisms of different species (Griffin and West, 2002). Therefore, determining experimentally the effect of these two factors on the evolution between parasitism and mutualism has historically been challenging.

To overcome this difficulty, Vostinar and Ofria created a simple digital evolution system to investigate these dynamics (Vostinar and Ofria, 2019). The Symbulation system currently supports a population of host organisms and a population of obligate endosymbionts. Organisms reproduce when they accrue enough resources, making fitness implicitly defined. An organism's phenotype depends on both its own genotype and its partners, and there are many possible genomes. In addition, organisms can evolve behavior along a spectrum from antagonistic (where the symbiont is parasitic and the host is defensive) to mutualistic (where both partners actively contribute resources to the other).

In this system, Vostinar and Ofria experimentally manipulated vertical transmission rate and whether offspring were placed immediately next to their parent or randomly in the world. They found that both with and without spatial structure, vertical transmission rate did increase selection for mutualistic behavior between the host and symbiont when there was sufficient starting variation in the population. However, regardless of vertical transmission rate, mutualism did not evolve when the hosts and symbionts started from an antagonistic relationship and parasitism did not evolve when the hosts and symbionts started from a mutualistic relationship (Vostinar and Ofria, 2019).

While the effect of vertical transmission rate was independent of spatial structure, Vostinar and Ofria found that the effect of spatial structure was not independent of the vertical transmission rate (Vostinar and Ofria, 2019). At high and low amounts of vertical transmission, the presence of spatial structure increased the amount of mutualistic behavior or decreased the amount of antagonistic behavior, respectively. However, at intermediate values of vertical transmission, the presence of spatial structure significantly decreased the amount of mutualistic behavior, in some treatments to the point that an antagonistic relationship evolved instead. Vostinar and Ofria hypothesized this to be due to a decreased amount of local symbiont diversity, which caused a reduction in the average mutualistic quality of symbionts that could infect naive hosts, selecting against mutualistic behavior (Vostinar and Ofria, 2019). These results indicate that vertical transmission does select for mutualistic behavior, but it is insufficient on its own, and it has complex interactions with spatial structure and population diversity which warrant further study as discussed in the following section.

As discussed in section 3, digital evolution systems typically do not attempt to simulate specific biological systems. Instead, they aim to find general processes and patterns likely to also be found in biological systems. Because of this, digital evolution can be viewed as a faster and less expensive way to initially test hypotheses about evolutionary processes. If the results found in digital evolution systems are promising, researchers may then choose to invest the time and resources to test if the results are also observed in biological systems.

The insights gained regarding symbiosis in digital evolution systems have impacted the biological literature several times. For example, Rocabert et al. (2017)'s findings were cited by Zhao et al. (2019) to potentially explain dynamics observed in human gut microbiomes, and Crombach and Hogeweg (2009)'s findings were included in Hindré et al. (2012), which reviewed bacterial adaptive dynamics. Finally, Zaman et al. (2014) has been quite influential on research regarding bacteriophage-host evolutionary dynamics (Masri et al., 2015; Burmeister et al., 2016; Stroud and Losos, 2016; Bento et al., 2017; Queller and Strassmann, 2018; Furubayashi et al., 2020; Ryan, 2020).

However, it is not necessarily possible or practical to replicate experiments from digital evolution in biological systems with current technology. For example, preventing any mutations from occurring, or only allowing mutations of a certain type, while allowing the populations to continue otherwise as normal is currently impossible in biological systems. Other forms of analysis are also currently infeasible, such as collecting a complete genetic sequence on every individual organism, identifying all mutations that have occurred in all populations, and determining what effects (if any) those mutations had (such as the evolvability analysis done by Zaman et al., 2014).

Further, the goal of many digital evolution systems and studies is to create an instantiation of evolution, not a simulation. Therefore, results do not need to be verified in biological systems, because they are accurate observations of the system in which they were found, in the same way that results from any particular model biological system are valid observations of the model system. As with any biological system, it is possible for the results in a digital evolution system to be due to artifacts or idiosyncrasies of the system. Ultimately, the aim is usually to find generalizable processes and patterns, and the only way to know if a process is generalizable is to look for it in many systems, both digital and biological.

The first set of open questions we will discuss focuses on the interaction between co-evolving symbionts. The presence of a co-evolving symbiotic partner is likely to influence the evolutionary dynamics of a species in many ways. Here we discuss eight dynamics—complexity, diversity, competition, sexual reproduction, eusociality, range expansions/shifts, and open-ended evolution—that have been hypothesized to be influenced by a co-evolving symbiont and are well-suited to being studied further using digital evolution.

Another set of questions that digital evolution is well-suited to address focuses on the evolution of symbiosis itself. As previously discussed, a symbiosis has several axes upon which it can vary:

These categories are all independent spectrums, and systems can shift along them over both individual lifetimes and evolutionary time. A symbiosis can also be a precursor to an egalitarian major evolutionary transition where the symbiosis becomes so tightly linked that the partners become a single organism (Estrela et al., 2016). Here, we propose specific hypotheses and more general questions surrounding the evolution of symbiosis that could be addressed with present or future digital evolution systems.

Endosymbiosis, a relationship where an organism lives inside its symbiotic partner, is an interesting phenomenon with both benefits and costs (Chomicki et al., 2020; Puri et al., 2021). It offers a potentially sheltered environment inside a host (or a more hostile environment, if the host is defensive) and the possibility of vertical transmission to the host offspring. However, it brings the risks of overspecialization and the complications of coevolution, as discussed previously. Therefore, a general open question is when would endosymbiosis be favored to evolve, and how does the form of the symbiosis influence that evolution?

Once inside a host, an endosymbiont could evolve away from its initial relationship to be mutualistic, parasitic, or commensal, or it could retain the initial effect it had on its host. As discussed in section 6, previous results indicate vertical transmission is insufficient for de novo evolution of a mutualistic endosymbiosis, because the endosymbiont went extinct before mutualism evolved. Clearly, the endosymbiont species must have the capability to survive long enough for any evolutionary pressures to affect them and their hosts.

There are many possible paths that an endosymbiotic species could follow, such as from a facultative mutualist toward an obligate parasite, or an obligate parasite to a mutualist (the Co-opted Antagonist Hypothesis, Johnson et al., 2021). Their relative likelihoods and contributing factors could be investigated in a digital evolution system that supports de novo evolution of endosymbiosis. Further, transitions from endosymbiotic parasites to mutualists appear to be fairly common in biological systems and determining the underlying mechanisms for this transition would be valuable for the field (Drew et al., 2021). Similar questions can be asked regarding ectosymbiotic dynamics, especially those that enable a symbiont to simultaneously interact with multiple hosts.

Finally, previous results have shown that spatial structure can affect the evolution of symbiosis in complex ways (Vostinar and Ofria, 2019). However, these results are only for obligate endosymbionts. Determining how spatial structure influences the evolution of symbiosis in other cases will also be important.

Early digital evolution systems (reviewed in section 5) were designed without a focus on enabling symbiosis. While some preliminary symbiotic interactions were observed, they were necessarily limited, because all organisms were required to occupy cells in a two dimensional grid. Additional functionality needed to be added to enable endosymbiosis, as was done by Zaman and colleagues in Avida (Zaman et al., 2011). Symbulation, created by Vostinar (Vostinar, 2021), was designed with symbiosis specifically in mind, but is a fairly recent addition to the software landscape and therefore missing many desirable features.

Ideally, improvements to existing systems, or the development of a new digital evolution system that supports symbiosis, will contain support for many forms of symbiotic interaction. To answer the questions discussed in the previous subsection, it will be necessary for such systems to support the evolution between different symbiotic dynamics. Specifically, we recommend that digital evolution systems for the study of symbiosis contain:

In addition, while improvements in underlying technology have enabled digital evolution systems to support fairly large populations and the evolution of hundreds of thousands of generations in long experiments, there is always room for further improvement. In particular, most digital evolution systems do not take advantage of the parallel processing capabilities of modern machines, though efforts are being made in that area (Moreno et al., 2021).

Finally, future digital evolution systems for the study of symbiosis should aim to improve access and ease of use. Many previous and current systems can only be used for evolutionary experiments when run on high-performance computing systems, which are not universally accessible. Improvements are being made in this regard, particularly with the emergence of digital evolution systems that can run in an internet browser, such as those supported by the Empirical platform¹. Further, digital evolution systems usually rely on advanced programming languages such as C++, making it challenging for less-experienced programmers to make modifications. However, researchers could design the systems with this knowledge in mind, making efforts to improve the modularity and readability of the systems' codebases.