Cellular Competency during Development Alters Evolutionary Dynamics in an Artificial Embryogeny Model

Here, we focused on a specific and novel question: the implications, for the rate and course of evolution, of a developmental process that exhibits competency at problem-solving in anatomical morphospace [59]. We produced a minimal simulation (Figure 1) that abstracted away many biological details to focus on a simple architecture: a phenotype easily optimized by traditional genetic algorithms, and a new component: competency of the individual cells to move based on interactions with neighboring cells, simulating a single body axis morphogenetic gradient of positional information [59].

Our approach is related to the work on the role of learning in evolution [58,60,61,62,63,64,65,66,67]. What is similar is the emphasis on genetics as the specification for a system that will then exhibit diverse behaviors that are not themselves hardcoded in the genome. Additionally, similar is the fact that DNA, as a code for specifying protein sequences, actually cannot encode directly for behavioral repertoires any more than it can directly encode morphology. What is crucially different is that traditional approaches focus on animal-scale behavior, which requires a novel and specific mechanism to evolve, such as nervous systems wired so as to facilitate specific types of learning. In our model, there is no learning needed; moreover, our system’s competency arises from a basic property of single cells: the ability to sense their neighbors, prefer those of similar type, and migrate as needed to reduce stress based on unmet expectations (e.g., intercalary regeneration in the limb [68,69,70,71]). Because cells were once themselves individual organisms and are organized into networks with homeostasis, allostasis, and homeorhesis properties [72,73,74,75,76,77], evolution is working with an agential material [24,26], which has competencies that do not need to be evolved directly (are present from the start).

However, it is likely that the cellular competency we examine here, and the behavioral learning that has been modeled by others, interact in ways that have evolutionary impact. For example, cellular activities can be the subject of behavior shaping by signals (implemented by evolutionarily-sculpted properties of the subcellular hardware such as signaling machinery and GRNs). In other words, much as organism-level learning enables an individual’s function to be molded by signals from conspecifics and parasites; similarly, cellular competencies open cells up to beneficial or detrimental signals from other cells in the organism itself that can control them via real-time triggers. It will be interesting in future work to understand how many results from the evolutionary learning field carry over to the evolutionary implications of competency, and how these two different aspects of the divergence of genotype from phenotype interact with each other in hybrid models that have both features.

Our framework was more complete than many evolutionary simulations because it included an explicit developmental layer between the genotype and phenotype. It was multiscale in the sense that important changes occurred on an evolutionary scale across individuals, but also ones driven by components of those individuals within their lifetime—the cells, which had their own perspective and local goals. However, our system clearly omitted a huge amount of biological detail with respect to cellular mechanisms of sensing, competition, cooperation, etc. We intentionally designed a minimal model to specifically focus on a few sufficient dynamics, and this likely under-emphasized the difference between cellular competencies and, for example, effects of learning at the organism level on evolution. Fundamentally we explored a toy model virtual world in which the individual roles of selection and competency could be quantitatively dissected in the absence of confounding complexity—we sought generic laws and dynamics [78,79,80,81,82], not a simulation of the detailed trajectory of any existing biological species.

Future work will add physiological layers, diverse cell types, computation at gene-regulatory and cellular-network levels, and a multi-dimensional target morphology (e.g., 2D or 3D pattern instead of just one primary axis) to more closely model biological reality. There is also much that can be improved with respect to the specific mechanisms that cells use to implement their competency: a rich set of diverse genes will be added in the future to enable evolution to manipulate different types of local goals and competencies. Moreover, recent discoveries in transgenerational inheritance [78,79,80,81,82] suggest that barrier between the genome and the phenotype is at least somewhat porous, and the effects of propagating sort order to offspring should be investigated.

We found that providing cells with a minimal homeostatic competency to improve their position in the virtual embryo results in better performance of the evolutionary search. Populations reach better fitness values faster when cellular activity is able to make up for genetic deficiencies (Figure 2). Indeed, in mixed populations, competent individuals tend to dominate and rapidly take over (Figure 4), as long as they have a minimal level of competency and/or are present in adequate numbers (Table 2). The simulation highlighted the distinction between two properties of each individual that are often conflated or obscured in simulations that do not include an explicit competency step: genotypic vs. phenotypic fitness.

Indeed, biology has many examples of evolution’s attempts to gauge genomes that it cannot see directly, for example by fluctuating asymmetry [83,84] and the near universal standards of sexual selection for left-right symmetrical features (which in turn is correlated with lack of genetic damage) [85,86,87]. In this system, we see that competency results in good phenotypic fitness but takes selective pressure off of genotypic fitness, which settles at a sub-optimal level (Figure 3).

Perhaps the most interesting aspect was the role that competency plays in exacerbating the inability of selection to evaluate the genetic material that gets passed on to subsequent generations. We observed that increases in competency made it harder and harder for selection to pick the best structural genes. Specifically, the correlation between genotypic and phenotypic fitness drops to insignificant levels very rapidly (Figure 5C). This could be expected to result in complex dynamics, because competency improves fitness of individuals but impairs the ability of the evolutionary hill-climbing search in fitness space to pick out the most elite structural genomes. Thus, we studied what happens when evolution is also allowed to control the degree of competency, which is biologically realistic since cellular capacities for sensing, computation, and action are themselves under evolutionary selection. We observed that the population drives towards picking the highest competency gene value in the population (Figure 5A), settling at a value close to 470. While this value of competency is sufficient to boost an embryo’s fitness to maximum, it is not necessary. We propose the following explanation.

Initially when evolution begins, the ordering of the cells is far enough from ascending order that a high competency gene value is required to create individuals with high fitness. At generation 20 or so, maximum fitness is achieved by choosing high competency gene values, and by simultaneously improving structural genome quality to 57% genotypic fitness. After generation 20, the genotypic fitness drops to a value of 52% and stabilizes with no further improvements. From this value of genotypic fitness, a competency value of 364 would theoretically be adequate to reach peak fitness. As a result, there is no selection pressure for evolution to always pick the highest possible competency value (i.e, a value of 480 seen in the shaded area of Figure 5A), because a value of 480 confers no additional phenotypic fitness over a value of 364. They are perceived as equal by the selection process and hence a random walk between these values would suffice. The reason we notice evolution picking values close to 480 at the end of 1000 generations could be because of a stochastic component to selection of embryos with competency gene values above 364. In our experiments, if >10% of the population have a fitness of 1.0 before selection, we pick the first 10%. This leads to the random selection of competency genes within the range of [364, 480].

In our models, we had to make a number of quantitative choices with respect to the evolutionary process. Thus, we checked how sensitive our conclusions were to these decisions via a hyperparameter scan: re-running the simulations with different choices for various hyperparameters (see Supplement S2.3.1). Specifically, we identified mutation probability and selection stringency as key hyperparameters which could influence the results of evolution. In an effort to probe their influence on the final competency gene value attained, we ran this experiment for 132 different combinations of mutation probability and selection stringency in the range of [0.2, 0.8] and recorded the stable-competency value attained for each hyperparameter combination (Figure S1 in Supplement S2.3.1). Correlation analysis revealed that a correlation of −0.4 existed between mutation probability and stable-competency-gene-value. However, no relationship was found between selection stringency and the stable-competency-gene-value. A possible reason for this could be that after generation 20, almost every embryo in the population achieves maximum phenotypic fitness, therefore there is no difference in choosing the top 20% of the population or the top 80% of the population. Mutation probability on the other hand has a direct influence on changing individual fitness, which explains its moderately significant relationship with the stable-competency-gene-value.

The competency of a population can be seen as granting robustness against perturbations, i.e., competency resolves aberrations in the genome and lessens the burden on evolution. The role of robustness in evolution has been a popular topic of discussion [88,89,90,91,92]. A population’s robustness is hypothesized to cause an evolutionary reduction in its adaptive performance; a sort of maladaptation caused when improved robustness traits layer on top of one another over evolutionary time and hide the underlying adaptive traits. This paradox has been shown to have broad implications on organismal design and is supposed to be a key aspect of evolution. Our results are in line with this paradox. Figure 3 is a clear depiction of the role robustness plays in hindering the quality of the genome. At each generation, increasing competency adds robustness that shields genomes which otherwise would have been culled by evolution. When compared to a population with no competency (hardwired), genomes in competent populations stabilize to a mediocre value whereas the untampered hardwired genomes rise steadily to maximum fitness. However, it must be noted that this is not necessarily a disadvantage. The paradox reveals the efficiency of competency: genomes need not be perfect; a stable threshold value of the structural genome is all that is required for competency to boost an individual’s fitness to maximum. Genetic information and problem-solving capacity of the cells work together to achieve a perfect solution to this fitness function.

This dynamic relationship between genotype and cell competency demonstrated in our simulation uniquely explains the remarkable example of planarian biology described in the introduction [19]. How can animals with a chaotic genome have such robust anatomies? We propose that planaria are an example of runaway competency: when cells get really good at making up for deficiencies in the structural genes, evolution has such a hard time selecting for the best genomes that further improvements instead increase generic competency to reach their target morphology despite perturbations. This positive feedback loop results in biological hardware that is highly successful at maintaining a specific morphology in a wide range of circumstances.

Tolerance to genetic and environmental insults is seen to some extent in other species; for example, human embryos are tolerant to being split at early stages, creating normal monozygotic twins, while mutations in important genes can sometimes be overcome by development [93,94,95]. However, in the amazingly regenerative planaria the effect was apparently much stronger. We propose this also as an explanation for another curious aspect of planaria. In every other model species, mutant lines are available—fruit flies with different number of wings or color of eyes, mice with abnormal tails, and many more genetic strains that are available from stock centers. In planaria this does not exist—no morphologically abnormal genetic strains have been reported. In fact, the only available abnormal line of planaria is a permanently two-headed form [96,97,98], which was produced not genetically but by manipulating bioelectrical signaling—the modality that is used to coordinate cellular competency [99,100,101], as is predicted by our model for species like planaria. Given their resistance to mutation, it’s unclear how speciation in planaria happens, but it should be noted that the same bioelectrical strategy that controls computation and cognition (i.e., behavioral competencies) in brains has been shown to coax genetically wild-type planaria to grow the heads appropriate to other species [102,103].

Of course, competency is itself carried out by molecular hardware which itself is subject to evolution and is encoded in the genome. However, it has long been clear that genotype does not uniquely determine the phenotype [104,105,106]). Development (and thus, evolution) can make use of many principles of physics (bioelectric computations, biomechanics, GRN memory, and other inherent properties [14,15,16,107,108,109,110,111,112,113,114]) that are not directly encoded anywhere but are exploited by the genome-specified machine. Our simulations study the effects of one type of such “free lunch”: cellular positional preferences and ancestral capability of motility during development, which are distinct from the environmental influences studied during typical epigenetics research. Our distinction is between structural genes (which directly specify phenotypic features) and competency genes (which specify a problem-solving machine that can exert context-sensitive activity). This is a powerful distinction for the same reason that the hardware-software distinction has driven a revolution in information technology. While the hardware (genome) is essential and important, software (competency) harnesses novel laws of physics, computation, and information processing that are not directly encoded by the hardware. This is akin to the way a logic table is implemented, but not directly encoded by, the specification of the transistors that make up a logic gate. This has been emphasized by fascinating work on the “arrival of the fittest” (evolutionary exploitation of “free lunches” provided by generic laws such as network properties [107,113,114,115,116,117,118,119]). While genes determine enzyme function fairly directly, the relationship between genes and complex morphology and behavior is extremely indirect [120]. However, the distinction between these modalities is not binary.

Thus, a more nuanced future framework will quantify (and exploit) a continuum of degrees of directness with which a generative encoding determines form and function from a given informational seed.

In planaria, most of the evolutionary “effort” seems to have gone into perfecting the algorithm (the ability of cells to create a normal worm morphology), vs. keeping a clean genome, because of the vicious cycle of competency increases. We found that the competency gene is changed significantly more often over the course of evolution than any structural gene (Figure 6). When gains can no longer be made efficiently by tweaking the genome (once selection cannot reliably pick out the good genotypes), all the effort goes in to increasing the competency level. This suggests the existence of a powerful ratchet mechanism in which evolution progressively becomes locked into improvements in the intelligence of the agential material with which it works, with reduced pressure on the structural genes. A positive feedback loop in which evolution increasingly puts more effort into the developmental software than perfecting the hardware points to a possible drive for scaling intelligence in morphological and other spaces [53,54,55,121,122,123]. It is possible that a drive for increased competency is an ancient and ubiquitous pressure [72], which plays out to different degrees in different biological lineages based on other aspects of their environmental and reproductive hyper-parameters.

One factor resisting the runaway positive feedback for multiscale intelligence is the cost of competencies. When included in our models (Figure 7B), it induced a classic Baldwin effect of assimilation into the genome and subsequent lessening of the drive for competency. However, whether this is realistic remains to be determined by measurements in vivo that have not yet been done. On the one hand, it is reasonable to posit that specific developmental computations (that might be needed for anatomical homeostasis for example [124,125]) could carry a metabolic or other cost. On the other hand, these may be capacities that cells are already doing regardless—they may be impossible to turn off, and may represent a use of internal processes that carries no extra penalty. Examples include bioelectric signaling that controls morphogenesis via ion channels needed for housekeeping physiology and cancer suppression [99,126], and learning properties of gene regulatory networks [127,128], which are emergent and require no new mechanisms for structural plasticity. Moreover, some properties (such as behaviors and morphogenetic outcomes) may simply be too hard to encode genetically, since genes directly specify proteins—not complex anatomical states. Our simple model of a 1D positional information axis did not enable that distinction (which may have otherwise limited the Baldwin effect and kept up the pressure for competency, for the same reason that brainy and highly morphologically plastic animals have advantages, despite the possibility of assimilation).

Another reason to include a competency penalty is to account for the extra developmental time that may be needed for the cellular activities to take place. However, it is not clear that this is a fair adjustment. We know of no data to suggest that the cleverer activity of competent morphogenetic processes takes longer than is required by minimal, feed-forward developmental mechanisms. Thus, giving a hardwired individual credit for completing development faster (equivalent to the penalty for competency in our Baldwin effect experiments) may not be an accurate modeling of the biology. Thus, we believe conclusions about the Baldwin effect and the limitations on competency observed in Figure 7 should be re-investigated in future work, when the real-world costs of these processes can be measured.

These results suggest a diverse research program on the evolutionary interplay between biological hardware and software. We suggest that the field of basal cognition [52,53,54,57,121,122,129,130,131] is an important part of understanding evolutionary developmental biology [73,132,133,134,135], and that intelligence (problem-solving competency) was an evolutionary driver long before complex brains and muscle-driven behavior arose [52,53,56,136,137,138,139,140,141,142,143]. Beyond understanding natural evolution, we suggest that the design of autonomous robotics [144,145], synthetic life [146], and interventions for regenerative medicine [35] can all benefit from deciphering and exploiting the multiscale competency architecture so richly exhibited by living forms.