Darwin's agential materials: evolutionary implications of multiscale competency in developmental biology

A critical aspect of evolution is the layer of developmental physiology that operates between the genotype and the anatomical phenotype. While much work has addressed the evolution of developmental mechanisms and the evolvability of specific genetic architectures with emergent complexity, one aspect has not been sufficiently explored: the implications of morphogenetic problem-solving competencies for the evolutionary process itself. The cells that evolution works with are not passive components: rather, they have numerous capabilities for behavior because they derive from ancestral unicellular organisms with rich repertoires. In multicellular organisms, these capabilities must be tamed, and can be exploited, by the evolutionary process. Specifically, biological structures have a multiscale competency architecture where cells, tissues, and organs exhibit regulative plasticity—the ability to adjust to perturbations such as external injury or internal modifications and still accomplish specific adaptive tasks across metabolic, transcriptional, physiological, and anatomical problem spaces. Here, I review examples illustrating how physiological circuits guiding cellular collective behavior impart computational properties to the agential material that serves as substrate for the evolutionary process. I then explore the ways in which the collective intelligence of cells during morphogenesis affect evolution, providing a new perspective on the evolutionary search process. This key feature of the physiological software of life helps explain the remarkable speed and robustness of biological evolution, and sheds new light on the relationship between genomes and functional anatomical phenotypes.

The basic workhorse of evolutionary theory is the cycle between the genotype (the target of mutations) and the phenotype (that which selection acts upon). While many models and analyses focus on these key elements, another is often neglected: the physiological processes that underlie morphogenesis. This is the control layer that sits between the genomically specified cellular hardware (proteins) and the form and function that selection sees: anatomy and behavior (Fig. 1). In effect, the behavior of cellular collectives in morphogenesis is the software of the system—the functional outcomes of the molecular machines encoded by genomic information [1]. This is relevant not only for embryogenesis, which converts compressed genomic information into a rich emergent set of large-scale structures, but also for regeneration, metamorphosis, remodeling, and other processes which establish and modify growth and form. Much work has addressed the evolution of developmental mechanisms, the evolvability of specific architectures, and the emergent complexity of epigenesis [2–6]. Moreover, recent work has begun to emphasize the active, cybernetic, problem-solving capacities of this process beyond feedforward emergence [7–11] and explore ways in which evolution increases the functional intelligence of cellular collectives [12–14]. Here, I focus on the complementary side of the evolution–intelligence feedback loop. This is fundamentally distinct from earlier efforts in the adaptationist/selectionist paradigms, and emphasizes problem-solving, unconventional embodied agency, and creativity that are specifically not due to adaptation. I first overview the data that illustrate the functional competencies of morphogenesis, casting multicellular growth and form as the behavior of a collective intelligence.

I then focus on a fascinating knowledge gap: what implications do the computational properties of the biological software layer have for the evolutionary process itself?Fig. 1Multiscale architecture of evolution and morphogenesis. A Biological systems exhibit complexity at every scale of organization, from molecular networks to whole colonies of organisms. Crucially, these are not just structural levels, but functional: each level’s components are problem-solving agents which navigate physiological, transcriptional, anatomical, behavioral, and other spaces. B Many organisms go through a bottleneck that forces a compression of form and function into a generative encoding which does not specify the final structure, but rather specifies a process which creates that form. This is used in machine learning as the “autoencoder” architecture (top sub-panel), which compresses data into a compact representation and then decompresses it to result in an output that is similar at large scale, but not identical in the lowest-level details. The process of reproduction via an egg stage (bottom sub-panel) has the same architecture, except that evolution not only learns the encoding, but also simultaneously encodes the structure of the developmental machinery—the decoder itself. C The evolutionary process has not only a genome (the target of mutation) and the anatomical phenotype (the target of selection), but also a critically important layer between them: the morphogenetic physiology that guides embryogenesis, maturation, metamorphosis, remodeling, regeneration, and suppression of cancer and aging. The circuits that implement the computational features of this layer themselves have a bottleneck (or “bow-tie”) architecture, exploiting biochemical, bioelectrical, and biomechanical modalities.

These use higher-level “virtual governor” [275, 295] nodes such as calcium patterns, resting potential, or tension—powerful dynamical control points that are complex functions of the underlying molecular details and do not map 1:1 with any gene or gene product. Images in all panels by Jeremy Guay of Peregrine Creative; Panel A used with permission from [14]

One key property of developmental morphogenesis is its emergent nature, in which the relationship between genotype and phenotype is highly indirect. It has long been clear that genomes do not directly code for anatomies; instead, DNA encodes for proteins—the nano-level hardware made available to each cell. The behavior of cells, in a “social” context of multicellularity [15, 16], is what gives rise to functional anatomies (Fig. 2). Cellular behaviors include proliferation, migration, differentiation, shape change, and apoptosis, operating in parallel over millions or billions of cells that are signaling to each other via chemical, electrical, and mechanical modalities—coordinating directly, at long range [17], or by using their microenvironment as a stigmergic scratchpad. This generative process is critical not only during embryonic development, but also in maturation, metamorphosis, regeneration, and suppression of cancer and aging—establishing and maintaining order across multiple scales [18, 19].Fig. 2Emergence and genetic encoding architectures. A The incredibly complex anatomical structure of an organism (shown here is a cross section through a human torso) is produced by the blastomere descendants of a single fertilized egg cell and its genome, which directly specifies protein sequences (some of the cellular hardware components), not large-scale geometrical features or functionality (nor aspects such as cytoskeletal structure or membrane composition, both of which are inherited parallel to the DNA [206]). B One way to achieve complexity via a compressed encoding is through “emergence”. Here are shown four simple rules of Conway’s “Game of Life” that specify the states of a specific cell at each time point based on the state of its neighborhood [296].

Such cellular automata produce patterns that, to a large-scale observer, reveal movement patterns (B’) such as ‘gliders’, and even structures (B”) that continuously produce these virtual moving dynamical patterns (entities), despite the fact that the rules are extremely simple and do not say anything about global patterns or their properties. C In the problem space of a complex coordinate plane (2-dimensional grid of positional information), a very simple, local policy can be defined for determining the state of every cell at a position in the grid: a color can be assigned based on how well a root-finding method, such as Halley’s method performs (a simple algorithm [297]). This gives rise to very complex biological-looking forms (C’) complete with multiscale structure and subtle left–right asymmetry. However, all of these kinds of emergent complex systems have a fundamental problem: they are irreversible. What changes could be made in the rules of (B) or the short complex number formula that encodes the image in (C’), to make a glider that moves at a different angle or a pattern that had a different number of blobs and spacing of “hairs” on them, respectively? Such open-loop (purely feed-forward) schemes do not facilitate control interventions (e.g., regenerative medicine) and limit evolution to a simple search of the micro-space of generative rules which is very rugged because the final features are an extremely complex, tangled result of the elements’ following of those rules. D A different architecture is suggested by considering that the subunits may be themselves complex components that implement feedback—cybernetic goal-directedness in which energy is used to optimize specific higher-level metrics.

The level of control of the subunits that build the organism, as well as of the organism itself (and any swarm of which it may be a part), can span many different competency levels on a scale such as proposed by [99]. E Simple picture of genes being responsible for specific features of the organism (which, had it been true, would have made the goals of genomic editing straightforward) has been supplanted with one in which each genetically encoded piece of hardware ends up contributing to many different components of the whole [248] (E’), which raises the inverse problem described above. In contrast, thinking of the genetics as specifying the hardware of computational agents that have competency to solve specific problems (e.g., build an appropriate structure despite changes in cell number, or starting position) offers a much more direct architecture. Closed-loop (homeostatic) agents enable behavior-shaping signals that alter setpoints, not just micromanage hardware components, allowing for a much more linear, tractable, and robust control policy that facilitates evolution as well as biomedical control. Images in panels A, B, E by Jeremy Guay of Peregrine Creative. Panel D used with permission from [99]

A major implication of this architecture is that it is irreversible—while it is straightforward to watch (or potentially to simulate) how biology follows local rules of chemistry and physics and thus to discover what anatomy emerges from a given genome, the inverse problem [20] is in general unsolvable: determining which protein sequences must be encoded to produce an arbitrary, desired large-scale anatomical form. This irreversibility of the recursive, highly emergent process of morphogenesis is what limits full-scale Lamarckism: the difficulty is not how to penetrate Weismann’s barrier and edit the genome in light of somatic experience—mechanisms exist for this [21–23]. Rather, it is how to know what to change in a genome to produce a desired feature based on physiological events (e.g., a longer neck). This is a direct consequence of the fact that while the central dogma could, mechanistically, be reversed, the output of the DNA- > RNA- > protein cycle is not anatomy, so reversing it does not solve the problem of going from anatomical features back to DNA. This is why we cannot predict the anatomies and morphogenetic capabilities of chimeras made of different cell types [24], despite having genomic information for both. For example, while we have the genomes for frog and axolotl species, one of which makes embryonic legs, we cannot predict if a frogolotl (chimeric 50/50 mix) would have legs, and if so, whether those legs will consist only of axolotl cells or both, or whether they will be regenerative like adult axolotl legs. This is because this is fundamentally a question of collective decision-making, which is still poorly understood. In fact, we cannot even predict a single-species anatomy from a genome without first comparing it to a genome whose anatomical outcome we already know.

This also sets a ceiling on the biomedical applications of technologies such as CRISPR and other forms of gene therapy: without being able to determine what we have to change to get a specific system-level outcome, our ability to control genomic information to solve complex injuries and disease states will be highly limited. Workers in the field of AI and machine learning are well familiar with this issue as “credit assignment”. However, as will be argued below, this limitation is actually a hugely important intelligence ratchet in biology, because it forces evolution to increase and exploit the competency of its material and the subsequent top-down control (biological innovation via behavior shaping).

The indirect nature and complexity of morphogenesis raises questions as to how evolutionary exploration can give rise to coherent organisms: the search space seems extremely rugged and hard to navigate using the hill-climbing algorithms we associate with evolution. This has given rise to a number of proposals [1, 25–36] to expand the neodarwinian synthesis in various ways. Here, I take a different direction, focusing on one specific aspect of morphogenesis which may provide a critical missing piece in our understanding of evolution: the affordances supplied by the plasticity and problem-solving capabilities of the cells which form the substrate of morphogenesis. Metazoan cells have numerous adaptive behaviors because they derive from ancestral unicellular organisms that needed a full range of behavioral capabilities to survive. Thus, the evolution of metazoan anatomy operates not on a passive material, but on an agential one [37]. I argue that what evolution is really searching is not the enormous space of all possible local rules, but instead the space of behavior-shaping signals by which cells hack each other’s functionality, and that the collective intelligence [38, 39] of cellular swarms has major implications for the rate and course of evolution.

New developments in the science of collective intelligence, and classic concepts from behavior science, are beginning to provide a crucial new lens on morphogenesis that complements the current view of patterning as a complex dynamical system (such as a cellular automaton) with open-loop emergent complexity. This stance emphasizes plasticity and computational competencies for problem-solving in diverse spaces—aspects which need to be dealt with using ideas from cybernetics [8, 9, 40], as well as physics. William James defined intelligence as the ability to reach the same goal by different means [41]. The emphasis is not just the well-known importance of modularity [2], but also the critical cybernetic properties of biological modules and the ways these enable reliability and task delegation (i.e., problem decomposition) across scales in vivo. Together, these features connect readily to concepts such as re-programmability, hacking, predictability, and multiscale polycomputation [42, 43]. This has massive implications for evolution, which are described below.

It is critical to emphasize from the outset that this research program does not support non-naturalistic explanations for evolution, does not make use of exotic influences on mutation such as quantum effects [44, 45], and does not require any goal-directedness in the evolutionary process at the large scale. Instead, it builds on progress in connectionist machine learning [34, 35, 46–50], basal cognition [12, 51–55], and an extension of neuroscience to developmental bioelectricity [56–58] to reveal how the collective intelligence of cells serves as an affordance for evolution. In that, it is complementary to current efforts to understand how network and other generic mathematical and physical properties affect the evolutionary process [59–68]. It should be noted that the below discussion is specifically focused on the traditional view of evolution, avoiding the fascinating and critical roles of epigenotype, epimutations, and other mechanisms that enrich the dynamics of inheritance. My goal here is to examine the major impacts that cellular intelligence has on even a mainstream view of the Modern Synthesis. Thus, I avoid the interesting (and widely debated) aspects of epigenetic inheritance [28, 69–73], which will be incorporated in future work, to focus attention on a number of unconventional hypotheses in the context of an uncontroversial view of evolution.

The implications of developmental architectures for evolution have begun to be explored via the understanding of phenotypic plasticity [74–78] and the way that behavioral agency of whole organisms shapes how evolution explores the latent space of phenotypes [77–82]. For example, the changes in skeletal features, behaviors, and physiology of fish and amphibia in novel stressful environments [83, 84] have been suggested to bias evolution and strongly complement traditional mutational variation as a source of change [77, 78, 85]. However, the crucial aspect is not simply that the complexity of development enables changes by real-time events, and that the genome does not uniquely specify one possible outcome. More than that, a view of morphogenesis as simply a recursive, massively parallel process limits bottom-up control via genetic editing in biomedical engineering for the same reason it limits the efficiency of evolutionary search across micro-level properties of a passive material. It is proposed that for the same reasons that bioengineers are now turning to top-down control policies and interventions [37, 58], evolution has discovered an architecture that exploits its unique substrate: cells and tissues with specific competencies, behaviors, and input-driven decision-making computations.

It is proposed that the plasticity of development and its implications for evolution cannot be optimally captured by an exclusive dynamical systems perspective in which different outcomes simply occur as the result of different environmental inputs working their way through a mechanical process. It has been suggested that morphogenetic mechanisms exist on a continuum of problem-solving capacities that exists across scales within the biosphere [54, 77, 81]. Evolution can thus be seen as not only shaped by organism-level learning and intelligent behavior [27], but also by physiological and anatomical level intelligence [55]. This perspective links the plasticity of the morphogenetic process with the field of basal cognition, which seeks early evolutionary origins of classical intelligence and problem-solving in behavioral science by looking at lower level examples of adaptive behaviors that rely on information processing. Cells and their components have many competencies for adaptive behavior when navigating physiological, metabolic, and transcriptional spaces [86]. Examples include the ability of gene-regulatory networks and chemical pathways to learn from experience [87–89], for transcriptional machinery to rapidly adjust to compensate for entirely novel physiological stressors [90], and for unicellular slime molds to explore their environment at a distance and make decisions using memory and comparison of alternatives [91–95].

The impact of intelligence on evolution has been studied [36, 96]. However, it has largely focused on intelligence of the organism (classical behavior in 3D space), and is only now beginning to be extended to higher levels such as the intelligence of swarms [97, 98] and entire ecosystems [49]. The framework discussed herein extends the study of the relationship of intelligence and evolution in several basic ways. It extends the notion of intelligence to sub-organismal scales, casting morphogenesis as the result of collective intelligence at the molecular cell, tissue, and organ levels. It operates within a gradualist [14] perspective on intelligence and goal-directedness [57, 99], which are both used here in a naturalistic, cybernetic, engineering sense of varied degrees of competent problem-solving in diverse spaces by unconventional agents (i.e., not restricted to higher-level cognitive capacities in brainy animals). Thus, it emphasizes morphogenesis as a computational process, not merely the emergent outcome of highly parallel local rules that could be dealt with by established complexity theory tools. Moreover, it expands the concept of intelligent behavior across a key invariant: effective navigation of diverse problem spaces, which includes problem-solving in physiological, metabolic, transcriptional, and anatomical spaces [12, 13, 86]. With these expansions in hand, it becomes possible to see how evolution pivoted some of the same strategies across these domains, and to then explore the implications that this multiscale competency architecture has for the evolutionary process itself.

Several open problems are addressed below. What are some of the features of morphogenesis that enable evolution to search a very difficult space (with pleiotropy, degeneracy, redundancy, etc.)? How can evolution be so rapid and effective at using mutation to address opportunities of form and function? How do we explain the ubiquitous presence of novel capabilities in living things for which no direct selection forces are plausible? Analysis of morphogenesis as the behavior of a collective intelligence of cells leads to the following key proposals. First is that the space which evolution actually searches is not only the space of microstates of the genome, but also a much more tractable space of behavior-shaping signals: evolution exploits cellular intelligence as a highly exploitable affordance. Second is that by using a variational autoencoder-like architecture of compression into a generative genome (Fig. 1), evolution is freed from over-training on past conditions and pushed to evolve general-purpose problem-solving machines which inevitably display robustness, plasticity, and adaptive success in entirely novel circumstances. Interestingly, these ideas are more familiar concepts in neuroscience; consistently with the recent advances in basal cognition [100–104], these concepts apply long before complex brains appear. My hope is that the following analyses extend an active research program on these ideas in regenerative medicine [58, 105–109] and synthetic bioengineering [37, 101, 110–113] to also impact evolutionary biology and engineering of artificial systems via evolutionary algorithms [96, 114–119].

Basal cellular competencies did not disappear during the transition to multicellularity: instead, they scaled up as a collective intelligence to operate in larger and more complex problem spaces [13]. Models of mechanisms by which single-cell competencies potentiate problem-solving capacities in larger and different problem spaces include gap junctional merging of intracellular milieus and other forms of tissue-level signaling (e.g., stress propagation). This promotes sharing of chemical engrams, which also help to wipe the individuality of cells, and enlarges the memory, predictive capacity, and sensory/effector radius of cellular networks, enabling them to pursue larger goal states in spaces like anatomical morphospace [14].

The key feature of scaling up individuality (whether evolved or engineered) is for higher levels of control to get their components to do things they do not do when operating as individual units [46, 120]. This can be discovered by extracting the parts of organisms and examining their behaviors in new contexts [121]. A cell-level example is that fragments of keratocytes move in a specific direction in an electric field, but whole keratocytes (collections of such fragments) actually move in the opposite direction [122]. A morphogenetic and behavioral example is seen in Xenobots [123, 124]. Frog embryo skin cells, in vivo, form a two-dimensional, passive layer on the outside of the animal that protects it from pathogens. However, when liberated from the instructive influences of the other cells, frog epithelial cells instead form a Xenobot—a functional, self-motile construct with many novel behaviors that are normally suppressed and hidden by the instructive signaling of other cells during development [125]. Human tracheal cells from adult donors also form such biobots, with the remarkable ability to traverse and heal neural wounds [126]. Thus, it is not obvious what the default morphogenetic behaviors and capabilities of cells are, because of the ubiquitous dominating controls of other cells in their environment. In some cases, embryonic capacities (e.g., formation of appendages) are “lost” at maturity (e.g., limbs in adults of many species), but can be activated in non-regenerative contexts by specific stimuli that can be provided by bioengineers in the context of regenerative medicine [127–129].

These and other examples indicate that biological components are themselves, to varied degrees, autonomous, but are controllable by signaling from other cells [37]. The dynamic of using simple signals that take advantage of the recipients’ complex, reliable repertoire generalizes to the concept of hacking, which is applicable at multiple scales and in many contexts (Fig. 3), ranging from chemical signals in cellular induction to colony-scale behavioral phenomena driven by acoustic signals [130]. The crucial focus in this concept is on the role of an agent that takes advantage of affordances in its own way, not necessarily in the ways “intended” by an engineer, or by evolutionarily-prior functions. Past discussions of relevant phenomena include developmental niche construction [131, 132], cancer cells reprogramming normal neighbors [133], host-altered behavior due to parasites that exploit neurotransmitter mechanisms [134] (and the reverse—the host evolving machinery to manipulate the parasite [135]), zombie ants controlled by fungi [136], the explore/exploit strategies of young mammals [137], and many others. An amazing example of morphogenetic hacking is the formation of galls, where signals from a parasite force the leaf cells away from their normal flat, green tissue phenotype and into building spiky, three-dimensional colorful forms [138].Fig. 3Morphogenetic hacking. A Morphogenesis is implemented by cells, which were once organisms themselves. Thus, the process of anatomical homeorhesis [252, 298] is implemented by a set of behavior-shaping signals in which some cells control the behavior of other cells. Image by Jeremy Guay of Peregrine Creative.

B One example is seen in the behavior of frog skin cells, which produce a passive, two-dimensional outer layer in a normal frog embryo, but when freed from the instructive interactions of other cells become Xenobots—spherical constructs that move autonomously and even exhibit entirely new behaviors such as kinematic self-replication (B’), in which they assemble loose cells into clumps that become the next generation of Xenobots and repeat the cycle in subsequent generations [124, 125]. Panels B, B' by Douglas Blackiston, used with permission from [125]. C When bioelectric signals that indicate “build an eye here” are reproduced in other locations of a frog embryo via potassium channel modulation [152], a complete eye can be formed even out of gut cells without needing to micromanage the construction of a complex organ—the modular trigger induces a morphogenetic process that can reliably take place in novel environments. Panel C used with permission from [148]. D The process is more than just triggers of master-regulator signals. When an insufficient number of cells is modulated (indicated in blue beta-galactosidase tracer), they recruit their normal (unmodified) neighbors to complete the task of building an ectopic lens of the right shape and size. This ability to complete a particular action in morphogenetic space by controlling the behaviors of other cells in a context-sensitive, dynamical manner is a key feature of controllability and reliability—it is how the problem-solving capacities of cells, not just ability to execute the same hardwired local rules, facilitate adaptive robustness in novel circumstances. Panel D courtesy of Sherry Aw, used with permission from [37].

E The controllability can be exploited even in cross-kingdom interactions (not just by bioengineers), in which plant cells are induced to form morphologically highly novel structures known as galls by signals from fungi or insect embryos: the novel structures, which differ greatly from the default, are an example of natural hacking of morphogenetic machinery (E = urchin gall on a scrub oak, by Timothy Boomer at WildMacro.com; E’ = hedgehog gall, by Andrew Deans)

Many such phenomena work because of the agency of the material being hacked. This includes viral infections since viruses have very limited capabilities and rely on the cell for all of its physiology; the use of toxins to manipulate hosts [139]; and the phenomenon of kinematic self-replication by Xenobots. These motile synthetic organisms make copies of themselves in an evolutionarily novel manner: when provided with loose cells, the biobots corral them into piles which self-compact and thus spontaneously form the next generation of biobots which go on to repeat the cycle [125]; this only works because the material itself is competent to form a viable Xenobot when rearranged (Fig. 3). At higher levels of organization, one can see classic organism-level communication and signaling as examples of hacking the more agential elements of one’s environment (i.e., the conspecifics, predators, and prey contained in it).

Thus, the concept of hacking extends well beyond the typical strategies used for traditional engineering with inert materials, or computational matter (as in, the computer science notion of hacking [140]), to systems with significant agency: hacking is conceptually linked to behavior shaping. In this perspective, everything in biology is a hacker, reaping rewards of efficient manipulation of its environment (and its internal components) using the appropriate tools (from direct chemical effects to subtle signals meant to be interpreted by complex agents): cells, parasites, viruses, transformed cells defecting from the morphogenetic goals of a tissue (cancer), tissues and organs during morphogenetic induction and competition (even within the same body [141, 142]), whole organisms, and swarms. Evolution can thus search the space of behavior-changing signals, exploiting the complex, agential nature of the cells which are its substrate as a hugely powerful set of affordances [143].

One especially powerful interface. which is exploited by evolution to accomplish morphogenesis, is the bioelectric layer of control (Fig. 4). Cells perform numerous computations in controlling their behavior via the voltage dynamics of ion channels and pumps [107]. But evolution discovered very early on, around the time of bacterial biofilms [144], that electric networks are also extremely convenient ways to integrate and process information across space and time, and to scale the functions of individual computing units towards a large-scale task (the same reason that brains, and computers, use this modality [145]). Several examples from the field of developmental bioelectricity illustrate this point.Fig. 4Bioelectric machinery underlying morphogenetic control. A The computational architecture of the nervous system relies on a network of cells, which set their resting potential via ion channels in their membrane, and communicates that state via controllable electrical synapses known as gap junctions [299]. B This machinery is ancient, with all body cells having resting potential across their membrane (Vmem) mediated by the same ion channels and most cells making gap junctional connections to form bioelectrical networks [188]. C This interface can be manipulated in any tissue in the same way as neuroscientists probe brain tissues: using molecular reagents to control connectivity (via opening and closing gap junctions) or directly setting voltage states by opening and closing ion channels (using drugs or optogenetic light stimulation). Also, the downstream messengers (neurotransmitters) operating in many tissues can be controlled directly [300, 301].

D An example of imaging of the bioelectrical prepatterns that define morphogenetic setpoints is the “electric face” revealed by signals from a voltage-sensitive fluorescent dye that indicates the future positions of the gene expressions, and subsequent anatomical structures, of the early frog embryo anterior ectoderm. This pattern is instructive because if it is shifted artificially and induces the predictable changes in subsequent development. Note the simplicity of the encoding—the bioelectrical states represent high-level order (elements of the target morphology for regulative development), such as the location of an “eye” or “mouth”, not the state of gene expression or individual stem cell fate. E The bioelectric representation of the target morphology guiding the cellular collective can work as a counterfactual memory, as in the example of the cryptic planaria: animals with normal one-headed anatomy and gene expression, but a two-headed target morphology representation in the bioelectric circuit (shown here using the voltage dye imaging), which causes them to build two-headed animals if cut into fragments [193]. F The bioelectric prepattern is a true memory because, despite the wild-type genetic sequence, fragments from two-headed animals continue to generate two-headed animals in perpetuity (with no more treatments) [195]—this is an example of the software layer that is reprogrammable and enabled by the genetically specified ion channel hardware, which does not need to change to be shifted to an entirely different large-scale target morphology. G Not only can head number be reprogrammed by a brief physiological stimulus to the morphogenetic agent, but also head shape: an animal with a triangular head shape can be shifted toward morphologies (including brain shape and stem cell distribution) [197, 198] of other species with round or flat heads.

This enables the cellular collective to explore attractors in morphospace (G’), using the exact same genome, which eventually could become canalized into genetically distinct species. Panels in A, B, C are courtesy of Jeremy Guay of Peregrine Creative; A, B used with permission from [188]. Panel D is used with permission from [150]. Panel E is used with permission from [106]. Panels F, G are used with permission from [107]. Panel G’, by Alexis Pietak, is used with permission from [197]

For example, in frog embryos, a set of cells known as “instructor cells” [146, 147] normally suppress melanocytes’ native behavior via serotonergic signals. In the absence of these signals, the melanocytes exhibit an over-proliferative, hyper-invasive phenotype strongly resembling melanoma [148]. Even more interestingly, the bioelectric signaling scales independent decision-making of the cells toward a collective outcome. When the control system is perturbed, stochastic outcomes (e.g., 70% conversion) result, but only at the level of the organism—each animal’s melanocytes are either all normal or all converted, at some stochastic rate: they toss a coin, but all the cells are tossing the same coin. It required a machine learning algorithm to analyze this control system and discover an intervention that breaks the concordance and enables individual melanocytes within the same body to make independent decisions [149]. This bioelectric signaling, which involves serotonergic control, parallels what nervous systems do to bind individual cell behaviors, even stochastic ones, toward a system-level collective behavior.

The vertebrate face [150, 151] is patterned by a set of bioelectric gradients that regionalize gene expression to determine the locations of eyes, mouth, and other components. Reproducing these patterns elsewhere in the body via misexpression of specific ion channels enables the induction of whole ectopic organs, such as eyes [152] on the gut or tail of a tadpole (thus unlocking novel capabilities of body regions, since only the anterior neurectoderm is competent to make eyes via signals from the chemical master regulator Pax6). Remarkably, when only a small number of cells are injected, they recruit their normal, unperturbed neighbors to participate in the process to make a normal-sized lens (Fig. 3). Thus, in response to the bioengineer’s instructions to “make an eye”, the prompted group of cells not only activates all of the required downstream molecular events (which are not directly specified by the simple, trigger-like stimulus supplied), but also determine how many additional cells are needed for large-scale morphogenetic setpoints such as size control and signal to activate them as well. This ability to scale recruitment of elements to task size is similar to that seen in collective intelligences such as colonies [153, 154].

Crucially, this is not just modularity [3] (“do the same thing again, in another place”), but takes advantage of cellular swarms’ competency to adjust dynamically to circumstances as needed, because the modules are not structural or hardwired. The cybernetic lens is appropriate here (and fruitful) because cell groups demonstrate diverse degrees of ability to accomplish a defined set of outcomes despite changing circumstances and can be manipulated top-down. This system for integrating the behavior of cells and tissues offers many advantages for phenotypic richness, but like any programmable interface is amenable to exploitation. For example, the bioelectric system for control of head shape and location in planaria [155, 156] can readily be hijacked by microbes that can control the number and structure of their host’s heads [157]. It is straightforward for evolution to also exploit these modular competencies: to reproduce an eye in another location, evolution does not need to reproduce all the underlying mechanisms.

“Intelligence: the ability to reach the same ends by different means.”― William James