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

The multiscale competency architecture (MCA) was foreshadowed in the notion of semiautonomous processes and dissociability by Needham, Gould, and others [242–245], but it has not heretofore been sufficiently appreciated. The ubiquitous embedded agency of evolution’s substrate means that the layer between the genotype and selection can make up for a wide range of errors and novel circumstances. In some cases, the correct species-specific target morphology is recreated (e.g., scrambled embryonic frog faces), while in other cases a different functional morphotype emerges (e.g., Xenobots). The same is true in physiological and transcriptional cases (e.g., planarian barium adaptation). These are not merely environment-induced differential developmental outcomes, but the ability to reach specific outcomes despite internal or external perturbations. Beyond the advantages of encapsulated modules and their triggers, the problem-solving capacity of morphogenesis enables the larger system to delegate tasks and rely on them being completed in a wide range of circumstances that do not have to be anticipated and micromanaged. The following are ways in which the MCA potentiates, accelerates, and shapes evolutionary processes (Fig. 8).By providing an agential substrate, MCA speeds the evolutionary search for better solutions in several ways. Active, modular problem-solving capacity enables the following properties for the evolutionary process:oGeneralization. By increasing the number of genotypes that all map to the same functional phenotype, the partially autonomous generative layer helps evolution generalize [35, 50]. When functionality and thus fitness are not tied to a single molecular implementation, significant plasticity and robustness arise via exploration of different ways to achieve the same goal, generating alternatives which might prove to have other advantages.

For example, the instructive property of resting potential (Vmem), distinct from the specific ion levels that underlie it, means that a bioelectrically controlled morphogenetic function can continue unaffected, while evolution experiments with other consequences of swapping out different channel proteins in that circuit.pReliability. Modular agential systems can persist and be continuously improved and built upon, because the ability of the parts to achieve specific design specs enables the other parts to functionally trust that they will accomplish their task even when changes occur. This makes it practical to invest energy in complex regulatory systems. MCA facilitates cooperation (which is not guaranteed by mere genetic relatedness [141]) between components and across levels of organization by encapsulating complex behaviors in a simple module with a specific goal. Systems that have a clear, reliable goal state that they can be depended upon to pursue are easier to cooperate with, and thus present a lower barrier for investment in cooperation.qA more tractable search space. By smoothing the fitness landscape, MCA makes the search process much more efficient, improving evolvability by enabling the search to overcome local maxima. By reducing phenotypic differences between different genotypes, MCA enables selection to explore the consequences of mutations which would otherwise be maladaptive. For example, a mutation which displaces the mouth, but has some other beneficial impact elsewhere, will not result in a dead embryo (and thus failure to exploit the other consequences of that mutation) because the mouth will self-correct [178].rCryptic variance. MCA makes many mutations neutral that would be deleterious in a direct genotype → phenotype architecture.

Such mutations whose effects are mitigated by developmental regulation are not weeded out; instead, they persist, generating over time a rich pool of diversity that may be phenotypically dormant, but can be exploited in future. Cryptic variation [74, 246] has been proposed to underlie adaptation to previously unencountered conditions or environments.sFunctional intermediates. Goal-directedness (in the cybernetic sense) of developmental modules helps resolve the problem of useful intermediates, because the autonomous problem-solving capacity of the parts pushes partial solutions toward an attractor. In On the Origin of Species, Darwin wrote, "If it could be demonstrated that any complex organ existed which could not possibly have been formed by numerous, successive, slight modifications, my theory would absolutely break down." Many scenarios which, in a flat architecture, would result in an in-between state with low fitness, instead, end up in a useful attractor state because of the “do until condition X is met” character of homeostatic and homeodynamic loops.Implications for the inverse problem. An intervening developmental control layer can help mitigate the nonlinear relationship between genotype and phenotype at the micro level by providing a more linear relationship between control signals exerted at higher levels of organization and phenotypes.oMore linear controls. It is likely that a layer of cybernetic competency makes the relationship between genotype and phenotype more linear [20, 247]—addressing the difficult inverse problem of selecting which genes must mutate to reach a system-level outcome. This potentially empowers epigenetic mechanisms that go beyond a completely blind search of the genetic space.

By encapsulating the chaotic (in the mathematical sense) relationships between changing local rules and emergent outcomes into modules with distinct goals, the system becomes easier to control via a combination of top-down encoded goal states and bottom-up emergence.As described above, morphogenetic goal state encodings (such as bioelectric prepatterns) can be re-written despite a wild-type genome. Keeping such goal state encodings functionally orthogonal (independently modifiable) from the machine that executes them is a powerful architecture exploited by computer science. This separation, complementing feed-forward emergence, progressively minimizes the inverse problem at each pair of layers, allowing much more efficient functional architectures.pFacilitating credit assignment between changes and positive outcomes. Evolution does not have to wait until a solution is found that improves one property while not impairing others—it can exploit the much easier route of testing out properties, while the error-correcting competencies of underlying modules mitigate the fact that most genes have many different roles [248]. MCA helps handle the fact that it is very hard for any optimization process to make progress when each change has many different consequences, by reducing the need for determining individual contributions for each control signal.MCA’s modularity confers both robustness and hackability.oMore control at the higher levels, less at the lower. The advantages of robustness (making some components controllable by other components) result in encapsulation of cybernetic modules which are more controllable in specific ways (goal rewriting, behavior shaping), but less controllable in their details (they resist noise and intervention, and sometimes have novel unpredictable behaviors—a hallmark of the primitive roots of agency).

It may be far easier for evolution to exploit complex outcomes by making changes in the setpoint information (e.g., by manipulating ion channel structure) than by trying to micromanage local rules that are many steps away from system-level outcomes. Future work will measure the relative improvement in edit distance [249] between highly fit organisms in bottom-up vs. behavior-shaping evolutionary approaches.pEvolutionary pressure to detect ‘hijacking’. However, the increased controllability, and thus reprogrammability, of components is also a potential target for exploitation by parasites and various cheaters [250], as it makes it that much easier to control complex outcomes in host or conspecific behavior. This in turn sets up an evolutionary pressure for systems to model their own functionality (causality) to detect when their control circuits are being changed internally or externally. This reinforces the boundaries between self and world (individuation) and may have many consequences for cognitive capacities in more advanced life forms.MCA facilitates evolutionary creativity.oPivots to novel problem spaces. Another way evolution can exploit the benefits of the MCA is by changing the measurement component of the homeostatic loops, which generates novel functions (and thus jumps across the fitness landscape) by providing new perspectives—new ways one component (or the larger system) can control and exploit on an existing mechanism—polycomputation [42]. More broadly, useful computational functions, such as associative memory, can be derived from a gene-regulatory network without changing the structure of the network (and thus without any possibility of adversely affecting any dependents).

A mechanism that interprets its outcomes in a particular way (maps specific nodes to the functional elements in an associative conditioning paradigm for example, [88, 89]) is readily evolved and modified, and provides a way for evolution to squeeze additional benefits from existing components without the negative consequences for other internal observers (subsystems), thus keeping prior gains. Future work must examine the role of perception (in the active inference sense) and polycomputing in potentiating evolutionary advances, to see how much of the effort of evolutionary optimization is spent tweaking the measurement, setpoint, or actuation machinery of homeostatic loops.pThe competency ratchet described above [240] and the facilitation of complex high-level control loops by the modularity of the test–operate–exit loops of homeostatic mechanisms go beyond simple problem-solving. They also help evolution address novel opportunities—open-ended [251] exploration of new spaces and selection of novel problems to solve, not just better ways to solve the same problem. One example of this is the ability of planaria to find transcriptional paths to resolve an entirely novel physiological stressor (barium) [90]. Another is the ability of Xenobots made of wild-type Xenopus laevis cells to perform kinematic replication (which, to our knowledge, no other organism does), suggesting that more focus needs to be placed on how allostasis, homeorhesis [252], and autonomization [253] not only guide navigation in an established problem space using developmental motifs (a.k.a., behavioral modules in anatomical morphospace), but also enable pivoting across novel spaces—a hallmark of early agency. Fundamentally, problem spaces and goals exist from the perspective of an observer (a subsystem exploiting a perspective for adaptive advantage).

Thus, it is essential to develop an understanding of the dynamics which enable, scale up, and transform the proto-cognitive/computational boundaries of evolved and designed agents.Fig. 8Summary of the evolutionary perspective on morphological competency. A It is well appreciated that evolutionary dynamics have scaled intelligence, from the primitive capacities of microbes to the metacognition of modern humans. However, it is suggested here that the process is symmetrical: the intelligence of the components (i.e., problem-solving capacity) potentiates the evolutionary process, forming a ratchet that enhances this capacity while it makes it harder for selection to see the structural genome [240]. B It is suggested here that the multiscale competency architecture (the cybernetic nature of each layer) enables evolution to pivot across problem spaces, using the same navigation policies to solve problems in metabolic, physiological, transcriptional, morphogenetic, and 3D spaces (the latter being the traditional space of “behavior”). For example, the (bioelectric) computations of neural networks that control muscles to move the body through 3D space emerged from much earlier usage of those same mechanisms and algorithms to control other cell types [145] to navigate the configuration state of the body through anatomical morphospace despite novel circumstances (examples in Figs. 5,6) [86]. Blue arrows symbolize how evolution scaled up each type of control into a new, larger space. The intelligence-focused lens on the co-developmental process suggests that the competency of the layer between genotype and phenotype (developmental physiology) greatly smooths the rugged (C) evolutionary landscape, resulting (D) in a faster search and exploration for novel capabilities.

The ability of cellular and tissue-level agents to accomplish their tasks in their respective problem spaces despite unpredictable changes enables many mutations that would otherwise have been deleterious to be neutral, as the system adjusts and enables evolution to explore other, potentially positive effects of pleiotropic mutations. Mutation is thus just one of many internal and external perturbations that deviate the body from a target morphology—a ubiquitous stressor that needs to be dealt with on developmental and evolutionary timescales. On this view, regenerative capacity is not a novel solution to external damage, but an ancient, ubiquitous aspect of homeostasis that enables developmental progression, regulative morphogenesis, and also efficient evolutionary search. Panels B, C, D used with permission from Jeremy Guay of Peregrine Creative.

It is no coincidence that Alan Turing was interested in both the question of intelligence in radically unconventional embodiments [254], and the question of self-organization of biological form during embryogenesis [255]. In a deep sense, they are the same question, seeking to understand the ability of matter to support reprogrammable, software-level (physiological) information processing that provides context-sensitive behavior for changing inputs. Turing’s modeling of self-organization in a chemical medium dovetails with the fact that an egg genome does not specify one embryo in a hardwired manner. Cell sheets that express genetically provided hardware are capable of giving rise to 0, 1, or several autopoietic systems that set their own boundaries in the excitable medium of the early blastoderm [256]—the most basic example of the flexibility of the product of the evolutionary process.

Biological systems exhibit cybernetic competencies in diverse problem spaces, which scale up from the most minimal behaviors of molecular components [257–260] via a process of evolutionary enrichment for computational flexibility and controllability. The fact that the substrate on which evolution works is an agential material [37] means that evolution is searching not only the space of micro-level phenotypic outcomes, but also the smoother, lower-dimensional space of behavior-shaping signals for control of modules that encapsulate complexity behind simple interfaces like the homeostatic control loop. This also allows evolution to bridge the sim-to-real gap—a problem that plagues roboticists and engineers, in which the environment is constantly surprising no matter how rich the prior training set was [261, 262]. In overloading [42] the same genetic hardware with multiple interpretations [125, 126, 197], the physiological machinery of chemical and cellular networks ensures that evolution does not over-train on prior data, but compresses and generalizes. Evolution crosses the gap between prior experience and adaptive novelty [261] by not relying too much on a lineage’s prior experience (which, like our simulations, often fail to recapitulate everything we need to know in the future), but rather by exploiting problem-solving competencies of modules at different scales.

The evolutionary process seems to have been enriched for computational capacity, starting at the level of molecular pathways [88, 89, 263]. It is likely that molecular and tissue-level intelligence is favored for the same reason that nervous systems have been so successful: they enable context-sensitive computation that cannot be pre-specified in the protein-level encoding of the DNA. Biology has many examples of remarkable compression of complex events into molecular representations [264–267], and the converse—cross-scale decoding: these need to be better understood. In fact, even living systems that do not seem to offer much plasticity and whose MCA capacities seem very low (e.g., C. elegans) cannot be guaranteed to be as hardwired as they first appear. It is an essential aspect of intelligence that its detection (or lack thereof) says as much about the observer as it does about the system itself [14, 86]; moreover, the vast bulk of research in development and even physiology in model systems has not specifically probed for problem-solving in diverse spaces. It is entirely possible that more sophisticated attempts to uncover such dynamics in novel spaces will yield surprises in model systems that are not known for their MCA capabilities.

How do the molecular events brought on by injecting a frog’s egg with an odorant molecule become expanded into a nervous system architecture which controls muscles to behaviorally seek out that odorant as an adult frog [268]? All of the above suggests that MCA enables evolution to search a simpler, smoother, more powerful space—the space of signals and rewards—rather than the high-dimensional space of individual local hardware states. Examples that support this extend all the way down to the cybernetic properties of chemical pathways, such as learning and habituation [87–89, 269]. By shaping signals that are interpreted by modules with various degrees of memory and decision-making ability, both evolution and engineers can take advantage of the MCA [270]. The ability of biology to move salient information across scales, problem spaces, and mechanistic implementations is likely key to the efficiency and power of evolution.

Homeostatic control cycles may facilitate such scaling of complexity and intelligence because of their functional and semantic modularity: for example, a setpoint can easily be changed from a metabolic (cell-level) metric to a measure of planar polarity alignment (tissue level), while keeping everything else the same. The ability to independently swap out the measurements, comparison function, and setpoints of a homeostatic loop means that a very wide variety of high-level strategies can be readily tested out by evolution. This is the same advantage engineers and end users reap from homeostats: overall function can be rapidly changed without needing many alterations of the underlying hardware by for example rewriting the setpoint (such as when the genetically wild-type planarian’s bioelectric pattern memory is briefly re-set to a two-head pattern by a simple ion channel-targeting stimulus, resulting in a permanent line of worms which reliably build to the new setpoint pattern when fissioning or amputated [194]).

The modulation of the evolutionary process by behavioral intelligence in animals (and science-capable humans) was long predated by a similar process taking place outside the brain: the diverse intelligence of morphogenetic processes operating in anatomical space. The fact that evolution not only finds solutions to specific problems, but also creates somewhat generic problem-solving machines, with multiple diverse (simultaneously existing [42]) capabilities, has many implications for the evolutionary process itself. It facilitates credit assignment for the evolutionary search process, enables exploration and novelty [4, 271], and hugely accelerates the process of increasing complexity. This feedback between evolutionary scaling of intelligence and the acceleration of the discovery of novelty by evolution forms a powerful ratchet [240], which is compatible with the emerging picture of a continuum of basal cognition across the tree of life [14].

This perspective suggests a unifying lens for aspects of evolutionary developmental biology and engineering [272], focused on the idea of hacking the agential interfaces of components instead of micro-managing design. Like parasites, body tissues, and evolution itself, bioengineers and roboticists have the opportunity to develop a science of control by exploiting competencies as affordances, applicable equally to the direct products of evolution and to the second-order creative process in which evolved humans partake. Future appliances in this field will be potentiated by universal APIs that go beyond object-oriented programming to goal-oriented programming to exploit biological architectures and offer similar advantages to designed digital systems (as well as hybrid platforms). A number of future research programs can now be pursued.

With respect to theory, much needs to be done to connect these ideas to the body of work on evolutionary units of selection [273, 274], and on natural induction [47]. Tools need to be developed to identify virtual governors [57, 275] as evolutionary forces and to quantify what spaces evolutionary processes are really searching [86]. Computational modeling and synthetic biological systems are now primed to address the dynamics of cognitive scaling and the impacts thereof on the competency of the search process itself, via computational models that explicitly blend low-level competencies with an evolutionary cycle [240, 276]. Some of the biggest questions concern the latent space of homeodynamic goals at different scales, and where these setpoints come from, when not directly specified by genome or environment [271, 272, 277–279]. It is likely that generic principles of inherent laws of biomechanics, bioelectrics, and information processing [62–64, 132, 280–284] will be key parts of the roadmap to flesh out the bi-directional relationship between multiscale intelligence and evolution.

It should be noted that the evolutionary consequences of agential matter were here explored from the perspective of an addition to what is otherwise treated as a fully mainstream picture of the evolutionary process itself. However, in future work, this perspective should be merged with currently developing alternative views of evolution, to determine what implications it has for those frameworks. Other non-traditional [28, 60, 285, 286] evolutionary dynamics could perhaps also be greatly facilitated by the ability of cells to solve novel problems using transcriptional responses. One example is the ready fit of these ideas into the emerging framework of natural induction [35, 47], which seeks invariants between the evolutionary process and that of learning broadly conceived. Another has to do with so-called Lamarckian effects [28, 60, 286]. For example, consider the case of regulating the small number of relevant genes to resolve a novel stressor, as in the barium-exposed planaria [90]. This provides clear evidence that cells can sometimes solve the hardest part of reversing the central dogma—the credit assignment needed to know which genes to edit to enrich for specific outcomes. Once the relevant transcripts are identified by computational processes at the transcriptional level, making genomic edits at those loci could in principle be done by straightforward, familiar mechanisms (akin to reverse transcriptase, CRISPR, trans-generational chromatin modifications, etc.). Whether this in fact occurs in vivo represents an exciting future direction for experimental research.

Specific impacts of the above ideas will go far beyond evolutionary biology, for example, to applications in regenerative medicine which will increasingly exploit the decision-making capabilities of cells and tissues, in addition to traditional bottom-up molecular rewiring [160]. Beyond the life sciences, engineering is now heavily reliant on evolutionary approaches to design [287–289], and robust future robotics and AI architectures must mimic the multiscale competency architecture to achieve (and surpass) the functionality of natural systems. Importantly, connectionist AI architectures are not only the beneficiaries of these ideas, but are also a key developing toolkit for the emerging field of basal cognition because of the “impedance match” needed between a phenomenon and the tools used to study it. Physicists use low-agency, mechanical tools to observe the natural world, and inevitably arrive at low-agency mechanistic models. Taking full advantage of virtual governors, proto-cognitive modules, and other aspects of the software of life requires tools that recognize and learn to hack these capacities. Agency cannot be directly observed with a microscope, but brains, evolutionary processes, and emerging machine learning tools are primed to detect and exploit it via agential models of control because they themselves are higher-agency systems.

As foreseen by many workers since the beginnings of thought on evolution, intelligence is key to evolutionary change. Recent consilience of a range of disciplines are giving rise to the field of diverse intelligence, which recognizes a spectrum of problem-solving and creative competencies in unconventional, basal media that goes beyond the old dichotomy of “dumb mechanical machine vs. high-level true intelligence”. This finally makes it possible to incorporate a naturalistic, non-reductive, empirically useful conception of intelligence into modern theories of evolutionary change. It also naturally leads to hypotheses that whole lineages, and perhaps even the evolutionary process itself, can be viewed as proto-cognitive, cybernetic agents, by characterizing the competencies of the search dynamics using the language of cognitive and behavioral science such as active inference and Hebbian learning [35, 48, 290–292]. What is becoming increasingly clear is that intelligence is not some latecomer that arrives with the appearance of big brains—it is baked in at the very beginning [293, 294], present at multiple scales of the biological substrate of evolution, and continuously shapes its course in a fluid dance that potentiates all participants.