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

The current paradigm for understanding morphogenesis focuses largely on ideas from emergence and complexity theory. It is clear that some local rules, executed repeatedly and in parallel by large numbers of cells, can result in an emergently complex structure (Fig. 2). This is true, and we now have good mathematical formalisms for understanding emergent complexity. However, this is a limiting paradigm because it is entirely feed-forward, or open loop: causality is thought to move in only one direction (from the molecular details upward) [158, 159], and thus all modifications have to be made at the level of the rules. Because the inverse problem is not generally solvable, this view locks in a model of evolution limited to searching the genotype space (which may be extremely rugged, due to its nonlinear relationship to the phenotype space). It also limits workers in bioengineering and regenerative medicine to exclusively targeting the molecular hardware in hopes of improving system-level outcomes (which in turn results in the difficulties with drug discovery [160]).

Fortunately, there are now tools available to begin to think in different ways about morphogenetic control, which facilitate new research and make possible a new roadmap for discovery of intervention policies, both by bioengineers and by the evolutionary process. The first is cybernetics [99, 161]: by emphasizing the information-processing capacities of multiscale components of living systems (with all of the attendant implications of circular control, multiscale causality, etc.), it becomes possible to recognize the reliability of morphogenesis as a consequence of the goal-directedness of underlying processes. It is essential to abandon the traditional scientific teleophobia [162] because cybernetics and control theory now provide a mature, naturalistic, quantitative science and engineering approach to mechanisms with goals. Goals do not require magic, they require a specific causal structure [163, 164], which is ubiquitous and not restricted to complex brains. The critical importance of the cybernetic worldview for developmental biology has been recognized already [7–9, 165–167]. Additional tools are provided by advances in information theory which have now rigorously shown that higher levels of organization can have causal power and drive lower-level events in the only sense that matters: by serving as the most effective control points at which to make changes in the system [168–172]. This is as relevant for regenerative medicine as it is for evolution, and is becoming a very useful lens through which to relate to biological systems.

The crucial transition is recognizing that homeostatic cycles, the atoms of cybernetic systems, are not merely feedback loops (which are widely accepted as ubiquitous in biology), but are the first rung on a spectrum of intelligence [99]. Intelligence is used here in William James’ definition, not limited to advanced metacognition in primates. The field of basal cognition [53, 54, 173] seeks to unravel the evolutionary origins of the brain’s remarkable trick—unifying the activity of millions of cells (neurons) toward a common purpose in behavioral space. By examining the transitional forms that smoothly continue all the way down to microbial molecular networks, a wider understanding of diverse intelligence has had to be forged [14, 55]. From this, the view of the brain as a collective intelligence has been enlarged to understand the morphogenetic transformations of the body as a collective intelligence of cellular swarms, which solve problems in other spaces [39, 86]. Thus, the robustness of development is not of first order (do the same thing reliably each time), but of higher degree (achieve the same target morphology, by various means, despite various perturbations).

Taking evolutionary continuity seriously and applying concepts from cybernetics reveals an important invariant: navigation in problem spaces as a common set of capacities that range across a continuum (Fig. 5) of competency. Different degrees of intelligence are thus cached out in engineering terms as abilities to overcome barriers in the problem space, in a wide range of diverse inorganic, organic, and neural-level systems. This generic formulation enables recognition, prediction, and manipulation of systems across a continuum that covers all the intermediate levels of sophistication that lie between two magnets’ ability to get together and Romeo and Juliet’s ability to get together [14, 86]. But, these competencies cannot be definitively ascertained from observations of the default course of morphogenesis, which obscures the ability of living tissue to handle novelty (of both, external environment and internal composition) and lulls the observer into a limited expectation that genomes code only for specific outcomes and no more. A very rich set of examples belie this misconception and instead support a view of morphogenesis as a goal-directed, homeodynamic process [11, 174].Fig. 5Examples of morphogenetic and functional competency. A Many embryos, including human, can be split in half at early stages and give rise to multiple normal whole bodies (monozygotic twins), making up for very drastic damage to nevertheless achieve the correct target morphology. B The same is true of regeneration, in which for example a salamander limb can be amputated at any level, and the cells will rapidly rebuild exactly what was missing, stopping when the correct target morphology has been achieved. C A remarkable example of top-down control occurs when newt cells are made to be larger than normal, by forcing polyploidy [181, 184].

The same overall kidney tubule structure and diameter results, despite the incorrect details at the lower level (excess genetic material, and wrong size of building components), as the cells adjust their number to the new size—this happens in real time (developmental timescale) not requiring evolutionary changes in the genome. Most amazingly, a different underlying molecular pathway is triggered (cytoskeletal bending, vs. cell–cell communication) when the cells are made so large that the only way to build the structure is for one single cell to wrap around itself, leaving a topological lumen. D The plasticity is not only structural, but also functional. When eye primordium cells are transplanted to the tail, they not only make a correct eye despite abnormal environment, but the resulting animals (which can be made to lack primary eyes) can efficiently see out of these eyes, which do not connect directly to the brain [176]. This does not require evolutionary adaptation to the new sensory system architecture, illustrating the amazing ability of body components to make up for unexpected changes induced by mutation or other means: a property that has huge implications for the evolutionary process. Panel A photo by Oudeschool via Wikimedia Commons. Panels B, C by Jeremy Guay of Peregrine Creative. Panel D by Douglas Blackiston

The most obvious examples are seen in regulative development and regeneration [57, 175], where cells work to implement and maintain a large-scale form (target morphology) despite surgical, genetic, and physiological sources of defects. But it goes much further than that. Tadpoles, in which the native eyes are prevented from forming and an ectopic eye is instead placed on the tail, can see and perform well in visual behavioral training [176], even though the ectopic eye connects to the spinal cord (or just to peripheral tissue) rather than to the brain—this radical change to the sensory–motor architecture does not require generations of adaptation to produce successful behavior. Tails grafted to the flanks of amphibians slowly turn into limbs [177]—a structure more appropriate to the large-scale target morphology (even though, for example, the local environment for the tail-tip cells is perfectly correct); and craniofacial structures that start out in odd positions are often corrected [178]. Frog skin cells, forbidden access to the normal reproductive and other capacities of an entire embryo, nevertheless create a functional proto-organism that makes copies of itself through kinematic replication, a mode unknown within the natural tree of life [125]. Perhaps the most remarkable example of this is Slijper’s goat [179], in which the effort of trying to walk upright (due to lack of forelimbs) drove, in one generation, many of the anatomical and physiological changes usually thought to require long periods of evolutionary adaptation to bipedalism.

These kinds of abilities to achieve target morphology are not unique to damage-repairing programs that evolved specifically to handle injury. They occur at all scales of organization, suggesting they may be a general capacity arising from our microbial ancestors’ need to respond to changes in their environments and in their own genomes. For example, mice with mutated semaphorin proteins still produce (through a novel path) correct connectivity of migrating thalamocortical connections [180]. Even more radically, animals artificially made with extra copies of their entire genomes, which produces huge cells, still exhibit normal size [181–183]. In the kidney system of polyploid newts, the normal cell:cell communication that creates a tubule from eight to ten cells working together is swapped for a single enormous cell that wraps around itself to create the same final outcome [184]. This is a good example of downward causation, because in the service of a large-scale anatomical goal, different molecular mechanisms (cell–cell communication vs. cytoskeletal bending) are deployed as needed for an unexpected change (a drastic increase in the copy number of all genes). The ability to handle, without periods of evolutionary change, an unexpected alteration in something as basic as one’s own genome and cell size echoes the theme that has been seen in the study of behavioral learning and evolution [36, 185]: degrees of behavioral intelligence can produce adaptive outcomes on timescales that are much faster than evolutionary search for hardwired phenotypes for each problem, and on landscapes whose ruggedness is too high for the basic evolutionary search process.

Seeing development, metamorphosis, regeneration, and cancer suppression all as manifestations of an anatomical homeostasis loop (Fig. 6) raises the obvious question of how and where the setpoint (target of homeostatic error minimization) is stored. Having been predicted as far back as the 1940s [186, 187], it has recently been found and manipulated using the same concepts and tools that have allowed neuroscientists to manipulate goal-regulating behaviors in living brains.Fig. 6Pattern homeostatic properties of the morphogenetic collective agent. A The normal process of rearrangements of the tadpole face to make a frog face is not hardwired to specific movements. Embryos in which the craniofacial structures are scrambled also make largely normal frog [178], which shows that these components form an error-minimization system for the correct target morphology and has the competency to reach it even when the initial positions are incorrect. B This kind of ability to navigate anatomical morphospace with policies beyond hard-coded actions (see the scale of possible competency levels in Fig. 2D) adds an important component to the mainstream open-loop scheme of traditional gene regulatory networks driving interactions, which result in emergent complexity at the anatomical level. The view of the cellular collective as a cybernetic system, which represents goal states for morphogenesis (see their direct visualization in Fig. 4D,E), enables specific research programs to detect, decode, and re-write target morphology information without needing to rewire the genetic hardware—an attractive prospect for regenerative medicine [57, 58, 303].

C The concept of morphological homeostasis during regenerative repair could be extended naturally to the broader concept of morphological homeorhesis, in which developmental progression is a collection of regenerative repairs [57]: each stage is in effect a “birth defect” from the perspective of the subsequent stage and is “repaired” by regulative development which seeks to minimize error (i.e., system-level stress) relative to the bioelectric target morphology. This perspective hypothesizes that the bioelectric target pattern changes more rapidly than the transcriptional and anatomical patterns, pulling them along by stress-minimizing loops that have been scaled up to act on larger metrics than single-cell stress states [304]. Maturation and adulthood result when the bioelectric and biochemical prepatterns stop changing and the anatomy essentially catches up and future changes are small-scale maintenance and resistance to aging and carcinogenic defections. Panel A taken with permission from [178], and courtesy of Erin Switzer. Panel B by Jeremy Guay of Peregrine Creative. Panel C by Brenda B. de Groot

Evolution broadly exploits bioelectric networks for the same reason they underlie cognition in the brain and computation in our information technology: they have some very attractive properties with which to build generic learning and problem-solving systems [58]. All cells participate in electric networks, which allows them to compute and increase the size and complexity of simple homeostats (Fig. 7). Neurons speed-optimized the process and took advantage of direct long-range connections, but other than that, many of the tricks that brains use to handle novelty are in fact prevalent across the body, long predating classical neuroelectricity on both evolutionary and ontogenic timescales [145]. Many aspects of connectionist models of neural networks’ computational capabilities apply to developmental events [34, 46, 58, 120]; for example, regeneration is a kind of pattern completion in anatomical space, and stochastic phenotypes arising from the same genome can be studied as a kind of perceptual bistability [106].Fig. 7Scaling of the morphogenetic agent. A The components of a complex morphogenetic system can be thought of as cells with minimal computational abilities, for example the homeostatic capacity to keep cell-level parameters (e.g., pH) within certain levels. B For the collective intelligence of morphogenesis to competently navigate anatomical morphospace, in addition to cell-level physiological space, these subunits have to connect into networks that computationally store much bigger goal states (e.g., rough representations of coordinates in anatomical space, or target morphology shapes). It is hypothesized here that this occurs via gap junctional dynamics, but the details of the scaling of setpoint states are an active area of current research [295, 305–310].

Of course, biochemical and biomechanical modalities are also very likely to be involved in the scaling of competencies from subcellular to multicellular problem spaces. C In effect, the scaling enables cellular collectives to enlarge their “cognitive light cone”—the spatio-temporal scale of the target states they are able to pursue, increasing their memories of the past, ability to predict future states from past experience, and the spatial extent over which they can measure current states [13]. D Current work is ongoing on the open problem of linking these cybernetic perspectives to the biophysics of the state space of the electrical circuit of a multicellular tissue, and to central ideas in connectionist machine learning which show how a network (excitable medium) can store large-scale patterns that are robust to perturbation and can be recovered after damage [57, 58, 270]. E The consilience of computer science approaches to cognition (connectionist models of memories in networks), dynamical systems views of systems navigating landscapes, and developmental bioelectricity provide an emergent unification of the notions of memory and recall in morphogenesis and behavior. Panels in C,D,E by Jeremy Guay of Peregrine Creative. Embryo panels in D used with permission from [311]

Bioelectric networks mitigate the inverse problem by serving as the “hidden layers” of a control network linking genetics to anatomy (Fig. 1). The relationships of ion channel genes to the bioelectric pattern, and bioelectric pattern to its resulting anatomy, are each easier to invert than the direct genotype–anatomy link, so they help to decompose the control problem into two easier problems. This not only helps natural evolution, but also helps engineers design and evolve complex proto-cognitive circuits, and increases control capacities for workers in regenerative medicine. Bioelectric networks provide modularity (triggers of complex subroutines, such a simple voltage state that triggers the “build an eye” or “build a leg” subroutine [188])—a known component of evolvability [3, 189]. They also provide an important kind of coarse graining, since voltage is a high-order parameter over ion channel gene and protein microstates, and individual ion concentrations: electrogenic proteins can be swapped out as needed (e.g., V-ATPase complex in the tadpole tail can be phenocopied by a yeast proton pump with no sequence or structural homology, because it produces the correct physiological signal), and everything still works as long as the bioelectric state is correct [190]. Like in the central nervous system, bioelectric interfaces that control tissue- and organ-level structure [107] are very attractive control nodes. It is no surprise that viruses, whose genomes are under significant pressure to remain minimal and concise, devote some of that precious space to ion channel genes [191, 192].

Bioelectric circuits also provide another crucial competency: memory, which is a specific instance of their more general capacity for software-level reprogrammability. As in the brain, bioelectric circuits can store information in a way that is readily changeable in the lifetime of the animal and enable the exact same (genomically specified) hardware to have different modes of activity depending on prior experience. For example, in planaria, the bioelectric circuit that controls the number of heads can be targeted with a brief (2-day) physiological intervention that re-sets the state of the circuit to specify “two heads” instead of the default one-headed bioelectric pattern memory. In this case, the exact same hardware (genetically wild-type planarian cells) can harbor multiple representations of the bioelectrically encoded target morphology that guides the cellular collectives during regeneration [193–195]. The question of what determines the number and shape of heads in the planarian is subtle, paralleling the question of what determines behavior. It is genetics in the sense that genomes specify the ion channels needed to form the circuit, but it is not genetics in the sense that the actual outcome is controlled by experience (in physiological and behavioral space, respectively). Evolution could make use of this in precisely the same way as it exploits classical learning (brain bioelectric dynamics), via genetic assimilation and Baldwin effects [29, 196] that can later canalize adaptive outcomes of morphogenetic and physiological plasticity.

Indeed, it has been shown that existing planarian species’ head shapes are readily recapitulated by a genetically wild-type animal experiencing changes to its bioelectric circuit during regeneration—100–150 million years of evolutionary distance in morphospace can be crossed in a few days because of the dynamics of bioelectric pattern memories [197, 198]. If proven advantageous, this could eventually be transferred into the genome by ion channel mutations that produce the same bioelectric circuit-driven developmental behavior (assimilation of physiological plasticity into the hardware [199]).

As ideal as bioelectric networks are for the plasticity and control dynamics that potentiate evolutionary change, they do not have a monopoly on this role (Fig. 1). Other mechanisms, operating at other scales, provide the same benefits [60]. For example, the richness of cytoskeletal information processing [200–202] enables cortical inheritance in unicellular organisms: there, as in planaria, DNA is not the only source of morphogenetic information. Changes made to unicellular form can persist across generations [203–205]; for example, surgical rotation of a small patch of the cell surface creates a stable line of animals in which a portion of the ciliary beat pushes food out of their mouths—a condition which their perfectly wild-type genome cannot correct. The cytoskeleton serves as a plastic information and control medium stretching (parallel to DNA) all the way back to our last common universal ancestor [206]. The availability of a rapid, stable but re-writable, internal scratchpad is an important component of any system that is to exhibit ability to deal with novelty and to navigate problem spaces in a flexible, dynamic fashion.

Another subcellular component with interesting proto-intelligent competencies is the class of pathways, which includes both protein interaction and gene-regulatory networks. These have been shown to underlie anticipation and probabilistic inference [207–209], six different kinds of memory [87–89, 210, 211], and other capacities that evolution can exploit in the control of individual cell behavior and thus organ-level morphogenesis. An example of transcriptional problem-solving is seen when planaria are cultured in barium [90], a non-specific potassium channel blocker. At first, their heads degenerate (due to the many cells that require K+ flux), but they rapidly re-grow new heads that are completely barium insensitive. Transcriptional analysis reveals that these tissues up- and down-regulate just a handful of genes to get back to physiological and anatomical setpoints despite the inability to regulate potassium. The remarkable thing is that barium blockade is evolutionarily novel—planaria very likely have never encountered the exotic stressor of barium during their evolutionary history, and there have been no selection pressures to develop specific responses to it. Their ability to navigate physiological and transcriptional spaces to solve an entirely new problem is an example of subcellular competencies for novel scenarios (and perhaps, include the ability of generalization from scenarios that they did encounter, such as epilepsy-induced excitotoxicity). Finally, the use of chromatin as a computational resource has been suggested to be an important driver in evolutionary transitions [212, 213].

Increased understanding of the capabilities of cells and their components dovetails with efforts in modern robotics in which morphological computation (via biomechanics [214, 215] and bioelectrics [216]) is heavily exploited—the emphasis is not on developing a specific controller that does all the hard computational work for a fixed body, but in evolutionary approaches that exploit bodies’ innate competencies [217]. By being easy to control, the body does much of the hard work of the evolved controller [96, 218, 219]. This is an important dynamic for understanding how multiscale intelligence shapes evolution: much as we examined existing pathways to find a perspective from which they appear as trainable systems amenable for example to Pavlovian conditioning without any structural change [88, 89], evolution can implement behavior-shaping signals that are equivalent to a control perspective on its substrate that recognizes its agential affordances and seeks to exploit its inherent competencies, not only micromanage the underlying details. Indeed, evolved pathway networks show much more capacity for learning than random ones [88, 89], serving as an example of evolutionary enrichment of such properties.

The evolutionary origins of these competencies are an active area of study: the field of basal cognition seeks to understand how evolution gives rise to increasing competencies in navigating diverse problem spaces [14]. Here, however, the focus is on the second half of the loop (Fig. 1): how do these problem-solving competencies affect the evolutionary process itself? The roles of basal agency in evolution are now beginning to be discussed [77, 220, 221]. Models are being formulated for understanding how each layer of the multiscale competency architecture (MCA) of life deforms the option space for the layers above and below it (Fig. 7). The above-mentioned examples of induced lens cells recruiting others, and of artificially large cells using a different molecular mechanism to complete tubulogenesis, demonstrate behavior shaping and top-down control. We can exploit this in engineering contexts, but it is certain that evolution discovered the power of exploiting these affordances as well. Next, we consider specific ways in which this dynamic drives an evolutionary ratchet for problem-solving capacities and makes sure that the results of evolution are often not specific solutions to specific environmental niches, but rather general purpose problem-solving machines, whose capacities cannot be guessed, or inferred from the invariant, default course of development.

“If your mind is empty, it is always ready for anything, it is open to everything. In the beginner's mind there are many possibilities, but in the expert's mind there are few.”― Shunryu Suzuki

The standard view of evolution emphasizes two things. The first is the cumulative nature of the process—each change is made in the context of prior events and thus is strongly dependent on history. Concepts like ontogenic recapitulation of phylogenetic events stress the fact that developmental mechanisms must operate with whatever components and signals had been established before (were adaptive within the prior historical context of that species). This is true, but there is a complementary aspect, because of the ubiquitous unexpected scenarios that every embryo has to face: environmental changes, genetic mutations, physiological stressors, parasites, and numerous other challenges that cannot be planned for in advance. Thus, biological systems evolve under pressure to remain flexible enough to accomplish coherent morphogenesis despite a wide range of perturbations.

This means that the most successful, robust embryogenic and regenerative processes must not assume prior states strongly. The above-mentioned examples show how little embryos can take for granted. They cannot assume how many cells they have (ability to regulate pattern despite available cell number [222, 223]), the genetic complement and size of those cells (polyploid newts [181, 184]), the starting position of specific structures (unscrambling frog faces [178]), or that earlier molecular signaling steps were completed correctly (repair of mis-steps in the left–right asymmetry pathway [224]). Evolution uses a mix of pre-specified features and “wildcard” phenotypes which are handled by real-time problem-solving [36]. When evolution is seen as a long-term learning process [35, 48–50, 225], it becomes clear that it does not over-train on its priors: past experience shapes development, but for most creatures (with exceptions like the mosaic C. elegans), the mechanisms that survive are ones that are not hardwired to specific expectations but are ready to operate despite a wide range of noise and both internal and external uncertainty [35].

This concept can be encapsulated with the term Play the Hand you’re Dealt (PHD): successful developmental systems are primed to operate with whatever internal and external conditions they are faced as they come into the world (within some broad, but limited, range). PHD is why biological systems are highly interoperable with exotic synthetic materials [226], why huge mechanistic diversity exists within individuals of an evolutionary type [227], and why viable chimeras can be made at every level (from mixing DNA to cell/tissue grafts, to parabiosis [24], to hybrots in which brains can be used to drive novel robotic bodies or live entirely in virtual worlds [174]). This extends to behavior and neuroscience: we are all “brains in a vat”, and our neural systems can readily adapt to sensory–motor augmentation with engineered devices or novel sense organs [228, 229] because of the fundamental plasticity that requires each being to learn about its own form and function in real time. Functional behavior from tail-eyes [176] or third arms [230] or video game embodiment all works because developmental processes are primed to figure out their sensors, effectors, and architecture in real time [231–234].

Genomes not only do not encode large-scale features, but they also do not even exclusively encode machinery that always makes the same large-scale features: they specify cellular hardware that solves problems at multiple scales. This includes dynamically setting boundaries between the self and external world [108], self-modeling [235, 236], and exploiting available problem spaces [86]. In addition to homeodynamic loops, biological systems exploit affordances of the laws of physics and computation to benefit from the learning capacity of network architectures [34, 46, 49, 88, 89], the informational scaling provided by strong cell:cell coupling [13], and the coordinating capacity of competition [141, 142]. All of this could result from the multiscale competency architecture, in which systems from molecular pathways to whole animals (though not being units of selection directly) solve local problems in an active manner with diverse degrees of competency [14]. This allows evolution to readily pivot through problem spaces (Fig. 7), using the same general computational tricks to handle navigation in metabolic, physiological, transcriptional, morphological, behavioral, and linguistic spaces [86]. Like the principal components describing the activity of artificial neural architectures, these spaces are often not obvious to us as external observers. For example, bacteria that can take actions such as physical movement, or transcriptional changes of enzyme expression, to address dissatisfaction with local sugar levels, operate in a hybrid problem space with effectors that operate in what we call gene expression space and 3D space, respectively.

Successful creatures must operate with a kind of “Beginner’s Mind” with respect to the collective intelligence of morphogenesis. There is an interesting parallel here with neuroscience, in which behavior is a combination of past history and on-the-fly decision-making. Much as each organism must interpret its physical engrams in real time to select actions, morphogenesis is a dynamic process that acts by interpreting the genetic material with which it must work. For example, Xenobot cells interpret the same Xenopus laevis genome in a way that results in kinematic self-replication and other behaviors, while standard frog embryos exhibit the more common frog-like interpretation. Like in learning, the ideal memory is not a recording of microstates (which often cannot be reused in other contexts), but a compressed, generative representation of what and how to optimize—second-order policies that are usable even when details change.

Morphogenetic problem-solving capacity must be seen at a level higher than that of cells. A focus on cellular plasticity and cell-cycle checkpoints leads to predictions that animals with easy access to undifferentiated cells (e.g., regenerative ones) should be highly prone to cancer; in fact, the true situation is the opposite [237]. Animals with robust morphogenetic control due to high levels of plasticity are resistant to changes of circumstance (injury) at many scales: they regenerate after loss of limbs and organs (degradation of body-scale information), they suppress cancer readily [238] (degradation of tissue-scale information) and they resist aging (degradation of cell-level information).

The best examples of this counter-intuitive dynamic are planaria, which not only recover their entire bodies from even small fragments, but are very cancer resistant and apparently ageless. This raises a crucial puzzle for the traditional view of genomes as specifying form and function. Why does this extremely “long-lived” animal (greatly outshining animals like humans or elephants, in which long-term cancer suppression is commonly touted) avoid cancer despite consisting of about one-third its cell number as stem cells? Moreover, because planaria often reproduce by fission, any mutation that does not kill the stem cell is propagated into the next generation and expanded, resulting in animals that are mixoploid chimeras with an extremely messy genome [239]. How does the animal with the messiest genome have the best morphological control?

A possible answer to this puzzle merges the above concepts of evolution producing versatile problem-solving machinery. Of course, the problem-solving competencies are themselves produced by genetically encoded hardware, suggesting the view of two kinds of genomic information: that which directly specifies phenotypes (e.g., sequence of protein enzymes, or structural genome) and that which specifies a problem-solving competency (second-order computational capacities). This in turn provides a potential explanation for planaria. In hardwired individuals, in which the genome directly specifies various features of the phenotype, the fitness reveals a lot of information about which genomes are most adaptive: selection can see the quality of the structural genome. However, in individuals in which the morphogenetic process has competencies (such as adjusting the position of a misplaced mouth, establishing functional vision with a tail-eye, or adjusting physiology in light of a stressor such as barium), selection has a hard time picking the best genomes because some of its successful instances are in fact not due to a great structural genome, but to a competency of the parts to adjust. Because cells can navigate morphological and physiological spaces, fitness carries less information about the structural genome.

It has been suggested that when evolution cannot make efficient gains by optimizing the hardwired components, the remaining targets for optimization are the competency mechanisms themselves [240]. This starts a feedback loop, because each gain in competency makes it even harder to judge the structural genome, which exacerbates the drive toward improving competencies—a ratchet for multiscale intelligence that can readily be seen in computational models of the process [240]. Planaria, salamanders (which regenerate, but are not immortal), and mammals all represent different degrees of how far this ratchet has operated in their lineages, because other forces oppose it (e.g., complexity drain [241]). This phenomenon is familiar for example in human evolution, in which case evolutionary pressure for the largest muscles has been lifted, because the most successful reproducers are ones with high computational capacity which use manipulation (e.g., tools, language, and medicine) to increase their reproductive success, making it hard for fitness to select for the ones that are physically the most robust.

This model explains a number of very puzzling observations, beyond the fact that the messiest genomes (400 + million years of somatic inheritance) have the most robust anatomies—a striking disconnect between genomic and morphological stability. For example, it predicts the confirmed fact that there are no known mutant planarian strains with abnormal morphologies (the way there are for fruit flies, mice, etc.) and explains why the research community has had such a hard time (despite decades of efforts) generating transgenic planaria. In this lineage, the ratchet has run all the way forward, optimizing mechanisms to create a functional body (almost) no matter what the genome looks like: all of the effort has gone into polishing a set of algorithms that produce a functional anatomy despite expected noise in the components, which then makes it very difficult to create change by targeting the genetic level. In fact, the only known permanent strains of planaria with an abnormal anatomy are two-headed [195] and destabilized (random 1 vs. 2 head) forms [193], which were not made by genetic change but by manipulating the physiologically implemented software that stores pattern memories and guides planarian tissues in their navigation of physiological and morphological problem spaces. This illustrates the profound link between the course of evolution and the dynamically changing competencies of its substrate material [37]. In the next section, specific implications for the evolutionary process itself are discussed.

“It is not the strongest of the species that survives, nor the most intelligent that survives. It is the one that is the most adaptable to change.”— Charles Darwin