The Multiscale Wisdom of the Body: Collective Intelligence as a Tractable Interface for Next-Generation Biomedicine

Conventional intelligence results when brains exhibit problem‐solving by a collection of neurons bound together into a network which has memories, goals, and preferences far beyond those of the individual cells. In this sense, the brain is a collective intelligence too. The evolutionary precursor of this remarkable capacity is the ability of all cell networks, not just neurons, to solve problems by navigating anatomical morphospace in the progression from egg to adult. Perturbative experiments in this space, like those done to probe the intelligence of conventional behavior in 3D space, clearly indicate that this is not a hard‐wired process and reveal the ability of cell collectives to achieve their anatomical goals despite interventions that deviate them from their normal path.

A few instructive cases of morphogenetic problem‐solving are shown in Table 1. They have several fascinating aspects in common. First is the notion of anatomical homeostasis: the ability of systems to reach and maintain a specific region of anatomical morphospace despite deviations. These examples indicate that these processes are not entirely open‐loop: large‐scale feedback loops exist that measure distance from a given target state and execute diverse molecular steps to implement those goals (reduce error). If we can accept this evidence for the existence of setpoints and competency mechanisms, it opens the field up to the use of tools from cybernetics and control theory—sciences of physical systems with true goals. Second, they expand the well‐known examples of homeostasis in which the setpoint is a simple scalar (hunger level, blood pH, etc.), demonstrating that networks can store setpoints that serve as complex data structures (like rough morphogenetic specifications). A remarkable fact about homeostatic loops is that they implement valence (desirable outcomes) and preferences [98, 99]. Chemistry does not make mistakes—every chemical reaction is equally correct in following the rules of chemistry. But developmental biology, while consistent with the rules of chemistry underneath, brings in the notion of a birth defect—an outcome in which the system could not reach its target morphology despite efforts to do so.

A kind of creative problem solving enables life to handle novel scenarios by using the tools at cells’ disposal in new ways (a classic definition of intelligence). For example, “scrambled” tadpole faces remodel during metamorphosis to form a normal frog face, and polyploid newts made with multiple copies of the genome, and therefore larger cells, are of normal size because structures such as kidney tubules use fewer cells to make the same overall shape [71, 100]. Most remarkably, when the cells are made to be truly gigantic, a single cell may create the tubule, wrapping around and leaving the needed lumen in the middle. This can be analyzed as a kind of downward causation, in which under different conditions, distinct mechanisms (cell:cell communication and tubulogenesis vs. cytoskeletal bending) are triggered toward the same specific large‐scale outcome. It arises because evolution does not just make specific solutions to specific problems, it makes problem‐solving systems that do not over‐train on prior history, knowing that both environment and genetics will change over time [56, 101]. This enables incredible robustness and noise‐tolerance for development [56] and has fascinating implications for how we think about both evolution and applied biomedicine.

Conventional cognitive systems operate in problem spaces, and with goals, unknown to their parts. When a rat learns to press a lever to receive a reward, no individual cell had both the experience of interacting with the lever and that of getting the treat. The associative memory belongs to the collective, which is more than the sum of the millions of cells that comprise it. What underlies our memories, plans, and preferences is an electrochemical network that binds the functions of cells such as neurons into an emergent whole with novel properties. These networks comprise cells running electrical circuits determined by ion channel proteins (which set resting potential of each cell) and electrochemical synapses such as gap junctions (which determine how changes in those potentials propagate across the network).

One of the key properties of such networks is the storage of memory because voltage‐gated ion channels provide a kind of historicity in which transient physiological stimuli can induce long‐term changes in the bioelectrical property of the circuit [102]. Memories are ideal keepers of the setpoint information used by homeostatic behavior. Another key feature is top‐down control, which enables stimuli to kickstart complex downstream autonomous cascades. The most widespread application in bioelectricity—the cardiac defibrillator—works because it is possible to provide a stimulus and depend on the organ to take it from there. In terms of behavior, it is essential that simple stimuli can trigger multistep behavioral responses, including autonomous homeostatic loops that perform actions until specific conditions are met.

The final crucial thing about bioelectric networks is that they allow information to cross levels of organization and be remapped across problem spaces, for example between linguistic space and 3D motion space in active behavior. A human being's top‐level career and interpersonal goals are effectively pursued because those goals result in the movement of ions across muscle membranes that enable the organism to move. This ability of mental structures to control biochemistry is not limited to rare, exotic forms of biofeedback and placebo effects—such “mind‐body medicine” is the everyday miracle of voluntary motion, made possible by the bioelectric network that transduces high‐level mental patterns into the action of muscles and glands.

While it is tempting to think of the cross‐level transduction as a unique capability of neural hardware and the electrochemical software that it enables, this architecture is ancient, being present in microbes [103]. Evolution discovered the immense benefits of electric networks by the time of bacterial biofilms [104, 105], using them to integrate physiological information across space and time in a colony. What did pre‐neural electrical networks think about before nerves and muscles evolved, enabling conventional behavior? Networks that managed an organism's position in 3D space evolved from non‐neural somatic precursors whose job was managing the navigation of the organism through anatomical morphospace [13]. Just as bioelectricity in the CNS binds neurons into a collective intelligence for motile behavior, somatic bioelectricity functioning from the time of fertilization binds all cells into a collective intelligence that solves morphogenetic problems.

Thus, brains and CNS function are the result of a fascinating evolutionary pivot. What changed was the space these networks represent and manage, and the time scale at which they operate (from the hours and days of morphogenetic change to the milliseconds of motile behavior). What stayed constant was the molecular machinery: ion channel proteins, electrical synapses (connexins), and neurotransmitter downstream targets that eventually regulate gene expression (Figure 3). Also, many of the algorithms (such as active inference [97, 106], perceptual multistability [107], dynamic rewritable memory [108], etc.) are highly conserved between morphogenetic and cognitive functions, enabling tools of computational cognitive science to be used in developmental biology contexts [15, 96].

Thus, the symmetry between developmental biology and neuroscience is deep [96, 109]. It can be seen in numerous channelopathies that result in patterning defects (see Table 1 in [110], and [111, 112, 113]), and in the applicability of the tools of neuroscience—from drugs to optogenetics to training protocols—to morphogenetic decisions.

It has been clear for over a century that endogenous bioelectric phenomena play a functional role in the control of dynamic anatomical outcomes [114, 115]. But the development of molecular tools to read and write bioelectric state information in non‐neural tissues has led to the identification of the native genetics underlying the circuit properties, the molecular targets of the voltage change, and the processes that exploit bioelectric networks as a self‐modifying control network [116]. Focusing on spatiotemporal patterns of resting potential (complementing older work on electric fields and ion fluxes) revealed several types of effects. First is the control of stem cell differentiation decisions [117]. But the importance of bioelectricity really shines at the organ scale because the bioelectric code mapping patterns into anatomical outcomes is not a just cell‐level code.

At the organ level, bioelectric patterns serve as critical prepatterns—informational scaffolds that store rough anatomical setpoints for tissue‐level order [90]. One example is the electric face (Figure 4B)—an endogenous distribution of resting potentials that determines the gene expression and anatomical regionalization of the vertebrate face [118]. Manipulation of this pattern via pharmacology, optogenetics, or ion channel misexpression results in predictable changes to craniofacial development and explains why ion channel mutations lead to such phenotypes in models ranging from frog to human [119]. Recreating specific bioelectric patterns in new locations can also induce ectopic organs; misexpression of potassium channels in Xenopus laevis results in the formation of ectopic eyes [120] by recapitulating the voltage eye spot seen in the electric face (Figure 4C,D). This underscores the fact that bioelectric patterns are both instructive and highly modular—a simple signal induces a cascade of events to build a complex organ. It also reveals an interesting competency of the cellular medium: if too few cells are injected with the channel mRNA to build an eye (Figure 4E), they recruit neighboring cells. We did not have to engineer this ability, the material already does this, challenging us to exploit such capabilities and discover new ones.

A key point revealed by these data was that numerous regions in the posterior of the animal could be induced to form eyes, for example, on the gut. It was always thought that cells outside the anterior neural field in vertebrates were not competent to build eyes. But prior studies prompted cells with the so‐called master eye gene Pax6 [123]. It turns out that a higher‐level signal (V mem change) can induce eyes almost anywhere in the body. Thus, higher‐level prompts can reveal new capabilities not apparent from manipulation of molecular levels.

The dynamic aspects of this organ‐level reprogramming have additional biomedical implications. At early stages in embryos injected with the potassium channel, many ectopic eye spots are detected, but just one new eye tends to appear at the end. This is the consequence of conversation between morphogenetic agents with different goals. The channel‐induced cells have been pushed toward a morphological setpoint corresponding to eye development, and try to convince neighbors to assist, via signals yet to be identified. Neighboring uninjected cells, whose morphogenetic setpoint is still “gut,” are sending signals to maintain the normal fate. This bidirectional communication to implement contradictory, physiologically specified goals provides important biomedical targets for challenges including birth defect repair, limb regeneration, and cancer suppression.

Cell collectives’ ability to read bioelectric state information can be used to repair birth defects in vertebrate embryos (Figure 5). Normal brain morphogenesis is determined by a specific bioelectric prepattern that is altered by chemical or genetic teratogens [124]. Forcing a return to the correct prepattern in the neural plate can correct brain morphology, gene expression, and learning capacity in animals exposed to alcohol, nicotine, or even mutations of the critical neurogenesis gene Notch [124, 125, 126]. Thus, at least some hardware defects (such as a dominant Notch mutation) can be fixed “in software” by a brief induced bioelectric pattern. The induction of this pattern does not require individual micromanagement of voltage state at every cell in the relevant region. Instead, a voltage‐sensitive ion channel—HCN2—can be activated, which causes different changes in depolarized and hyperpolarized cells. In effect, it is a “sharpen filter” for the bioelectric pattern that was blurred by the teratogens, establishing crisp lines between developmental compartments and leading to normal morphogenesis. This context‐sensitive property of HCN2 is the first step toward interventions that push complexity off the scientist and onto the system itself, prompting a primitive form of decision‐making within the system to communicate a complex goal.

The third mode of bioelectric network function encapsulates complex organ‐building cascades and binds them to specific voltage state triggers. One example is a very simple sodium flux to trigger the regeneration of a tail or limb in scenarios where they would not normally regenerate [128]. Another is the prepattern that determines the polarity of the planarian head‐tail axis [129]. Altering the bioelectric pattern for just a few hours can convert regenerating planaria into a 2‐headed form [130, 131]. Remarkably, the bioelectric circuit is a true memory because once flipped into a 2‐head state, it is permanent: 2‐headed worms will continue to regenerate with two heads in perpetuity without further manipulation [132]. This is not reflected by any genetic changes: it is an example of non‐genetic inheritance of morphology (previously shown only in single‐cell organisms [133, 134], though it also occurs in trophic memory of deer antler regeneration [135, 136]). Beyond reiterating the capacity for modular control, the two‐headed planarian is an example of rewriting the memory information that encodes the target morphology.

Cancer is a highly complex, heterogeneous systemic disorder [137, 138, 139, 140]. Current approaches focus on genetic damage—a perspective in which irrevocably broken cells must be killed via application of toxic chemotherapies or immunotherapeutics [1]. This focus on cell cycle checkpoints and dysregulated molecular pathways predicts that animals with ready access to large numbers of plastic, undifferentiated, proliferative cells should be especially prone to cancer. In fact, it is often speculated that human bodies’ limited regenerative potential is an evolutionary tradeoff to limit the cancer burden in such a relatively long‐lived animal. But if so, what to make of the fact that animals such as planaria and salamanders are both highly regenerative and cancer resistant? An important clue to this apparent paradox is provided by the striking observation that regenerative [141, 142] and embryonic [68, 143] environments can reprogram cancer.

Perhaps a better question would be not why cancer occurs, but why is there ever anything but cancer—why do cells cooperate to build complex structures in the first place? Focusing on the initial state of all cells as highly proliferative unicellular organisms leads to a view of cancer as a breakdown of the mechanisms of multicellularity [144, 145, 146, 147]. But it is more than simply growth inhibition by neighbors. A recent theory of cancer [148, 149, 150] focuses on the concept of the cognitive light cone: the size of homeodynamic goal states that any active system can pursue. Consider an amoeba: all of its physiological, transcriptional, and metabolic goal states are limited to a very small spatial diameter, with very limited memory and predictive capacity (time horizons). Everything this cell does is in service to tiny goal states, and everything outside of this horizon is considered external environment, at the expense of which the cell may survive. What happens during evolution and developmental morphogenesis is an enormous scaling of this cognitive light cone. The cells belonging to a salamander limb are working on grandiose goals—achieving a complex structure of a highly specific size and form. It is a goal in the sense that when deviated from this attractor, the cells will rapidly work hard to get back to it (by rebuilding) and then stop when it is achieved. Compared to the amoeba, the salamander limb cellular collective's goal is much more complex, with a much larger spatiotemporal horizon.

Thus, we can view cancer from the perspective of basal cognition as a pathological rolling back to the smaller cognitive light cones of the deep past. This has already been confirmed for transcriptional profiles, as proposed in the atavistic theory of cancer [151, 152, 153], and has been extensively discussed in the literature addressing the inadequacies of the mutation theory of cancer [154, 155, 156, 157, 158]. But more specifically, the cognitive light cone model suggests that cancer cells are not more selfish; they just have smaller selves. They begin using action loops with much smaller local goals, in effect shrinking the boundary between self and world, reverting to an identity where the rest of the body, including neighboring cells, is treated as an external environment, outside the set of variables that must be maintained in desired ranges.

Many things, including but not limited to mutations, can kickstart the transformation process in which cells physiologically disconnect from the tissue‐level network. For example, a long period of stress can cause cells to close off gap junctions to prevent bystander toxicity effects. Each reduction in physiological connectivity makes it easier for cells to perform computations at the local level. This makes it easier for cells to close off gap junctions even more, driving a positive feedback loop—a cycle of dissociation and isolation that makes it ever harder to achieve large‐scale cell collective goals (maintaining complex anatomical structure). This picture is consistent with the known early steps of transformation involving gap junction closure [159, 160] and the seemingly paradoxical tumors induced by geometric barriers between tissues even when those barriers consist of materials that themselves are not carcinogenic [161, 162, 163].

This “changing boundary of the Self” model reinforces the theme of divergence between the DNA‐specified hardware and the physiological software that drives outcomes and makes strong predictions for a research agenda. If cancer is a failure of cognitive glue mechanisms that normally bind cells to common paths through morphospace, then targeting these mechanisms should enable: (a) detecting incipient cancer by monitoring cell physiological connectivity; (b) induction of cancer in genetically normal cells by physiological stimuli; and (c) normalization of cancer despite genetic defects. As described above, bioelectricity [21, 112] functions as the cognitive glue that scales up the cognitive light cone for navigation in anatomical space by cell collectives. Ion channel genes are increasingly recognized as oncogenes and ion channel drugs as potential electroceuticals [164, 165, 166, 167, 168, 169, 170, 171, 172], with increasing recognition that bioelectrical parameters are important in cancer initiation and metastasis [173, 174]. It has now been shown that (a) cancer induced by human oncogenes in vivo can be detected early by bioelectrical imaging [175], (b) transient perturbation of bioelectrical and serotonergic signaling among cells can induce a melanoma‐like phenotype in the absence of carcinogens, oncogenes, or DNA damage [176, 177, 178], and (c) tumors induced by powerful human oncogenes can be prevented and normalized by managing their bioelectric state [175, 179, 180, 181]. This ability to reinflate the cognitive light cone of cells shows that physiological information processing, not genetic hardware, dominates outcomes and suggests a clinical roadmap quite different from the chemotherapy that dominates the field today [165, 168, 182].

The competencies described above reveal new fundamental principles for biomedical discovery. Biomedicine becomes more about communication and collaboration with an unconventional agent and less about rewiring the molecular hardware to force specific phenotypic states. Crucially, categorical distinctions between molecular pathways (as “real” hardware targets) and cognitive content (memories, world models) are dissolved [54, 101]; ancient questions on the relationship of mind and matter are not merely philosophical but of immense practical urgency. While neuroscience has long grappled with the rich spectrum between organic versus psychic disease, somatic medicine has too long grappled under the assumption that all its challenges can be resolved at the organic end. Below are a few examples of specific research programs that could extend and distinguish future medicine from the status quo by better navigating the space of the possible in ways not accessible to purely bottom‐up design of interventions (Figure 6A,B).

One feature of integrated large‐scale selves is that control signals are not local; the benefit of bioelectric information networks, from bacterial biofilms [105, 189] to brains [190, 191, 192, 193], is that they integrate across space and time. The causal structure of multi‐scale order in biology often means that diagnostics and interventions can be deployed far from the cells in question. Examples include planarian regeneration [194], the rapid change of voltage profile of a knee joint when the contralateral leg is amputated (Figure 4E, [122]), and the ability to control brain patterning [127, 195] and tumor incidence [180, 181] by modulating the bioelectric pattern of cells on the other side of the body.

Importantly, integrated control systems go beyond the typical examples of cells integrated into tissues. For example, the hypothesis of scale‐free biology [196] has led to the discovery that individual embryos in large groups form a kind of “hyperembryo” with its own unique transcriptome and an ability to solve teratogenic challenges far better than individual embryos (or small groups) exhibit [197]. Other recent work has identified social metabolic control of immune cells [198]. Short‐term opportunities targeting these kinds of phenomena, for example, by learning to fake the cross‐embryo morphogenetic assistance effect within one organism and deploying it in biomedical settings, are just a part of a much bigger effort of learning to communicate with, and thus control, larger levels of organization that are not apparent from molecular‐biology perspectives [14, 15].

Another aspect of biological agential material that simplifies therapeutic targeting is their context‐sensitivity with respect to interpreting stimuli. For example, tail regeneration is induced in tadpoles by bathing the entire animal in the bioelectric modifying drug, because only the wound cells are paying attention to the signal [128]. Similarly, only ectopic optic nerve (and not the normal nerves in the head) respond to drug‐based signals present systemically [199]. Even more remarkably, the exact same electroceutical—monensin—triggers tails in tail wounds and legs at leg wounds: the specificity in all of these cases is in the cell collective that responds to the stimulus [128, 200]. Thus, the intervention does not need to micromanage the spatial properties of the reagent, if we understand the circumstances under which the cells know what to do.

More broadly, the ability of cell groups to navigate anatomical and transcriptional space toward specific, encoded goals suggests that one powerful way to achieve complex goals is to rewrite the setpoints, as described above for planarian regeneration. Modifying setpoints as opposed to forcing specific states avoids the compensatory regulation described above for the multiple ectopic nascent eyes. This approach gets the system's “buy‐in” as to the correct state, avoiding triggering resistive responses because it makes the intervention appear as if it originated within the system itself.

While prepatterns of resting membrane potential are very convenient control knobs for complex downstream modules [90], bioelectricity is not the only modality. Proto‐cognitive competencies exist below the cell level, in the learning and optimization capabilities of molecular networks; thus, precisely timed stimuli (a.k.a., dynamiceuticals [201]) offer the possibility of benefitting from, or shutting down, cellular memories of past physiological events [23, 24]. Thus, the roadmap of the future includes not merely drug discovery, but behavioral discovery—we must learn to search the space of behavior‐shaping signals that motivate cells to adopt desired phenotypes. Many drugs eventually fail because the system finds a way around the intervention that seemed efficacious in 1st‐order modeling [202]. Second‐order interventions include: patterns of stimuli designed via behavioral control strategies (drug conditioning, training molecular and cellular components for desired outcomes using reward and punishment) [17], drugs that target homeostatic control loops, such as perhaps semaglutide [203, 204, 205], and anti‐cancer strategies that seek to reinforce cellular connections to the collective [149, 206]. Third‐order interventions could include psychedelics or plastogens to directly control the self‐models (and their plasticity) in tissue contexts [207, 208] and nootropics to help cells find better solutions to stressors.

Beyond new ways to use drugs to communicate with cells lie more complex agential interventions: reagents that are themselves context‐sensitive. This includes the HCN2 channel described above, which can be deployed systemically because the channel itself distinguishes cells in different physiological states and thus can sharpen voltage prepatterns (leading to a repair of birth defects) by selectively acting on depolarized cells [125, 126]. The next level of agential interventions are cellular constructs such as patient‐derived personalized biobots (Figure 6C), which can move autonomously and dynamically interact with their living or abiotic microenvironment, with context‐sensitive and tissue‐hacking competencies remaining to be discovered. These have been shown to induce repair of neural cells in vitro [188] and bile ducts [209]. The use of bespoke, personalized, patient‐derived biobots to exert repair within the body without need for genetic editing or immunosuppression is a major opportunity for biomedicine, as these share with the body a molecular understanding of inflammation, damage, cancer, microbiome, and health. Cell‐based biobots inherently contain a myriad of sensors and decision‐making machinery accumulated during a billion years of evolution and far beyond anything our nanotechnology can produce today.

One thing that needs to be expanded is the notion of “disease” beyond specific molecular markers/events to an understanding of undesired systemic states as learned attractors and dynamic patterns within physiological and other spaces [210, 211]. For example, the apparent inability of limbs to regenerate without neurons is learned—it is not an innate limitation but rather a kind of “nerve addiction” [212, 213, 214]. What other disease states and limitations of healthy functioning are the direct result of learned priors by cells and molecular networks?

The next stage of advances will go beyond conventional dynamical systems theory approaches to systems medicine by including not just a view of network states as passive features of a complex landscape but as patterns within a proto‐cognitive system that might themselves have minimal decision‐making and computational competency to facilitate their own persistence (in the same way that depressive and repetitive thoughts enact niche construction on the neural hardware of the brain to make it easier for such thoughts to exist and amplify [215]).

Given the many kinds of persistent, coherent, self‐reinforcing patterns of energy and information seen in dynamical systems theory and physics (solitons, autowaves, etc. [216, 217, 218, 219]), could some disease conditions be those kinds of persistent quasi‐objects—regions of transcriptional, physiological, or anatomical space that exist and exert causal power despite the fact that they are not classical objects? These have been found in physiological media, for example, as domain walls [220]—bioelectrical patterns propagating through homogenous tissue, and the bioelectric prepatterns discussed above are examples of these in vivo, as are “mirror foci” in the brain—epilepsy‐triggering physiological states that exist in a brain hemisphere opposite from the one that actually sustained damage [221, 222, 223, 224, 225, 226, 227]—a pernicious natural process that is the converse of bioengineers’ attempts to repair bioelectric state by modulating remote regions [127, 180, 181, 195]. Packets of stress, setpoints, self‐models, and estimates of safety may be agents in our body that live in physiological, metabolic, and transcriptional state space, impacting health and disease as their mental counterparts do for mental health.