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

The dominant paradigm in biomedicine focuses on genetically‐specified components of cells and their biochemical dynamics, emphasizing bottom‐up emergence of complexity. Here, I explore the biomedical implications of a complementary emerging field: diverse intelligence. Using tools from behavioral science and multiscale neuroscience, we can study development, regenerative repair, and cancer suppression as behaviors of a collective intelligence of cells navigating the spaces of possible morphologies and transcriptional and physiological states. A focus on the competencies of living material—from molecular to organismal scales—reveals a new landscape for interventions. Such top‐down approaches take advantage of the memories and homeodynamic goal‐seeking behavior of cells and tissues, offering the same massive advantages in biomedicine and bioengineering that reprogrammable hardware has provided information technologies. The bioelectric networks that bind individual cells toward large‐scale anatomical goals are an especially tractable interface to organ‐level plasticity, and tools to modulate them already exist. This suggests a research program to understand and tame the software of life for therapeutic gain by understanding the many examples of basal cognition that operate throughout living bodies.

While the current biomedical/pharmacological paradigm has become increasingly effective at controlling molecular‐level events with drugs, true regenerative medicine still eludes us, particularly in the area of anatomical control. We are unable to predict and repair most classes of birth defects or restore damaged or missing organs and appendages. We resort to toxic chemotherapies in the face of cancer [1]. We cannot build complex organs or new biobots to desired specifications and we remain powerless to stop senescence and the degradation of our functionality over time.

While existing physiological interventions are often effective, at least in the initial stages of application, they also have well‐known limitations: pharmacological therapeutics are plagued with differential utility across patients, diverse/unpredictable adverse events, and habituation. In addition, existing drug interventions mostly address symptoms, which often return or worsen once the drug is stopped, or even with continued treatment. The most successful current interventions target “invaders” in the body: microbes (antibiotics/antivirals and parasites (antiparasitics). Few treatments definitively and reliably induce permanent restoration of health by modulating the function of the host organism itself. How do we get to definitive repair?

I see the end‐game of this field as an “Anatomical Compiler.” Someday, we will be able to sit in front of a system and draw the plant, animal, replacement organ, or novel biobot that we want—at the level of functional anatomy. We will be able to specify its geometric structure, and the system will output a set of stimuli that can coax cells to build exactly that. It will also output a communications protocol manual, indicating what further stimuli will control its physiological functions and enable permanent restoration of health states; similar protocols will be usable to maintain and restore the health of native organs in situ. In this vision, the anatomical compiler is not a 3D printer or genomic editing system: it is a translator—a communication device that enables us to offload the complexity of producing the desired outcome onto the cells, communicating and collaborating with the living material by specifying target states for the native machinery, whose fundamental nature is to seek allostatic goals [2, 3, 4, 5]. The ability to communicate structural and functional goals to groups of cells could erase the burden of birth defects, traumatic injury, cancer, and degenerative disease.

Why do we not already have an anatomical compiler platform, or anything remotely like it? Individual successes have been achieved; but—despite the deluge of molecular information and omics data over the last few decades—we are extremely far from being able to exert full, rational control over form and function. This is the consequence of underlying limitations of the molecular biology paradigm, which has served as the exclusive basis for standard approaches in Western medicine [6, 7]. The following are a few examples among many that demonstrate how far we remain from understanding the control of complex anatomies.

Baby axolotls have forelegs, while early tadpoles do not. Despite having sequenced genomes for both species, we cannot predict whether frogolotls—chimeras made of frog and axolotl cells—would have legs or if those legs would be made of axolotl or frog cells or both. In fact, we cannot even predict the anatomy of a non‐chimeric species from its genome alone; we can only generate a rough prediction by comparing it to the genome of a species with known anatomy. While we study the developmental roles of specific genes, we are often surprised; for example, why do cells thrive without highly conserved cell cycle and genome integrity genes and pathways [8]?

Planaria provide another example of where the current paradigm is insufficient [9]. Due to accumulation of somatic mutations over 400 million years of asexual reproduction driven by fission and regeneration, planarian genomes are chaotic and their cells are mixoploid [10]. They also contain huge numbers of highly plastic and proliferative stem cells—a situation normally assumed to imply a high risk of cancer. Despite that, they are champions of regeneration, are cancer resistant, and apparently do not age. This is the polar opposite of the outcome that the current paradigm of “genome drives function” would predict for the animal with the messiest genome, full of undifferentiated cells in the adult stage. Examples like this emphasize how far we are from understanding the regulation of large‐scale properties, even as we drill down into better and better molecular details. This limitation will become increasingly stark as tools such as CRISPR move beyond single‐gene phenotypes and confront the problem of which genes to edit to get a desired complex anatomical or functional outcome.

Here, I argue that the above barriers can be overcome by augmenting today's focus on the hardware of life with a research program that seeks to exploit the inherent physiological software of cells and tissues. Numerous examples of non‐neural memory, problem‐solving capabilities, and collective decision‐making [11, 12] reveal that our bodies are constructed as an architecture in which each layer of organization navigates its own problem space [13]. I suggest that the tools of cybernetics, behavioral science, and neuroscience can be brought to bear on the deep problem of biological control to yield therapeutic and bioengineering solutions [14, 15] far beyond neurons and their control of conventional “behavior” of motile animals in 3D space. Specifically, I argue that the traditional modalities for engineering with passive matter are insufficient in the life sciences; living bodies are an agential substrate, full of competencies and agendas [16], and demand engineering approaches that are new to somatic biomedicine but have been used extensively and successfully in the behavioral sciences [17].

My position is pragmatic and naturalist (Box 1), with two main tenets: (1) we must understand and exploit the mechanisms appropriate to each level, and (2) empirical success in biomedicine and bioengineering should be the only arbiters of conceptual frameworks. At the same time, it is a deeply organicist position because the future of biomedicine and of basic biology requires us to come to grips with what is special about living materials and how they handle the scaling of intelligence across subsystems. Importantly, I am not arguing against the utility of the mainstream paradigm; rather, I hope to show how the focus on molecular events can be incorporated as one approach within a richer and wider set of powerful tools.

A fundamental argument has raged across the biosciences for centuries. The mechanistic approach sees living beings as a kind of machine, focusing on data about molecular components, research agendas that emphasize decomposition into parts, and emergence and complexity science as the key tools with which to predict and control systems [255, 256, 257, 258, 259]. In contrast, the organicist approach seeks proof [256, 260, 261, 262, 263, 264, 265, 266] that autopoietic living things are fundamentally different than machines, emphasizing top‐down causation and control, and unique features of life that cannot be captured by algorithmic models. Which framing is more conducive to the next generation of regenerative medicine? I argue that it is neither, as both take on unnecessary baggage that constrains future discovery by limiting the toolbox that workers in the life sciences can use.

The perspective presented herein could be critiqued from both camps. On one hand, I argue to introduce strong forms of teleology (goal‐driven behavior) and cognitive capacities into molecular and cell biology as well as developmental and evolutionary biology. The use of tools from behavioral science to understand molecular pathways and morphogenesis is squarely against the mechanist project.On the other hand, I also liberally make use of computational tools (notions of software, reprogrammability, etc.) to understand life and uncover novel bioengineering capabilities, which is anathema to the organicist project that holds that such attempts miss what is special about life and mind. In prior work, I have suggested that “machines” and “living beings” are in fact on the same continuum (a ″spectrum of persuadability, Figure 1A), but that the cognitive aspects of this continuum reach way down, at least to the level of molecular networks. This view does not fit comfortably in either of the major camps.

I suggest two main concepts that dissolve the dichotomy between these positions and make them compatible in a way that facilitates future research [22]: pragmatism and pluralism. We should embrace the idea that everything in science (including such mechanist favorites as “pathways”) is a metaphor, and that our goal is the empirical test of which metaphors enable what outcomes. We cannot rely on philosophical commitments and stale categories that were developed in pre‐scientific times to constrain us from importing tools across disciplines. The mechanist/organicist debate [262, 263, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276] rages largely because both camps believe they are describing things as they are. One especially divisive question is whether living things are computers, Turing Machines, and so forth. and thus demarcated by their known limitations. The knot is untied if we understand that cognitivist, computationalist, and other claims are about our formal models, not about the system itself. These are instead engineering protocol claims signaling the intention to use a particular framework to study the system. Nothing, especially living things, is objectively one thing—all we have are a multitude of approaches and metaphors, many of which can be appropriate (useful) depending on context. An orthopedic surgeon should see their patient as a mechanical machine. Their psychotherapist should not. Both are correct. As a key element of context, the notion of an “observer” is central; in biology, observers are scientists but also conspecifics, parasites seeking to hack a system, and the living subsystems of the body seeking to make sense of each other's signals [277].

I propose that it is critical to dissolve binary categories in favor of a continuum hypothesis with respect to cognition and a commitment to uncover the principles of scaling of competencies: not whether something is/is not cognitive, but how much and what kind of cognitive competencies it has that can offer a useful interface. The future lies in telling better stories about how cognition scales and about symmetries of deep concepts applied to novel substrates and levels of organization and size. In service of this goal, we need tools from both the mechanist and organicist toolboxes. Current computational paradigms do not capture everything that is important about life: self‐reference, self‐construction, blurring of the data/machine distinction, allegiance to on‐the‐fly salience of information, and much more. This means we must improve those paradigms for use with some kinds of systems. But other computational constructs—virtualization, abstraction layers, encryption, modularity, software, reprogrammability—are useful, and we should draw on them without fearing that using them commits us to the idea that they provide access to everything.Likewise, the utility of modeling an inner perspective with memories, goals, preferences, and decision‐making competencies is a perfectly rigorous approach to physical systems. The acid test of this framework, as with all mechanistic and organicist approaches individually, is empirical testing. Specifically, not whether each viewpoint can be fitted with epicycles to explain new biological discoveries post‐hoc, but whether it generates new research agendas and leads to the experimental discovery of new capabilities—whether it facilitates questions that could not be asked before.

While I freely make use of metaphors—perspectives on data that suggest next experiments—that have proven effective in areas such as computer and information sciences [18], I do not argue that living material is a computer in the sense that it uses anything like today's mainstream computer architectures, or that the typical Turing Machine model is sufficient to deal with self‐assembling, self‐modifying, multiscale goal‐driven living systems [19]. What I do claim is that our approaches to medical treatment have long been constrained by the implications of four connected, dominant ideas: (1) that the genome determines final phenotypic outcomes, (2) that intelligence only applies to brainy organisms and thus conceptual tools suitable for simple mechanisms at the level of chemistry are the only correct approach to the biomedicine of the body, (3) that goal‐centered frameworks are taboo because goals are the exclusive purview of advanced brains, and (4) that the control of biological form and function must occur via open‐loop emergence of complexity, arising bottom‐up from iterations of local mechanical rules.

The central hypothesis is that the body is a collective intelligence behaving in anatomical, physiological, and transcriptional spaces [13]. Thus, behavioral science tools developed to collaborate with complex systems should provide therapeutic advantages over biochemical micromanagement. Therefore, drugs and other interventions should be developed as communication messages, not low‐level controls. The focus for therapeutic discovery therefore must shift to understanding how proto‐cognitive systems interpret interventions, supporting a strategy of collaboration with endogenous capacities [20] to achieve desired outcomes.

Fortunately, we already have one well‐established modality for interfacing with complex goal‐driven systems: the bioelectric networks that tie neurons together toward conventional cognition also serve as a proto‐cognitive glue for somatic cells as they navigate anatomical spaces [21]. Powerful tools now emerging in other fields can be harnessed to communicate with these networks if we are willing to soften conceptual barriers that prevent the use of toolkits across categories. While it is very likely that other modalities (biochemical, biomechanical, and perhaps biophoton) also underlie similar dynamics in vivo, bioelectricity provides a uniquely tractable interface because of the advances of neuroscience in highlighting the ways in which electrophysiology can implement mind.

Here, I review concepts and data from the emergent field of diverse intelligence relevant to multiple scales in living organisms, frame morphogenesis and regeneration as behavioral outcomes of problem solving across these scales, and discuss how tools coming online in behavioral and systems neuroscience can revolutionize our understanding of molecular networks and cellular collectives as agential materials to enable transformative advances in definitive regenerative medicine, cancer, and senescence.

I define intelligence here in William James’ [22] sense as the ability to reach the same goal by different means: the focus is on problem‐solving competency, which in this case is the ability to adapt to conditions to construct and repair a specific morphology or function. This definition of intelligence is functional and generic, enabling its study in substrates much different than we are used to. It is supported by an emerging body of work on basal cognition and diverse intelligence [11, 15], which finds applications of behavioral science tools to a wide range of systems along the evolutionary path from simple biochemical systems to the kinds of animals in which we recognize this capability in full bloom.

All systems can be categorized along a continuum corresponding to the degree of autonomy they implement (Figure 1A) [18]. It is often assumed that cells and tissues are at the very left of such a spectrum, as physico‐chemical machines which are complicated but low in agency. Accurate placement on this spectrum is critical because it determines the kinds of tools—conceptual and practical—that can effectively be used, and the kind of outcomes that can be expected. Neuro‐behavioral sciences illustrate how the tools for systems towards the higher end of the autonomy spectrum capitalize on inherent system capabilities: achieving a complex behavioral outcome in an animal does not require manipulating every neuron and muscle individually. Instead, we can simply train it because the system itself offers an interface—learning—which serves as an abstraction layer providing a convenient way to encapsulate complex, integrated responses behind simple triggers. Could something like this be possible for morphogenesis?

For bioengineers and workers in regenerative medicine to use this kind of approach to produce specific and complex anatomical forms and functions requires a crucial conceptual transition: the realization that living tissue is an agential material.

A critical parameter for the continuum of agency is the degree to which an external observer must take into account the system's internal representation of the option space and its autonomous decision‐making. To know what a bowling ball is going to do on a bumpy landscape, your external, 3rd‐person view of the landscape, and calculations about energy minimization, tell the whole story. The same approach does not work for a mouse on that landscape—what is salient there is its view (internal model) of the landscape and free energy minimization with respect to its priors and goals, not yours [25, 26]. Where do cells fit on this continuum?

Single cells, from microbes to somatic cells, have been shown to have numerous proto‐cognitive capacities, including memory (learning), decision‐making, and anticipation [11, 27, 28, 29]. Especially relevant to future biomedicine are the examples showing that cells of multicellular organisms have not lost the basal features of adaptive information processing, including phenomena such as cardiac memory [30, 31] and active cell perception and signal processing with respect to stimuli [32, 33, 34, 35, 36, 37, 38, 39]. A diverse range of cell types exhibit learning via molecular signaling networks and bioelectric circuits [23, 24, 40, 41, 42] and context‐dependent decision‐making [36, 43, 44].

As will be seen below, the competencies of cell groups during morphogenesis arise from the scaling up of the single‐cell agentic repertoire. What underlies the competencies of cells? Does agential behavior first appear in single cells—are they the smallest unit of cognition [27, 45, 46, 47, 48]? It turns out that even below the single‐cell level, molecular components already have aspects that are tractable to the use of tools for the study of cognitive systems, and that what cells know (and what they can know)—their senome [45, 48, 49, 50]—is as important as their genomes, proteomes, and other such hardware specifications.

By applying standard approaches from behavioral theory [17], it was seen that even simple models of gene regulatory networks (GRNs) and molecular pathways show several different kinds of learning, including habituation, sensitization, and Pavlovian (associative) conditioning [23, 24]. By treating some nodes within the network/pathway as the unconditioned stimulus (UCS), other nodes as the response (R), repeated presentation of the UCS with an initially neutral node results in the network treating signals arriving on that node as a conditioned stimulus (CS) which can now trigger the response on its own (Figure 1B). This is a kind of dynamical state memory that does not require any hardware changes to the topology of the network or the strength of the edges (promoters). This is also a kind of “molecular placebo” because, due to its history of experiences, even a simple network can start responding to a neutral stimulus as if it was a much more potent one.

This shows how applying tools from other disciplines can reveal novel dynamics. The metaphor of gene networks as mechanical, low‐agency, dynamical systems [51, 52] is useful for many things but it did not on its own facilitate the discovery that GRNs and pathways should be trainable. Similarly, tools from cognitive neuroscience such as active inference are shedding light on reasons for the not‐infrequent failure of drug therapies [53, 54]. Importantly, for biomedical purposes, GRNs are an abstraction layer—an interface to cells that permits training them, just as brains and neural networks are an interface to animal behavior, one that was used by humans to train dogs and horses long before we knew how their brains worked. The implications of this molecular intelligence layer are numerous, spanning drug habituation, resistance of cells to therapeutics, unexpected side‐effects and differential efficacy across patients with different physiological histories, and the potential to use associative conditioning between drugs to induce outcomes using low‐cost and well‐tolerated trigger compounds [54]. The innate learning capacity of cells can lead to drug failure because the body's systems adjust to what they detect as a hacking attempt by an external exploiter. For this reason, first order models using conventional reagents to clamp a specific pathway state in place would not work—second order models, which take into account the cells’ ability to maintain goal states, and to facilitate re‐writing those goal states, are needed.

Given this learning capacity in chemical networks within cells, all the benefits of learning for evolution become relevant, modifying the speed and course of evolution beyond what standard cycles of random mutation + selection produce alone [55, 56, 57, 58]. Indeed, several computational studies have shown that evolution works quite differently over an agential material that has competencies and plasticity beyond a fixed genotype→phenotype relation [59, 60]. The more intelligent the software mapping, the more robust and rapid the evolutionary progress. Several studies have now identified ways in which exploratory learning and other problem‐solving capacities [40, 61, 62, 63, 64, 65] can affect evolution.

It is important to emphasize that we have only begun to scratch the surface of physiological problem‐solving. For example, when flatworms are exposed to barium—a nonspecific potassium channel blocker—their heads rapidly degenerate; this is not surprising given the necessity of endogenous potassium metabolism in the neurons and other cells of the head. Remarkably, they soon regenerate new heads which are completely adapted to the presence of barium [66]. Transcriptomic analysis reveals just a handful of genes that have been up‐ and down‐regulated to enable this feat of morphological and physiological homeostasis. The key question is: how do cells know which genes to regulate to resolve their physiological stressor? It is unlikely that they have a built‐in response to a frequent evolutionary history of barium exposure. As with the examples of polyploid newts and many other novel manipulations (Table 1, and the biobots described below), navigating a high‐dimensional transcriptional space to the right solution for barium exposure is an impressive problem‐solving capacity that we do not yet understand.

Thus, cells have the capacity to solve problems in metabolic, physiological, and transcriptional spaces. But that is just the beginning of the impact that collective intelligence has on the basic and applied life sciences. Some of life's most remarkable and impactful capacities are revealed when we turn to the goals and competencies of large groups of cells, which traverse the latent space of anatomical possibilities.

We previously argued that morphogenesis can be viewed as behavior in anatomical morphospace [13]; on this model, cellular collectives are systems that actively navigate from various starting states to a region corresponding to the species‐specific target morphology (Figure 2). As with all autonomous systems, the navigation process can exhibit diverse degrees of competency, ranging from a constrained random walk to highly advanced pathfinding and problem‐solving [89]. The practical benefit of entertaining this metaphor (in parallel to other equally metaphorical terms, such as “pathways” and “genes for traits”) is that it encourages a specific research program: the use of tools from behavioral and cognitive sciences that offer mature and powerful frameworks for understanding, predicting, controlling, and creating agents that navigate problem spaces.

From this viewpoint, neuroscience is not about neurons, but about understanding multiscale dynamics of emergent intelligence [95, 96]. If morphogenesis is decision‐making and goal‐directed behavior, then memories, preferences, measurements, disorders of perception, active inference, self‐models, and many other dynamics become potential therapeutic targets in development, regeneration, and cancer [15, 97].