Active inference, morphogenesis, and computational psychiatry

Active inference is a leading theory in neuroscience that provides a simple and neuro-biologically plausible account of how action and perception are coupled in producing (Bayes) optimal behavior; and has been recently used to explain a variety of psychopathological conditions. In parallel, morphogenesis has been described as the behavior of a (non-neural) cellular collective intelligence solving problems in anatomical morphospace. In this article, we establish a link between the domains of cell biology and neuroscience, by analyzing disorders of morphogenesis as disorders of (active) inference. The aim of this article is three-fold. We want to: (i) reveal a connection between disorders of morphogenesis and disorders of active inference as apparent in psychopathological conditions; (ii) show how disorders of morphogenesis can be simulated using active inference; (iii) suggest that active inference can shed light on developmental defects or aberrant morphogenetic processes, seen as disorders of information processing, and perhaps suggesting novel intervention and repair strategies. We present four simulations illustrating application of these ideas to cellular behavior during morphogenesis. Three of the simulations show that the same forms of aberrant active inference (e.g., deficits of sensory attenuation and low sensory precision) that have been used to explain psychopathological conditions (e.g., schizophrenia and autism) also produce familiar disorders of development and morphogenesis when implemented at the level of the collective behavior of a group of cells. The fourth simulation involves two cells with too high precision, in which we show that the reduction of concentration signaling and sensitivity to the signals of other cells treats the development defect.

Finally, we present the results of an experimental test of one of the model's predictions in early Xenopus laevis embryos: thioridazine (a dopamine antagonist that may reduce sensory precision in biological systems) induced developmental (anatomical) defects as predicted. The use of conceptual and empirical tools from neuroscience to understand the morphogenetic behavior of pre-neural agents offers the possibility of new approaches in regenerative medicine and evolutionary developmental biology.

Embryonic self-assembly requires large numbers of cells to cooperate and compete toward specific morphologies (Pezzulo and Levin, 2016). Moreover, the repair and remodeling of complex anatomical structures under novel circumstances (injury, changes in cell size or number, etc.) illustrates that this process is not hardwired. Numerous examples, such as tadpoles with scrambled craniofacial organs that move in un-natural paths to nevertheless construct normal frog faces (Vandenberg et al., 2012), and similar data even in early chordates (Voskoboynik et al., 2007), reveal control of growth and form to exhibit the ability to solve problems in morphospace—dealing with novel circumstances to arrive at the correct target morphology despite unexpected perturbations. It has been argued (Levin, 2019, 2022b) that morphogenesis is an example of a kind of basal intelligence, meeting William James' definition of “same ends through different means” (James, 1890), due to its ability to restore amputated structures (and then stop), produce the same size body using very large, or very few cells, or remodel after organs are placed in incorrect configurations (reviewed in Pezzulo and Levin, 2016; Levin, 2022a). Specifically, it has been proposed that the sorts of problem-solving capacities observed when nervous systems navigate the 3D world via behavior evolved by speed-optimizing much more ancient versions of the same system in pre-neural cells (Buznikov and Shmukler, 1981; Fields et al., 2020).

In an important sense, morphogenetic tissues and brains both represent a kind of collective intelligence, emergent from groups of cells (Fields et al., 2020; Fields and Levin, 2022). Much as neural cells connect in networks to process information and control muscles, collections of cells navigate anatomical morphospace by processing information toward the control of non-excitable cell types (resulting in cell migration, differentiation, etc.). Morphogenesis can thus be seen as the behavior of a collective in morphospace. Indeed, all of the machinery used in the central nervous system (CNS) to support adaptive functions (ion channels, electrical synapses, neurotransmitter signaling, etc.) is ancient, and is implicated in the control of growth and form (Levin and Martyniuk, 2018). Consistent with this are the many examples over the last decade of tools from neuroscience being used to probe and control animal shape in vivo: tools like optogenetics, ion channel pharmacology, etc. cannot distinguish neural from non-neural applications and work well in both arenas (Bates, 2015; Harris, 2021; Levin, 2021a). Given the high conservation between mechanisms regulating canonical behavior and those regulating morphogenesis (Pezzulo and Levin, 2015), it becomes interesting to ask: which conceptual models from neuroscience shed light on the collective dynamics of morphogenesis? This is especially important, both for regenerative medicine and for fundamental evolutionary developmental biology: progress on molecular genetics mechanisms must now be augmented by a better understanding of the algorithms and organ-level decision-making, to develop prediction and control strategies for large-scale structure and function. Here, we focus on one specific set of ideas from neuroscience—the active inference framework—Parr et al.

(2022) exploring interesting parallels that allow it to generate insights for open problems in the science of dynamic morphogenesis.

What paradigm can be used to understand the dynamic information processing that underlies both, adaptive behavior and morphogenesis? Active inference is a general framework in theoretical neuroscience that proposes a unified account of perception, learning, and action (Friston et al., 2006; Friston, 2010). The framework is built around the idea that biological agents must maintain homeostasis and have to reach a limited range of states defined by their phenotype and avoid “surprising” states (Friston et al., 2010). However, as the states of the environment and the internal states of the organism itself are hidden (i.e., not directly observable), a biological agent must infer them from incoming sensory information. The framework claims that an agent is a probabilistic machine that embodies an internal generative model of its environment and continuously tries to minimize the surprise of its own sensations. In other words, an agent minimizes the difference between its expectations of the world (that can be priors on its own sensorium for perception, or goals for actions) and its sensory evidence. In turn, this requires that organisms continuously try to predict the internal (proprioceptive or interoceptive) and external (exteroceptive) consequences of their behavior and others—hence the idea of the brain as a “prediction machine.”

More formally, in active inference, action, perception, and learning occur through a process of variational free energy minimization (where variational free energy is an upper bound on surprise) (Friston K. et al., 2017). In this context, perception resolves exteroceptive prediction errors by selecting the predictions that best explain the current sensations; action serves to change the world in such a way that it generates the agent's preferred sensations (as encoded in its priors, or prior preferences); and learning is the update of the generative model.

The framework of active inference has been initially developed in computational neuroscience (Friston K. J. et al., 2017; Walsh et al., 2020) but was recently extended beyond it, to address problems of morphogenesis and developmental biology (Friston et al., 2015; Kuchling et al., 2020). In this perspective, cells are minimal active inference agents that minimize their surprise (or free-energy) in order to reach (collectively) a target morphology and maintain anatomical homeostasis (see Figure 1). This novel perspective on the coordination of migration and differentiation of cells suggests an interpretation of cellular behavior as being driven by an error minimization mechanism akin to usual conceptualizations of goal-directed organismal behaviors in neuroscience, cybernetics, and philosophy (Millikan, 1987; Friston et al., 2015; Fields and Levin, 2019; Kuchling et al., 2020). This top-down approach to developmental biology may complement bottom-up strategies that currently focus on molecular pathways and could help understand regeneration, morphogenesis, and their associated disorders.

In which sense can this framework help understand disorders of morphogenesis? Within the active inference framework, psychopathological conditions such as schizophrenia, hallucinations, and depression can be understood as disorders of inference (Montague et al., 2012; Friston et al., 2014). Along similar lines, active inference can help formalize developmental defects or aberrant morphogenetic processes as disorders of inference—when the inference is about (for example) a target morphology, as in Friston et al. (2015). Similarly, in developmental biology, focus is beginning to shift toward an understanding of measurement, memory, and prediction in cells and tissues (Barkai and Shilo, 2009; Abouchar et al., 2014; Cervera et al., 2019; Levin, 2021a).

The aim of this article is three-fold. Our first aim is to establish a formal link between disorders of inference apparent in psychopathological conditions and disorders of morphogenesis/development, via active inference. We exemplify our ideas by focusing mainly on schizophrenia and autism, because these conditions have been studied deeply under active inference and they depend on a key mechanism (aberrant precision control) that we believe could help explain developmental disorders, too (see below). However, we will not attempt to establish exact, one-to-one correspondences between these two (or other) psychopathological conditions and developmental disorders. Rather, our goals are to establish more general formal correspondences between psychopathological and morphogenetic/developmental problems—for example, using the concept of precision control that might be heuristic in both fields—and simultaneously to offer some specific examples, which could guide more specific future investigations. Our second aim is to show how disorders of morphogenesis can be simulated using active inference. Our third aim is to discuss how the formalization offered here could help design novel intervention strategies for repairing developmental defects or aberrant morphogenetic processes seen as disorders of information processing. We recognize that psychopathological conditions are very complex and multifarious (e.g., fundamentally shaped and constrained by complex cultural and sociological factors). The examples we are offering in this paper do not imply a reductionist view of these phenomena but are limited to modeling some specific computational aspects of them.

Indeed, in this aspect our approach to cognition parallels the arguments made in our prior work about morphogenesis, where we suggest that a more effective understanding of collective cell behavior in embryogenesis and regeneration requires not only reductive molecular biology-based mechanistic information, but also a consideration of information processing and goal-directed action (Pezzulo and Levin, 2015; Levin, 2022a,c).

We present four simulations of disorders of morphogenesis as disorders of active inference. These simulations use active inference to model the behavior of groups of cells and show that endowing the cells with aberrant values of a key parameter (precision) leads to developmental defects in much the same way that manipulating the precision of more sophisticated active inference models could reproduce some aspects of psychopathological conditions, such as the positive symptoms of schizophrenia (Brown et al., 2013) or aberrant statistical learning in autism (Arthur et al., 2021). We also present a simulated biomedical intervention of the reduction of concentration signaling and sensitivity to the other cells signals of two cells having a too high precision. Finally, we present the experimental result of the test of one of the predictions of our model: thioridazine (a dopamine antagonist that may reduce sensory precision in biological systems) is shown to induced developmental (anatomical) defects in Xenopus laevis tadpoles, illustrating the ability of these conceptual models to drive novel empirical results in developmental biology.

Active inference is a computational framework for action and perception and has been linked to various psychiatric disorders (Montague et al., 2012). From this point of view, psychopathological conditions can be understood as disorders of (active) inference. False inference could explain symptoms, like hallucinations or delusions in schizophrenia (Fletcher and Frith, 2009), to the loss of coherence in autism (Pellicano and Burr, 2012).

While there have been several alternative proposals about what exactly causes psychopathological conditions (Friston et al., 2014; Paulus et al., 2019; Van den Bergh et al., 2021), many focus on deficits of a specific parameter of the (active) inference: the precision parameter that weights prior beliefs and sensory evidence (and hence prediction errors, which measure the discrepancy between predicted and sensed observations). The precision parameter encodes the expected uncertainty or in other words the degree of confidence of an agent in the sensory information it is receiving in any given context (Angela and Dayan, 2005; Brown et al., 2013; Iglesias et al., 2013). A correct regulation of the precision parameter is key for the success of hierarchical (active) inference, because the balance between prior beliefs and sensory evidence during the process of evidence accumulation is determined by the precision of prediction errors at each level of the hierarchy. If the precision is set incorrectly (e.g., if an unreliable source of evidence is assigned excessively high precision, or a reliable source of evidence is assigned excessively low precision), the resulting inference can be misleading.

More formally, precision is defined in active inference as the inverse of the variance of a given variable; and it can be interpreted as a measure of signal-to-noise, or the confidence associated to an incoming information stream. For the brain, it can be represented by the modulation of the gain or excitability of neurons that compute prediction errors (Clark, 2013; Moran et al., 2013). This is particularly important, because many psychological disorders implicate modulatory neurotransmitter systems and a putative failure of post-synaptic gain control (Feldman and Friston, 2010).

One of the reasons why the precision parameter is so important in explaining psychopathological conditions is that it mediates the so-called sensory attenuation that necessarily accompanies action. In active inference, actions are realized by descending proprioceptive predictions, which in turn engage motor reflexes that enact the intended or predicted movement (Adams et al., 2013). In other words, in order to realize an action, an active inference agent has to imagine the action outcome and then enact a movement that fulfills the prevision. In parallel, it has to attenuate the sensory precision—in order to disregard the evidence that one is not moving. The attenuation of sensory precision is necessary for action, because otherwise, ascending (proprioceptive) prediction errors would lead to the revision of predictions about actions (i.e., correctly infer that one is not moving) instead of realizing the intended actions. Phenomenologically, the attenuation of sensory precision produces sensory attenuation—or the transient suspension of attention to the sensory consequences of an action—that is commonly observed during movement (Brown et al., 2013).

Sensory attenuation is an important theme in many psychiatric disorders, too. It has been proposed that prior expectations on sensory precision are compromised in schizophrenic patients—and hence the patients are unable to attenuate sensory precision. More specifically, the sensory precision of schizophrenic patients might be too high in relation to the precision of their (prior) beliefs about the causes of sensations. Therefore, schizophrenic patients cannot ignore sensory evidence. Deficits of sensory attenuation in schizophrenic patients may also produce disorders of movement, given that movement itself depends on correct precision control (see Brown et al., 2013 for a detailed discussion). The failure in sensory attenuation might explain two types of false inference in schizophrenia: false positives and false negatives (Limongi et al., 2018). Positive symptoms, like hallucinations or delusions (Powers et al., 2017), correspond to prior beliefs that are accorded too much precision. Negative symptoms, like the resistance to illusions and the failure of slow pursuit movements, correspond to a compromised capacity to elicit predictions informed by prior beliefs.

One symptom that shows this deficit is slow pursuit eye movements. Patients with schizophrenia present a disorder in the anticipation of movement. This failure can be measured experimentally using a mask during the motion of an object, with the subject having to anticipate the re-appearance of the object behind the mask. A computational model of this process that used active inference and manipulated precision could account for several features of smooth pursuit in schizophrenia: a reduction in anticipatory eye movements during visual occlusion, a paradoxical improvement in tracking unpredicted deviations from target trajectories, and a failure to recognize and exploit regularities in the periodic motion of visual targets (Adams et al., 2012).

Disorders of beliefs about agency are also common in schizophrenia (in the sense of agency as the property of feeling that one is the agent in control of one's actions). One example is the resistance of patients with schizophrenia to the force-matching illusion (Shergill et al., 2005, 2014). During this task, a patient's hand is touched by a device and therefore the patient can sense an external force. Patients have to press directly on themselves, or use a robot to reproduce the magnitude of the perceived pressure of the external force. The force-matching illusion consists in the fact that healthy people underestimate the magnitude of the force that they apply (i.e., self-pushing power), which results in pushing the device with a larger force than the external pressure. Patients with schizophrenia, instead, present a better accuracy on this task, which suggests that they might not attenuate their sensations. This result has been modeled using active inference and corresponds to an attenuation of the sensory precision that would create this illusion (Brown et al., 2013). Similarly, depersonalization disorder has been linked to a deficit of somatosensory attenuation and precision-weighting (Ciaunica et al., 2022).

Autism has also been linked to an aberrant account of precision (Lawson et al., 2014). Autism is a neurodevelopmental disorder of variable severity that is characterized by difficulties in social interaction and communication and by restricted or repetitive patterns of thought and behavior. Aberrant precision may explain different features of autistic perception, action, and social behavior. One hypothesis is that autism is characterized by excessively high sensory precision (compared to the precision of priors) leading to the difficulty in contextualizing sensory information (Frith and Frith, 1999; Baron-Cohen, 2000). Furthermore, in autism, the failure to attenuate sensory precision has been interpreted in a developmental context—and particularly in the context of the acquisition of generative models that distinguish between self and other. It is proposed that an infant with autism may have difficulties distinguishing between autonomic responses elicited by the mother and those caused by its own interoceptive predictions. This could also explain autonomic over-responsiveness to interoceptive cues or interoceptive hypersensitivity and a deficit in engaging with social (exteroceptive) cues (Paton et al., 2012).

In sum, here we have highlighted that psychopathological conditions, such as schizophrenia and autism, can be described in terms of aberrant values of key parameters of active inference, and most notably the precision parameter that weights sensory evidence and prior information [see also chapter 9 of Parr et al. (2022) for the discussion of other psychopathological conditions in terms of maladaptive active inference]. Below we highlight that these information- and system-level disorders have parallels outside of CNS-mediated behavior—in the domain of morphogenesis.

A key fact of development and regeneration is that individual cells must cooperate in networks to achieve specific outcomes in morphospace, to very tight tolerances, despite noise, environmental influences, and imperfections in the components themselves. Doing this in real-time, to achieve the correct target morphology via novel paths, requires collective decision-making and active measurements of current shapes as well as the ability to recognize errors (perhaps through a memory of correct states). Morphogenesis has been simulated in the active inference framework (Friston et al., 2015; Kuchling et al., 2020). From this perspective, cells are information processing agents, where the driving force behind morphogenesis is the minimization of a cell's variational free energy. Each cell is equipped with a generative model that encodes its beliefs about what are the chemotactic signals it should receive or express, relative to its location in the target morphology. Cells reach their morphogenetic goal by performing actions (expression of receptors and other biochemical signals) according to their internal (i.e., generative) model of what type of receptors and other signals they should express.

Within this framework, Friston et al. simulated morphogenesis and dysmorphogenesis (e.g., induced birth defects) by manipulating the influence of extracellular signals, without changing the generative model (genetic and epigenetic processes) (Friston et al., 2015).

Using a similar strategy, Kuchling et al. simulated alterations of anterior–posterior axial polarity (i.e., the induction of two heads or two tails) as in planarian regeneration (Kuchling et al., 2020). Furthermore, they simulated the first steps of carcinogenesis by imposing false inference of a single cell on what it should sense and how it should act within the cellular ensemble. They also showed that mis-patterning of development and regeneration can be caused (or can be rescued) by simple modifications of the inference process without changing the implicit generative model of a cell as determined (for example) by its DNA.

However, the previous work on this topic did not investigate the impact of precision and sensory attenuation. In other words, the generative model had remained untouched in these prior simulations. Kuchling et al. manipulated external parameters to show disorders of morphogenesis (Kuchling et al., 2020), and Friston et al. (2015) manipulated the extracellular signals, both without changing the generative model or the precision of its parameters.

Here, we show that disorders of inference resulting from precision changes can produce disorders of morphogenesis in cell-like agents, analogous to the way aberrant precision control might produce psychopathological conditions in sophisticated (human-like) agents.

In this section, we summarize the mathematical foundation that underlies the Bayesian interpretation of non-equilibrium steady-state dynamics [based on Kuchling et al. (2020)] and introduce the formulations necessary to set up and discuss our morphogenetic inference model based on the minimization of variational free energy.

To that end, we begin by stating the dynamics of our system in generalized coordinates of motion, denoted with a tilde, where x~ is defined as:

This expression defines a state not just through its position, but also with its velocity, acceleration, etc. Generalized coordinates of motion allow the inclusion of temporal correlations in random fluctuations, assuming a sufficiently smooth dynamical system, through the Langevin equation:

where f(x~) is the generalized flow (or time evolution) of states under forces acting on the states and random fluctuations ω~ that are confined by the typical Wiener assumptions; in other words, the flow of states is made up of a process of independent, Gaussian increments that follow a continuous path.