Active inference, morphogenesis, and computational psychiatry
Low sensory precision is associated with hyporeactivity (Idei et al., 2021). In this case, the cells would not react enough to the chemicals they were sensing, leading to incomplete differentiation and migration of several cells of the collective; this in turn leads to failure of anatomical maturation.
This simulation shows the effect of the reduction of concentration signaling and sensitivity to the other cells signals of the two cells having a too high precision (see Figure 8). Without this intervention, the collective is unable to reach the target morphology and one head (red) cell stays close to intestinal cells and never fully differentiates. After the simulated biomedical intervention, as we can observe in Figure 8, the developmental defect is cured and all cells reach their location with the appropriate differentiation. The developmental noise we can observe in the pathological case is also removed.
This resembles the effects of antipsychotics for the treatment of schizophrenia. Most of them are antagonists of D2 receptors and therefore reduce the sensitivity of neurons to dopamine (Seeman and Kapur, 2000). In our simulation, we reduced the sensitivity to the other cell signals for the two cells having a too high precision, suggesting that dopaminergic drugs acting as antagonists could be tested in morphogenesis. Indeed, in active inference, the precision of sensory information is assumed to be mediated by neuromodulators, such as dopamine (Friston et al., 2012) or acetylcholine (Vossel et al., 2014).
Our framework emphasizes strong parallels between the dynamics of cognition and those of collective cell behavior during embryogenesis and regeneration. Due to evolutionary conservation (Fields et al., 2020), these parallels are not only functional but are mechanistic: the same machinery used to drive information processing in the brain should be implicated in the control of anatomical development. While this first generation of such models do not have the detail to be able to predict precisely that anatomical outcome of specific perturbations, they do make a surprising prediction not implied by other existing models of developmental events: that perturbation of dopamine signaling will cause errors in morphogenesis, including of non-neural tissues. Thus, we next tested a prediction of our models—that finely-tuned levels of sensory precision in cells are necessary for correct embryogenesis—using a loss-of-function pharmacological assay in early frog embryos. Thioridazine is a dopamine receptor antagonist and has been used to treat schizophrenia (Fenton et al., 2007). In the active inference framework, as a dopamine receptor antagonist, this drug may reduce sensory precision in biological systems (Friston et al., 2012) and according to our simulations this drug should induce different developmental defects. Thus, we applied this drug to Xenopus laevis embryos during early stages (see section 6) and scored their morphogenesis by the swimming larva stage. Compared to normal embryos (Figures 9A-left, Figures 9B-top), treated embryos exhibited abnormalities that included hypopigmentation, kinked primary axes, edemas, abnormal face shapes, malformed guts, and cleft cement glands (Figures 9A-right, Figures 9B-bottom). The percentage of embryos with each defect is shown in Figure 9C; raw counts are shown in Table 1.
These results are consistent with our models, which suggest that modulation of mechanisms, such as neurotransmitter signaling (known to be involved in information processing in the field of behavioral and brain sciences), can regulate morphogenesis of non-neural structures. More broadly, these results contribute to the study of morphogenesis as behavior of a collective cellular agent in anatomical morphospace (Fields and Levin, 2022).
Xenopus laevis tadpoles were raised in 0.1X Marc's Modified Ringers solution (MMR), pH 7.8, and were staged according to Nieuwkoop et al. (2020). All embryos (pooled from separate female frogs and then randomly divided) were raised at 14°C. Feeding stage animals were fed three times a week with Sierra Micron powdered diet, and media changes were performed on alternate days. Tadpoles were raised at a density of 100 tadpoles per 40 ml of media in 100 × 20 mm petri dishes. All experiments were approved by the Tufts University Animal Research Committee (M2020-35) in accordance with the guide for care and use of laboratory animals.
Thioridazine (thioridazine-HCl, Sigma-Aldrich) was dissolved in deionized water at 1mM, and were then frozen in single use aliquots to prevent continuous re-thawing. Drug exposure occurred at 90 μM in 0.1X MMR conducted on NF stage 12.5–26 embryos that had been reared at 14°C prior to exposure. During the treatment, experimental and controls embryos were kept at 18°C. After the exposure they were returned to 14°C until they reached scoring stage 45. Scoring was performed by inspection of morphology using transmitted light microscopy.
In this article, we simulated disorders of morphogenesis as disorders of (collective) inference. The decision of one cell is a function of the states of the other cells; or in other words, the collective sends the appropriate signals to each individual cell. In a functional sense, the collective “remembers” the body shape; if one cell diverges, the collective continues emitting signals that guide that cell until it again reaches the correct position. We implemented deficits of sensory attenuation and aberrant precision that are usually invoked as explanations of autism and schizophrenia and also considered a low precision simulation that has been linked to hyporeactivity. All of these simulations lead to developmental defects via aberrant sensory precision relative to the prior precision. We presented a proof of concept appealing to disorders of inference and showed that precision control is useful in understanding not only psychopathological conditions, but also developmental defects and morphogenesis (see Table 2).
We have furthermore shown how manipulations of the information processing modes, in terms of precision and prior strength, interplay with biochemical signals in developmental biology, such as signaling ligand concentration and receptor sensitivities. Crucially, our (simulated) results showed that the disrupting effect of individual cells' signaling aberrations on the overall morphology can be rescued by decreasing the diffusion of signaling ligands and receptor sensitivities, an effect that would have been hard to predict without our proposed view of certain morphological defects as disorders of inference and precision control (Pezzulo and Levin, 2015). This simulated result remains to be experimentally tested in vivo and is left as future work. This new framework that focuses on the precision of spatiotemporal signal progression during development comes at an opportune moment, when tools such as optogenetics (Bugaj et al., 2017) and microfluidics (Sonnen and Merten, 2019) have provided us for the first time with enough spatiotemporal precision to study and manipulate these events in real time on a single cell level.
If developmental defects are the results of disorders of active inference similar to what we can find in psychopathologies like schizophrenia or autism, it may be possible that the psychoparmacology used for hyper- or hyporeactivity, hallucinations, or other symptoms related to these diseases could be used to address developmental defects. Various case studies highlight this possibility. Indeed, by applying different reagents targeting the glutamatergic, adrenergic, and dopaminergic pathways to Xenopus laevis embryos from gastrulation to organogenesis stages, Sullivan and Levin observed numerous developmental defects, including craniofacial defects, hyperpigmentation, muscle mispatterning, and miscoiling of the gut (Sullivan and Levin, 2016). Serotonin pathways are also implicated in left-right asymmetry and the pre-neural morphogenesis (Fukumoto et al., 2005; Levin et al., 2006). Serotonin signaling has also been linked to autism and may be involved in this pathology during early brain development (Yang et al., 2014). In addition, usually, dopamine, serotonin, and adrenaline receptors, are molecular targets for antipsychotics and schizophrenia (Stepnicki et al., 2018); and we have now provided evidence that modulators of decision-making in traditional behavioral systems have predicted effects on morphogenetic behavior of cell collectives. Therefore, by targeting similar pathways during development, we may be able to reverse developmental defects.
These ideas are part of a bigger effort to solve the inverse problem in regenerative medicine (Lobo et al., 2014): strategies such as CRISPR and genome editing cannot reach their full potential because it is usually too difficult to know what to change at the hardware level (DNA and pathways) to achieve desired changes at the level of anatomy. Fortunately, neuroscience has blazed the trail in a pluralistic understanding of complex systems: work in that field (and types of intervention targets) ranges across studies of ion channel protein structure, synaptic plasticity, network roles in image processing and memory, behavioral responses, and psychological drives and analysis. Neuroscience would greatly benefit from understanding how non-neural cells work, because neurons share many molecular and functional properties with non-neural cells. Thus, while current biomedical approaches have focused entirely on manipulating the lowest level (gene regulatory networks and protein structure/pathways), we propose to begin to deploy interventions across scales in regenerative medicine, as commonly occurs in the behavioral sciences. We have previously proposed that high-level manipulation of tissues' perceptions and memories (via training for example) may be a better way to manipulate large-scale properties of morphogenesis in health and disease than bottom-up micromanagement of the protein-level hardware (Pezzulo and Levin, 2015; Mathews and Levin, 2018).
The active inference framework is a very powerful aspect of system-level functionality in cognition (Friston K. et al., 2017; Pezzulo et al., 2018; Parr et al., 2022), and a key target of strategies to exploit invariants between neuroscience and other fields (Friston, 2013; Fields and Levin, 2020; Fields et al., 2021). Thus, once we reach a better understanding of the role of precision in development and morphogenesis, it may be possible to apply “reverse inference” as we can see for computational psychiatry (Schwartenbeck and Friston, 2016). Reverse inference in development/morphogenesis reflects the probability of a developmental process being present given the knowledge of activation in a particular tissue region. The aim is to infer from a specific anatomy which precision parameters caused it (i.e., which precision parameters were incorrectly set). Similarly to Schwartenbeck and Friston (2016) where the researchers infer the model parameters of a single subject or a group of subjects from measured behavioral or neuronal responses, it could be possible to develop a “computerized clinical diagnosis” for developmental defects leading to specific biomedical interventions. These are just the first steps on a roadmap to regenerative medicine that borrows from neuroscience the highly fruitful approaches focused on modulation of systems at different levels, from molecular pathways to behavioral decision-making. A future computational somatic psychiatry will take advantage of the tools of behavioral and cognitive sciences to manipulate not only the hardware of cellular pathways, but also the decision-making of cellular collectives.
By taking seriously the native competencies of tissues (Levin, 2021b, 2022c), it may eventually be possible to effectively modulate outcomes in birth defects, traumatic injury, and cancer, not by micromanagement but by quantitative behavior shaping in morphospace.
The original contributions presented in the study are included in the article/supplementary materials, further inquiries can be directed to the corresponding author.
LP-L and FK wrote code, performed computational experiments, and analyzed data. AT performed Xenopus experiments and analyzed data. GP and ML developed the project, guided the experimental design and conceptual foundations, and provided oversight and funding. All authors contributed intellectually to the work, co-wrote the manuscript, and approve the submitted version.
This research received funding from the European Union's Horizon 2020 Framework Programme for Research and Innovation under the Specific Grant Agreement No. 945539 (Human Brain Project SGA3) to GP and the European Research Council under the Grant Agreement No. 820213 (ThinkAhead) to GP. ML gratefully acknowledges support by the Templeton World Charity Foundation (grant TWCF0606) and the John Templeton Foundation (grant 62212).
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.