Looking for keys in model organisms: interpreting and keying-up in basal cognition
Michael Levin and his lab, one of the hotbeds of basal cognition, study a variety of model organisms. Their work is rooted in developmental biology, and as such they use systems derived from that field, primarily the frog Xenopus laevis and invertebrate Planaria (see the Levin lab website7). Their perspective is based on the view of “morphogenetic systems as cognitive agents” (see the lab website8). This notion is embedded in the framework of “cognition all the way down” (Chis-Ciure & Levin, 2025; Levin & Dennett, 2020), which views biological systems as made up of agentic and cognitive parts—a notion that Levin has broadly explored (e.g., Levin, 2019, 2022, 2023). Here I wish to focus on a particular set of experiments on Planaria (in particular, Oviedo et al., 2010; which uses the species Dugesia japonica, see also the review in Levin, 2022).
Planarian flatworms are an important model organism for the study of development (with history stretching back to the research of Thomas Hunt Morgan). They exhibit astonishing regenerative capacities, as they can regenerate a whole body even from a thin slice (hence, they respond reliably to experimental manipulation, i.e., meet the responsiveness criterion in Dietrich et al., 2020). Genomes of several planarian species have been sequenced, and a variety of tools is available for working with them. Levin lab experiments specifically target the understanding of morphogenesis, the process and mechanisms underlying the acquisition of shape. In experiments reported by Oviedo et al. (2010), the worms undergo an amputation where the tail, along with a portion of the body, is removed. They are then transiently submerged in a mixture of water with a gap junction blocker (octanol), and left to regenerate in normal water. During the regeneration, some animals are further submitted to a surgical procedure to disrupt ventral nerve cords. This treatment impacts the regular regeneration, and the worm achieves a significantly altered body plan, in many cases regenerating a second (or even a third) head, either protruding from the side, or replacing the tail. These results indicate that morphogenesis does not rely directly on a genetically encoded body plan, but rather requires bioelectrical communication between the cells (altered by the gap junction blocker) and the inputs from the nervous system (in the presence of head post-amputation, signals through the ventral nerve cords can suppress the growth of an additional head, but the surgical procedure in the experiment further disrupts the anterior/posterior axis in regeneration).
Further experiments (e.g., Durant et al., 2017) confirmed these results, and explored in greater detail the role of bioelectric signalling between the cells and its role for morphogenesis. Crucially, the studies conducted by Levin and colleagues indicate that the modification of the body plan persists across multiple rounds of regeneration, without any chemical intervention beyond the first round. They interpret these results as indicating that the body plan is at least partially stored by the pattern of bioelectric signalling which persists over a longer period.
These studies are used in two research contexts. They are directly embedded in the set of questions from developmental biology. In this context, the aim is to capture an important biological mechanism for the control of shape during development (morphogenesis). More importantly for current purposes, Levin explicitly connects this work to central problems of basal cognition research. His research models the behaviour of individual cells through “cognitive lens”. In this context, similarly to the other two case studies discussed, individual cells in a developing (or regenerating) organism are interpreted as exemplifying information processing-properties. The existence of bioelectrical communication between cells has long been known, similarly to the fact that it’s involved in developmental processes. What this research aims to highlight is the specific functional role of such cells, as well as its sensitivity to the specialized electrical communication circuit, i.e., the nervous system. The intervention highlights the extent of control that bioelectrical signals have. Standard conceptualizations of developmental processes consider the target morphology of the organism to be encoded or stored by individual cells (in their genetic code—assuming an informational lens). This motivates an ascription of more complex information processing capacities to cells and the processes of bioelectrical signalling (potentially ascribing them some computational capacities, see Rorot, 2022).
Such an interpretation allows, as in the case of plants, viewing the developmental system as a model for basic principles of biological information processing (the relevant Z-domain), abstracting them from specializations likely to evolve along the different branches of the tree of life. But even more generally, this allows Levin and his collaborators to search for mechanisms underlying all flexible and adaptive behaviour, expanding how the relevant Z domain is conceptualized. In doing so they also construct a key-sketch which identifies relevant units of analysis through their agentic behaviours, independently of the scale of organization at which they exhibit those behaviours. Once this identification is made (partially through stipulation, partially motivated by the experimental intervention), the relevant agentic entities are mapped onto one another by composing the interpretation- and key-sketches. For instance, the activity of individual cells are mapped onto the activities of the organism as a whole. This allows for certain functional or organizational properties to be imputed between the different biological systems. For instance, decentralized processing and storage of information is one such property, identified in the experiments on developmental systems, and then imputed to other biological systems, for instance memory (e.g., Colaço, 2023; Colaço & Najenson, 2024). As in the case of plant signalling, while these are not strictly functional ascriptions, they follow a similar pattern to Sartori’s functional identity keys, in that the key maps coarser-grained properties that can be realized by different mechanisms across the model and target organisms.
In 2003, the Nobel Prize in Chemistry has been awarded to Peter Agre and Roderick MacKinnon for “discoveries concerning channels in cell membranes” (Nguyen, 2003). MacKinnon, in particular, was awarded for “structural and mechanistic studies of ion channels.”9 MacKinnon’s work has relied on analysis of bacterial potassium channels. His central achievement, which was the direct motivation for the Nobel prize, was uncovering the molecular basis of potassium conduction and the selectivity of ion channels in bacterium Streptomyces lividans (Doyle et al., 1998; see also MacKinnon, 2004). However, the function of ion channels for bacteria (beyond homeostasis) has been uncovered only in 2015, in the work from Gürol M. Süel and his lab (Prindle et al., 2015). Süel’s lab uses Bacillus subtilis, a widely accepted and extensively studied model organism (Barák, 2021). For the undomesticated strain NCIB 3610 used in the experiments discussed here (Prindle et al., 2015), the complete genome sequence is available (Nye et al., 2017). The experiments of Süel and colleagues consisted of imaging the potassium signalling in growing biofilms using voltage indicator dyes. The experimental intervention focused on the availability of nutrients. Further, they recorded voltage directly with patch clamps, and compared the results with genetically modified bacteria, where the expression of the ion channel proteins was suppressed. Finally, they used mathematical modelling to establish whether the observed patterns of electrical activity can be attributed to the proposed potassium channel mechanism. Their results point out that ion channels in the biofilms conduct long-range electrical signals. These signals coordinate the metabolic states among cells of the biofilm, using a mechanism of positive feedback loop.
Quite surprisingly, this positive feedback loop seems largely similar to the activity of ion channels in nerve cells (in fact, the relation between bacteria and brains is made explicitly on the Süel Lab website10).
The selection of bacteria for this study seems to depend on a number of factors (following the categorization in Dietrich et al., 2020), from available resources and institutional reasons (the well-established role of bacteria in the study of ion channels), through their experimental responsiveness, to the comparative and translational potential (established in previous work, and ultimately cemented with the Nobel Prize award to MacKinnon).
The authors begin key construction with a mapping already established in previous research. That key links the general mechanism of ion channels and the particular molecular make-up of potassium channels in bacteria to distinct ion channels expressed in various cells of other organisms—most importantly in neurons. This is a functional identity key, which maps the same functions of ion channels (e.g., voltage gating) across organisms, regardless of different mechanistic underpinnings (different proteins used).
Uniquely, the extension of the key seems to be constructed in the opposite direction, at least as it applies to higher-level function, the role of biofilm-wide signalling. The inference is drawn from target to model, as an inverse key of sorts. Ion channels in nervous systems have long been understood to be involved in electrical signalling. Hence, Prindle and colleagues draw on this view of channel function, assuming that the functional identity of the underpinning mechanisms allows for the functional identity at a higher level of organization, i.e., that the channels in biofilms also enable long-distance information exchange. In doing so, their inference follows the heuristic of the localization of function (Wimsatt, 2006; see also below in Sect. 4). If we assume structure-function correspondence, identifying analogous structures motivates the search for an analogous function–in this case, between the structure of ion channels in unicellular organisms and in neurons. Indeed, this extension of the key allows them to connect the results of their study to open issues in neuroscience:
For example, the connection between neuronal signalling and metabolic activity (neurometabolism) is an active area of research. Furthermore, depletion of glutamate, the most common excitatory neurotransmitter, also forms the initial trigger for these collective metabolic oscillations synchronized by potassium. Therefore, it is intriguing to think not only about the structural similarities between bacterial and human potassium ion channels, but also their possible functional similarities with respect to long-range electrical communication. (references removed; Prindle et al., 2015)
The presence of heuristics and idealizations in scientific modelling has long been underscored in philosophical accounts of model use (Nowak, 1980; Potochnik, 2017; Weisberg, 2007, 2013; Wimsatt, 2007; see also Cartwright, 1983). MOs are no different in this regard. In the case of idealizations, this is recognized in the literature (e.g., Levy & Currie, 2015; Sartori, 2023), though these authors use this term more broadly than is currently accepted, essentially conflating idealizations with standardizations.11 What is crucial, MO-based inferences do involve introducing distortions with positive representational function. This can be seen quite explicitly in the case of Drosophila melanogaster as a model of neurodegenerative disease (e.g., Feany & Bender, 2000). While Drosophila doesn’t exhibit neurodegenerative disease in the wild, the symptoms can be induced by genetic tools intervening on analogous and homologous pathways to those hypothesized to be involved in the development of disease such as Parkinson’s. This artificial insertion of symptoms is closely tied with important limitations of the model, such as the short lifespan of the fly which limits the development of the symptoms to early stages of neurodegeneration (see the overview in Vos & Klein, 2021). Together, these limitations distort how the model exhibits the target phenomenon. Properly constructed keys take these distortions into account.12
Heuristic reasoning, on the other hand, has not been explicitly recognized as playing a role in MO-based research. Quite the opposite, a large chunk of philosophical study of MOs—including the work within the DEKI framework—has focused on the justifications of the inferences from MOs to targets. While this is an important issue for established MOs, the adoption of species in a novel domain involves much looser inference schemas, as illustrated in the case studies.
Heuristics are cost-effective substitutions for truth-preserving algorithms, which do not guarantee correct outcomes. Compared to such exact inference schemas, heuristics offer significant improvements in complexity by transforming a problem into a “nonequivalent but intuitively related” one (Wimsatt, 2007, p. 77). Crucially, the errors that heuristics introduce are not random, but systematically biased, and as such they can be scrutinized and (ideally) corrected, with their limitations and conditions of failure being predictable.
For model organisms, heuristics enable a construction of the key while the exact relation between model and target is not established, and the extent of distortions in the model is not known. In the case studies analysed here, the reductionist heuristic of localization (Wimsatt, 2007, p. 82), embedded within the broader molecularization of life sciences (see Chadarevian et al., 2003), plays the primary role. This heuristic assumes that an identified function is implemented by a localized structure or mechanism. In the case of biofilms and developmental systems, electrical signals are localized to activities of ion channels, and in the case of plants—chemical cellular communication to the vesicular containment of auxin. Crucially, in all three cases (see summary in Table 1), it is difficult to justify the ascription of the function directly to the model organism. The structure, however, is known (except for the plant case, where structure is speculative too) and shared between model and target, justifying the inverse inference from target—where the functional ascriptions are commonly accepted—to model. This heuristic enables the construction of both an interpretation-sketch into the domain of information processing, and a key-sketch for imputing the properties exemplified by the model to systems commonly recognized as cognitive.
Indeed, the inverse reasoning is perhaps the most characteristic element of how the three analyzed model organisms are being extended to novel domains. In all of them, though in different forms, the research is partially guided by what we already know of the target system and establishing interpretations and keys partially depends on identifying analogous properties in the model systems. This process primarily stems from researchers’ scientific motivations and interest, as well as their background assumptions regarding e.g., the continuity of information processing in biology (see Lyon, 2006, 2020). This inverse reasoning (from target to model) has important normative consequences. Any property that is derived from target cannot be then used to justify inferences about the target, on the charge of circularity. Nevertheless, the development of a complete key can aim to provide independent backing for those properties. This notably differentiates the case of plants (Sect. 3.1) from that of developmental systems (Sect. 3.2) and bacteria (Sect. 3.3). Alpi et al. (2007) in their critical note point at the triviality of some claims of the “plant neurobiology” proponents (“Plant cells do share features in common with all biological cells, including neurons.”, emphasis added) and point that “Although at the molecular level the same general principles apply [...] this does not imply a priori that comparable structures for signal propagation exist at the cellular, tissue and organ levels.” This criticism points precisely to the lack of independent support for the interpretation-sketch proposed in “plant neurobiology”. For developmental systems and bacteria, the interpretation in terms of signalling motivates direct experimental intervention, which aims to provide such external support.
Importantly, viewing these reasoning schemas as heuristic enables tracing the (possible) errors and misrepresentations these models introduce. Wimsatt notes two particular risks of the localization heuristic: ignoring pleiotropy (multiple functions of a single structure) and division of labour (“missing other parts’ roles in the hypothesized function because they are a part of the constant context, so they are always there to provide it,” Wimsatt, 2006, p. 470). Both of those limitations are visible in the case studies analysed above:
In terms of pleiotropy, in all cases the signalling or information transfer activity is regarded as separate from metabolism, despite their close functional coupling, especially in the bacterial biofilm case. This tracks the standard views of the nervous system. Tellingly, this metabolism-information processing discontinuity is considered as a key feature in the evolution of cognition by some theoreticians (e.g., the view of cognition as “meta-metabolic activity”, see Godfrey-Smith, 2016; Moreno & Etxeberria, 2005; Moreno et al., 1997; Van Duijn et al., 2006). Nevertheless, there’s a growing body of research which posits that information processing activity of neurons cannot be decoupled or analysed separately from their metabolism (e.g., Jacob et al., 2023), especially in the case of cognitive impairment or neuropsychiatric disorders (see the review in Morella et al., 2022).13
In terms of division of labour, all three examples focus the construction of a key-sketch on a subset of neural interactions—spiking activity of the neurons. While this is in line with the dominant view of the nervous system, the keys are constructed to the exclusion of a number of different elements that are increasingly recognized as contributing to nervous system’s implementation of cognition, such as dendritic computations (London & Häusser, 2005), ephaptic coupling (Anastassiou et al., 2011; Han et al., 2018), or neuron-glia interactions (Rangel-Gomez et al., 2024).
Despite its revolutionary potential, in its focus on reasoning motivated by the localization heuristic, basal cognition (at least in the three examples analysed here) follows the received (and relatively conservative) view of what kinds of mechanisms can process information. Indeed, localization is widespread in traditional neuroscience, where cognitive functions are commonly attributed to well-delineated structures (either anatomical or dynamically connected) in the brain. But already in this traditional context it’s facing increased scrutiny (e.g., Noble et al., 2024; Westlin et al., 2023). While the flavour of localization in the basal context differs significantly, focusing primarily on a molecular (rather than anatomic) scale, it is still susceptible to the anti-reductionist arguments raised in that debate (e.g., Pessoa, 2022). And regardless of one’s views on the matter, it is important to keep track of the errors associated with structure-function localizations.
The suggested heuristic perspective on the use of model organisms is not in conflict with the stricter character of justification that Sartori (2023) argues for (or that Levy and Currie consider, though differently, 2015). While Sartori highlights how scientists go about justifying the conclusions drawn from established research organism, within a stricter inferential framework, my suggestion here applies to the early stages of research, when the organisms are first selected for a particular purpose. Nevertheless, as those heuristics are codified in the form of the local key that emerges from the key-sketch, they can continue to impact the stricter, justificatory inferences.
Interpretation- and key-sketches differ from actual interpretation functions and keys by their epistemic role in research. They are constructed as mappings, but often involve significant handwaving and heuristic justifications, which are unable to ground models’ (derivational) correctness as proper interpretations and keys do. Instead, they offer initial epistemic purchase that can guide empirical research. This is most clearly visible in the inverse reasoning patterns identified, as this target-to-model inference is initially unjustified or purely metaphorical, but can prove empirically productive. In this process, gradually, sketches are filled in or modified, and eventually can be regarded as proper interpretations and keys.
In the paper, I account for the epistemic operations involved in adopting a model organism in a novel field of research, focusing on the emerging research tradition of basal cognition. To this purpose, I have drawn from the DEKI account of scientific representation, which has been usefully applied to model organisms before. While DEKI is focused on established models, the framework it offers highlights the steps required for the introduction of a novel model.
In the case of model organisms, the introduction of novel models or adoption of models to new domains involves a number of heuristics, which justify tentative interpretation- and key-sketches. These sketches are often incomplete but structurally resemble the full mappings from the model to the domain and from the domain to the target, which figure as main components of the DEKI account, and are necessary for establishing the correctness of the models.
The concept of interpretation- and key-sketches allows for a better understanding of the relation between the research conducted within the field of basal cognition and in neuroscience more broadly. While the basal cognition explicitly tries to expand the scope of neuroscientific research, one of the most serious criticisms it faces is its inability (so far) to produce results that are directly informative for neuroscientists.
While there is little philosophical analysis of the role of model organisms in neuroscience (the important exceptions are Kaplan, 2017; Nelson, 2013, 2018), the field relies heavily on a very small subset of organisms, primarily mice (Mus musculus) and rats (Rattus norvegicus)—and research carried out directly on humans (Homo sapiens).14 While considering the domestication and captivity effects (see Chrzanowska et al., 2025; Nemati, 2024) or emphasizing the evolutionary and comparative considerations for justifying the inferences between model organisms and their targets (e.g., Cisek, 2019; Cisek & Green, 2024; Cisek & Hayden, 2022) helps in understanding the impact of this narrow evidence base, important limitations to neuroscientific knowledge follow from this small set of MOs (see Keifer & Summers, (2016); see also articles in Baldwin et al., 2015).
In this context, I propose to view the research carried out in basal cognition as expanding the collection of model organisms for neuroscience, extending the diversity of available data, and eventually strengthening the methodologies of the neuroscientific use of model organisms (in the way argued for by Cisek & Hayden, 2022). This introduces clear criteria for the adequacy and informativeness of MOs considered in basal cognition, and specifies how researchers working in other areas of neuroscience can directly benefit from the research carried out in that field. In that respect the debates over the proper definition of cognition, which tend to dominate the field, are not immediately relevant. The parallels between the models and targets can be constructed artificially, and the features may overlap only partially, while still providing important insights into the phenomena of interest. This is clearly visible, e.g., in the case of the use of D. melanogaster as a model for Parkinson’s disease.
It is so also in the three case studies analysed here. The reasoning patterns embedded in the interpretation- and key-sketches proposed by researchers enable us to see how the study of these simple model organisms can enhance our understanding of neuroscience, whether or not we are inclined to consider bacteria, plants, or individual metazoan cells as cognitive.
I want to thank Roman Frigg, Guido I. Prieto, Julia Sánchez Dorado, Lorenzo Sartori, and two anonymous reviewers for extremely helpful comments to an earlier version of the manuscript. I have also benefitted from the comments of the audiences at the symposium “The Rise of Basal Cognition” during the British Society for the Philosophy of Science conference at the University of Exeter in 2022, at the Animal Minds ECR workshop held at LSE in April 2024, and at the German Society for Philosophy of Science held at the Friedrich-Alexander-Universität Erlangen-Nürnberg in March 2025. In the last case, I’m particularly grateful to my co-symposiasts—Alejandro Fábregas-Tejeda, Roman Frigg, Sabina Leonelli, Guido I. Prieto, and Lorenzo Sartori. Further, I’m greatly indebted to Mike Levin and the members of the Levin Lab for allowing me to audit the lab meetings for a period in 2022 as an online visiting scholar, which helped me greatly in understanding the practical elements of research in basal cognition, and contemporary life science broadly. Finally, I want to express my gratitude to Nedah Nemati for our discussions of this topic and her impact on this paper and to Urtė Laukaitytė, David Harrison, and Taylor Beck for our many conversations on basal cognition.
The work on the article was enabled by the START scholarship from the Foundation for Polish Science awarded in 2024, and some of the research was conducted while I was a Writing-Up Fellow at the Konrad Lorenz Institute in Klosterneuburg, Austria, between October 2024 and March 2025.