The target article makes a case that metabolic constraints should play a more prominent role in cognitive modeling. While we agree with this premise, we examine why explicit metabolic modelling remains uncommon, using Marr’s levels of analysis as a clarifying framework. In particular, Marr’s perspective helps highlight how metabolic constraints can shape models, even when they are not represented explicitly.

David Marr famously proposed that information processing systems can be understood at three levels: the computational level (what abstract function the system performs), the algorithmic level (what representations and elementary operations realize this function), and the implementational level (how these operations are physically implemented) (Marr, Reference Marr1982). Importantly, in the case of brains, the implementation level can entail multiple biological dimensions, such as electrical activity, synaptic transmission, development, connectivity, and metabolism. Although this framework is not without limitations, it remains a useful way to separate different types of explanations in computational neuroscience (Lengyel, Reference Lengyel2024; Poggio, Reference Poggio2012).

Equipped with this terminology, we can rephrase the main thesis of the target article: the metabolic aspects of implementation should be included in cognitive models more often, because they provide useful constraints on other processes, both within the implementation level (e.g., membrane and synaptic biophysics) and across levels (algorithm and computation). While we agree with the diagnosis, we find it useful to consider why metabolism is not modeled more often.

One reason is methodological: researchers tend to model what can be measured and fitted. During the last few decades, neuroscientists have routinely collected vast amounts of data on neuronal activity in vivo (spiking, local field potentials, hemodynamic response, etc.). Recording energy expenditure (e.g., using ATP sensors) remains much more challenging, even despite recent methodological advances (Zhao et al., Reference Zhao, Tanourlouee, McCracken and Williams2025). As a consequence, the researchers willing to model metabolic cost often lack ground truth data. Let us consider a concrete example: estimating the energy cost of a reward-related burst in a dopamine neuron. Because direct metabolic data are not available, one might in principle use the ATP cost per spike obtained from cortical neurons (Howarth, Gleeson, & Attwell, Reference Howarth, Gleeson and Attwell2012). However, dopamine neurons likely consume much more energy due to their complex morphology (Pissadaki & Bolam, Reference Pissadaki and Paul Bolam2013). More importantly, it is challenging to estimate how much energy could be saved by reducing spiking (Kondrakiewicz & Nawrocka, Reference Kondrakiewicz and Nawrocka2025), as recent models suggest that dopamine neurons may need to fire periodically to prevent toxic buildup of free radicals (Chintaluri & Vogels, Reference Chintaluri and Vogels2023). This illustrates how basic metabolic questions remain unanswered, leading many researchers to avoid metabolic assumptions altogether.

Another reason for ignoring metabolism is conceptual. Computational models are often designed to capture simple, general principles that approximate the behavior of real-world complex systems, while remaining amenable to analysis and interpretation. A well-known example is the backpropagation algorithm, originally developed to update synaptic weights in deep neural networks for learning complex tasks (Rumelhart, Hinton, & Williams, Reference Rumelhart, Hinton and Williams1986). Although long assumed to be biologically unrealistic, it became the standard approach for training neural networks due to its effectiveness. Only recently have researchers discovered biologically plausible learning rules that can approximate aspects of the same computation (Lillicrap et al., Reference Lillicrap, Santoro, Marris, Akerman and Hinton2020). This illustrates how abstracting away from implementation details to extract general principles can yield insights relevant for understanding computation in the brain.

Despite these limitations, the relative scarcity of explicit metabolic models does not imply that metabolic constraints play no role in current modeling practice. Instead, metabolic constraints often influence models implicitly.

First, many computational models are fitted to biological data. Those data already reflect the energetic limits of real neurons. If a model replicates, for instance, spiking statistics from experimental data, we can conclude that the model output is broadly consistent with the metabolic constraints on spiking, even if these constraints are not modeled explicitly. However, such agreement with observed activity patterns does not uniquely determine the underlying mechanisms, as different parameter configurations can give rise to similar activity patterns (Prinz, Bucher, & Marder, Reference Prinz, Bucher and Marder2004). This in turn means that these configurations may differ in their associated energy costs, or achieve similar costs through different mechanisms, further complicating the relationship between function and metabolism (Deistler, Macke, & Gonçalves, Reference Deistler, Macke and Gonçalves2022). That said, model fitting remains a powerful approach for assessing whether a model is generally consistent with implementation-level constraints without having to commit to specific and potentially unknown quantitative assumptions about those constraints.

Second, computation- and algorithmic-level modeling often invokes efficiency as a guiding principle. In particular, the efficient coding hypothesis is built on the idea that, due to evolutionary pressure, biological systems use their limited resources to represent information as efficiently as possible, for example, by reducing redundancy in neural representations (Barlow & Rosenblith, Reference Barlow and Rosenblith1961; Simoncelli & Olshausen, Reference Simoncelli and Olshausen2001). The target article correctly classifies these approaches as “generic,” since energetic costs are rarely explicit. However, these approaches can remain effective even without ground-truth metabolic data. For example, efficient coding models correctly predicted that the early visual system reduces spatial and temporal correlations to improve coding efficiency (Atick & Redlich, Reference Atick and Norman Redlich1990; Dan, Atick, & Reid, Reference Dan, Atick and Reid1996).

To sum up, we argue that the common practice of omitting explicit metabolic details reflects not a lack of interest, but the difficulty of integrating metabolism with other levels of analysis. At the same time, many existing modeling approaches already capture metabolic constraints implicitly. Does this mean that explicitly incorporating metabolic constraints is unnecessary, or that it will remain challenging? We do not think so. Interest in brain metabolism is growing, which will hopefully lead to the collection of much needed data on the topic (Bergmann et al., Reference Bergmann, Mousaei, Rizzoli and Tchumatchenko2025; Padamsey & Rochefort, Reference Padamsey and Rochefort2023). Combined with advances in computational methods and the growing public sharing of resources, this progress should make metabolic modeling more accessible than ever before. Ideally, these technical advances will enable researchers to effortlessly add a metabolic layer to existing computational models. This, in turn, would open the door to addressing questions that directly link levels of analysis, such as the relationship between algorithmic and metabolic efficiency, and the extent to which metabolic constraints shape or restrict the space of possible computations (Jedlicka, Bird, & Cuntz, Reference Jedlicka, Bird and Cuntz2022; Pallasdies et al., Reference Pallasdies, Norton, Schleimer and Schreiber2021). We anticipate that the coming decade will bring meaningful progress toward answering these and related questions.