Haueis and Colaço provide a timely and compelling case for integrating metabolic considerations into cognitive modeling. Their distinction between evaluative and generative uses of metabolism offers an important advance beyond metaphor toward substantive constraint. However, their account remains underspecified with respect to scale. In practice, metabolic knowledge is not equally actionable for cognitive modelers, because it is often articulated at scales that are mismatched with the explanatory target of cognitive science. We argue that a scale mismatch problem pervades current discourse: metabolic data are typically either too molecular or too global to directly constrain models of cognitive systems. We propose that mesoscale metabolism, the metabolic phenotypes of functional neural circuits and networks, provides a more promising interface for metabolically informed cognitive modeling.

Cognitive models may operate at the level of functional systems, circuits, and networks. By contrast, much of metabolic knowledge resides at extremes of scales. At the molecular level, precise metrics such as ATP cost for synaptic transmission (Harris, Jolivet, & Attwell, Reference Harris, Jolivet and Attwell2012) or plasticity (Karbowski, Reference Karbowski2019) are indispensable for biophysical realism, yet they rarely map directly onto the functional descriptions employed in cognitive models. At the global level, whole-brain energy budgets (Attwell & Laughlin, Reference Attwell and Laughlin2001; Engl & Attwell, Reference Engl and Attwell2015) provide important upper bounds on information processing but they are too coarse to adjudicate between competing models of specific cognitive capacities. This mismatch helps explain why metabolism is often treated as “background” in cognitive modeling: not because it is irrelevant, but because it is typically presented at scales incommensurate with cognitive explanation.

A more promising interface lies at the mesoscale: the level of functional neural circuits, pathways, and networks that directly implement cognitive operations. Certain metabolic phenomena are inherently mesoscale. Aerobic glycolysis, for example, is preferentially upregulated in specific functional networks and covaries with their computational roles (Theriault et al., Reference Theriault, Shaffer, Dienel, Sander, Hooker, Dickerson, Barrett and Quigley2023). Likewise, neurovascular coupling and region-specific cerebral blood flow track the metabolic demands of distributed cognitive operations rather than local spiking alone. Crucially, recent large-scale mapping of mitochondrial respiratory capacity demonstrates that different mitochondrial profiles, including mitochondrial density, energy-transformation capacity, and oxidative potential, are systematically heterogeneous across the human brain (Figure 1) (Mosharov et al., Reference Mosharov, Rosenberg, Monzel, Osto, Stiles, Rosoklija, Dwork, Bindra, Junker, Zhang, Fujita, Mariani, Bakalian, Sulzer, De Jager, Menon, Shirihai, Mann, Underwood, Boldrini, de Schotten and Picard2025). This network- and region-specific patterning implies that mitochondrial energetics constitute a mesoscale property of neural systems, aligning with functional specialization. These phenomena are therefore not merely cellular processes, but metabolic phenotypes of cognitive systems.

Figure 1.

Mesoscale organization of mitochondrial energetics in the human brain. Brain-wide mapping reveals pronounced heterogeneity in mitochondrial respiratory capacity across the human brain. Data are referenced to the brain maps of mitochondrial profiles reported by Mosharov et al. (Reference Mosharov, Rosenberg, Monzel, Osto, Stiles, Rosoklija, Dwork, Bindra, Junker, Zhang, Fujita, Mariani, Bakalian, Sulzer, De Jager, Menon, Shirihai, Mann, Underwood, Boldrini, de Schotten and Picard2025). MitoD denotes mitochondrial tissue density. TRC reflects tissue respiratory capacity. MRC (mitochondrial respiratory capacity) indexes TRC normalized for MitoD.

Figure 1. Long description

Panel A: A heat map labeled MitoD showing mitochondrial tissue density across different regions of the human brain. The color scale ranges from 0.9 to 1.3, with lighter colors indicating higher density. Panel B: A heat map labeled TRC showing tissue respiratory capacity across different regions of the human brain. The color scale ranges from 0.9 to 1.5, with lighter colors indicating higher capacity. Panel C: A heat map labeled MRC showing mitochondrial respiratory capacity, which is TRC normalized for MitoD, across different regions of the human brain. The color scale ranges from 0.9 to 1.5, with lighter colors indicating higher capacity.

Mesoscale metabolic features therefore offer constraints that are both biologically grounded and cognitively specific. They allow evaluative metabolic considerations to move beyond generic efficiency claims toward capacity-specific adjudication. For example, models positing sustained neural activity can be evaluated against evidence for rhythmic or transient metabolic engagement at the circuit level. Likewise, generative modeling can treat mesoscale metabolic regimes as sources of architectural and dynamical hypotheses, such as metabolically informed gating or gain control in predictive coding frameworks. In this way, metabolism becomes a direct driver of model structure, not merely a post hoc plausibility check.

This perspective operationalizes Haueis & Colaço’s framework. For evaluative modeling, the specification of relevant metabolic constraints should prioritize mesoscale phenotypes tied to the target capacity. For generative modeling, network-level metabolic features provide natural parameters for model construction. Crucially, a mesoscale focus enables genuine multiscale integration: molecular cost and global budgets can be respected without collapsing explanation to either extreme.

We therefore agree that metabolism matters for cognitive modeling but argue that its impact depends critically on scale. Without an explicit mesoscale focus, metabolic considerations risk remaining either too abstract to constrain or too detailed to integrate. Mesoscale metabolism is where computation and energy genuinely meet, as cognitive algorithms are implemented in biological networks with finite, structured energetic resources. Future work should adopt a scale-sensitive framework that explicitly identifies which metabolic features constrain which levels of cognitive explanation. The path to metabolically constrained cognitive modeling runs through the mesoscale.