Commentary on Pang et al. (2023) Nature

Abstract Pang et al. (2023) observe that the geometric eigenmodes, derived from the shape of the cortical surface, are better at reconstructing patterns of both spontaneous and stimulus-evoked activity, when contrasted with three alternative connectome-based models including structural connectome derived eigenmodes. Based on this observation they propose that geometric eigenmodes offer a good model for explaining brain function, noting that “ wave dynamics offer a more accurate and parsimonious mechanistic account of macroscale, spontaneous cortical dynamics captured by fMRI ”. They then question the prevailing view that brain activity is “ localized to focal, spatially isolated clusters ” and it is driven by “ intricate patterns of anatomical connections ”. While the observation that geometric properties fit brain activity well is intriguing, we argue that accepting geometric eigenmodes as a model for brain function risks the logical fallacy of “affirming the consequent”. A representation that effectively describes the underlying geometry is inherently adept at fitting patterns within that geometric space; it does not necessarily shed light on mechanisms of the brain’s functional attributes. To this end, we provide two lines of empirical results: (a) Basic parcel-based representations, which capture localized structures, can reconstruct activity patterns as well. (b) Geometric eigenmodes demonstrate a high flexibility when fit to a range of manipulated patterns, which evokes the danger of overfitting. Based on those results, theoretical considerations, and previous data we argue that more consideration is needed regarding “parsimony, robustness and generality of geometric eigenmodes as a basis set for brain function”. While we recognize the potential role of the brain’s geometry in influencing its dynamics, assertions regarding its efficacy should be weighed against the performance of simpler models, an inherent risk of overfitting and anatomical evidence. Pang et al. 1 put forth harmonic modes derived from the brain’s geometry as a previously underrecognized model to explain brain-wide dynamics. Their reconstruction framework relies on multiple linear regression to fit brain patterns using a basis set and then calculating Pearson’s correlation between original data and fitted data, both parcellated using an atlas with 180 parcels in each hemisphere 2 . In addition to the overarching challenge of accepting any model as an actual reflection of the real world 3 , we here provide empirical results and theoretical arguments that highlight the need for further consideration regarding geometric eigenmodes as a model of macroscale brain activity.