Spatial Collinearity Constrains Multivariate Molecular-Enriched Network Estimation

Abstract

Analyses of neuroimaging data increasingly leverage the distribution of neurotransmitter receptors derived from positron emission tomography (PET) to bridge the gap between micro- and macro-scale brain function. However, these receptor maps are highly spatially overlapping, which can give rise to interpretive and analytical challenges. Here, we systematically investigate the impact of spatial collinearity among PET maps in the context of Receptor-Enriched Analysis of functional Connectivity by Targets (REACT), a method that uses receptor maps as spatial regressors to derive subject-level molecular-enriched functional connectivity networks. Exhaustive combinatorial analysis across 19 receptor and transporter maps showed that collinearity scales rapidly with the number of receptors modelled simultaneously, and that the underlying correlation structure was relatively stable across parcellation scales, reflecting the intrinsic organisation of neurotransmitter systems. Using test–retest fMRI data from the Human Connectome Project, we demonstrate that modelling greater numbers of receptors degrades the reliability of molecular-enriched networks derived from conventional multivariate REACT models, and that collinearity among receptor maps drives this degradation. An alternative univariate approach, in which each receptor is modelled independently, yielded more reliable networks and, when applied to a within-subjects study of LSD compared to placebo, better recovered the role of the 5-HT_2A receptor in LSD's neural effects. These findings identify spatial collinearity as a fundamental constraint on multivariate molecular-enriched network estimation. We therefore recommend univariate models as the default, with multivariate models used only where specifically motivated and reported alongside the univariate result.