Neuromodulatory systems partially account for the topography of cortical networks of learning under uncertainty

Abstract

Learning in dynamic and stochastic environments is notoriously difficult. Neuromodulatory systems may shape this process, thereby constraining where learning-related neural activity emerges according to the spatial distribution of receptors and transporters across the brain. However, the extent of this constraint, and which neuromodulatory systems contribute the most, remains unclear. Here, we focus on fMRI data from four probabilistic learning studies that was combined with a Bayesian ideal observer model. This model formalizes latent computational variables such as confidence and surprise that drive human learning. We show that the functional correlates of confidence, and to a lesser extent surprise, exhibit strong spatial invariance across tasks. This invariance suggests that these functional correlates of learning reflect a stable cortical organization largely independent of sensory modality and task structure. We found that this invariance aligns with the cortical chemoarchitecture using 20 PET-derived receptor and transporter density maps. Distinct receptor and transporter patterns mapped onto confidence- and surprise-related activity, including both known catecholamine pathways and novel opioid associations. Together, our findings provide evidence for a neuromodulatory account of adaptive learning and offer receptor-level hypotheses. More broadly, our approach provides a general framework to understand how neuromodulatory systems shape diverse cognitive processes.