The brain architecture of punishment learning

Learning to suppress actions that lead to harmful consequences is a fundamental survival skill. While much research has focused on Pavlovian fear—where environmental cues predict threat—less is known about how the brain learns to regulate self-generated actions that cause harm. This paper investigates the neural architecture of punishment learning, moving beyond simple valuation models to map the multiscale brain networks that allow animals to selectively suppress punished actions while maintaining rewarded ones.

Using whole-brain Fos mapping, the authors identified that punishment learning does not rely on a single brain region but rather triggers a large-scale reorganization of functional networks. While the brain maintains a modular, small-world topology, the specific community membership of various regions shifts significantly. A core network of hub regions—the basolateral amygdala (BLA), zona incerta (ZI), and the rostral linear nucleus of the midbrain—emerged as critical for this process. In silico network deletion and multisite chemogenetic inhibition confirmed that these regions are necessary for adaptive punishment learning. Interestingly, these hubs play dissociable roles: the BLA is essential for within-session adaptation, while the ZI and midbrain structures are required for between-session retention.

To understand the cellular basis of this network, the researchers used spatial transcriptomics to identify punishment-associated gene programs. They found that punishment recruits diverse neuronal subclasses, including specific glutamatergic neurons in the BLA and GABAergic populations in the ZI and midbrain. Furthermore, longitudinal calcium imaging of BLA ensembles revealed that these neurons dynamically encode punishment. Specifically, distinct ensembles partition actions from outcomes, while others flexibly remap their activity to represent punished actions within an aversive neural space. This suggests that the BLA acts as an integrative hub, transforming action representations based on their consequences.