Separate Neural Representations for Physical and Communicative Social Interactions: Evidence from Data-driven Voxel Decomposition

Rapidly recognizing and interpretating others' social interactions are fundamental abilities for humans. However, despite the wide range of social interactions encountered in real-world perception and emerging behavioral evidence suggesting qualitative differences between interaction types, neural evidence supporting this multifaceted nature of social interaction recognition has been largely absent in prior hypothesis-driven approaches. Here, we employed a data-driven voxel decomposition technique to dissociate distinct neural responses to different types of social interactions in the lateral visual pathway during dynamic social vision. Using an fMRI dataset of 200 three-second video clips depicting two individuals engaged in social activities, our analysis identified two components with distinct functional profiles in the lateral visual pathway that were consistent across subjects. One component, predominantly weighted in posterior regions corresponding to the lateral occipitotemporal cortex (LOTC), including the extrastriate body area (EBA), responded strongly to videos depicting physical interactions between people; while the other, weighted in anterior regions, specifically the superior temporal sulcus (STS), responded strongly to videos depicting communicative interactions. We replicated the findings using the BOLD Moments dataset, by analyzing fMRI responses to 601 three-second video clips featuring one or more individuals engaging in a broader range of everyday activities. Together, our findings suggest that in the lateral visual pathway, posterior regions in the LOTC and anterior regions in the STS differentially respond to physical and communicative interactions, offering new insights into the hierarchical functional organization of third-person social interaction perception in the lateral social stream. Furthermore, our findings also highlight the distinct advantage of data-driven approaches in uncovering neural representations that hypothesis-driven methods may fail to detect.