From cortical and white matter structure to meaning: a brain-constrained neural network of semantic grounding in...

The brain basis of semantic and conceptual processing is complex and difficult to explain. Specific word categories selectively engage some cortical areas, whereas others function as semantic hubs processing words across all categories. Beyond general neurobiological principles, cortical areas and their connectivity via white matter tracts are essential for determining an area's role in language and semantic processing. Yet, most neural network models fall short of capturing the brain's anatomical architecture, limiting their ability to provide mechanistic explanations. Here, we present a brain-constrained neural network model of 12 frontotemporal and occipital cortices constrained by tractography-derived structural connectivity, extending previous modelling work based primarily on literature-derived connectivity. Semantic circuits emerged spontaneously across the modelled cortical regions by means of Hebbian correlation learning, exhibiting distinct topographies: action words engaged fronto-central motor regions, while object words preferentially involved the primary visual area, replicating a range of neural activation patterns from neuroimaging studies. Crucially, regions central in the neural architecture, the anterior temporal and inferior prefrontal cortices, showed category-general semantic processing, consistent with a semantic hub function. A novel prediction concerns the potential hub-like contribution of the secondary temporo-occipital region, although this effect showed variability across tractography-derived structural connectivity variants. Correlation analyses further revealed that regions with richer inter-areal connectivity developed higher neural matter densities of the semantic circuits. Taken together, these findings demonstrate that brain-constrained neural models with increased biological realism at the white matter level can provide a mechanistic account of how distributed semantic representations emerge across multimodal hub, sensorimotor, and language regions.