Abstract
AbstractInter-subject modeling of cognitive processes has been a challenging task due to large individual variability in brain structure and function. Graph neural networks (GNNs) provide a potential way to project subject-specific neural responses onto a common representational space by effectively combining local and distributed brain activity through connectome-based constraints. Here we provide in-depth interpretations of biologically-constrained GNNs (BGNNs) that reach state-of-the-art performance in several decoding tasks and reveal inter-subject aligned neural representations underpinning cognitive processes. Specifically, the model not only segregates brain responses at different stages of cognitive tasks, e.g. motor preparation and motor execution, but also uncovers functional gradients in neural representations, e.g. a gradual progression of visual working memory (VWM) from sensory processing to cognitive control and towards behavioral abstraction. Moreover, the multilevel representations of VWM exhibit better inter-subject alignment in brain responses, higher decoding of cognitive states, and strong phenotypic and genetic correlations with individual behavioral performance. Our work demonstrates that biologically constrained deep-learning models have the potential towards both cognitive and biological fidelity in cognitive modeling, and open new avenues to interpretable functional gradients of brain cognition in a wide range of cognitive neuroscience questions.HighlightsBGNN improves inter-subject alignment in task-evoked responses and promotes brain decodingBGNN captures functional gradients of brain cognition, transforming from sensory processing to cognition to representational abstraction.BGNNs with diffusion or functional connectome constraints better predict human behaviors compared to other graph architecturesGraphic AbstractMultilevel representational learning of cognitive processes using BGNN
Publisher
Cold Spring Harbor Laboratory