Title: Functional Connectivity Graph Neural Networks
Speaker: Yang Li
Abstract:
Real-world networks, such as the human brain, feature complex topologies that balance localized specialization with global integration. The brain’s connectome exemplifies this balance through its structural and functional connectivity. These complementary modalities provide a comprehensive view of its intricate organization. Graph-based frameworks offer a versatile method for analyzing such complex systems, with multi-modal approaches proving essential for accurately representing diverse connectivity types. However, existing graph neural networks (GNNs) often focus on single-modality structural connections, limiting their ability to capture long-range functional interactions. This talk will propose a novel functional connectivity block that incorporates functional topology into neural network architectures using persistent graph homology, a mathematically stable representation of global graph topology. Building on this, I will introduce innovative functional connectivity graph neural networks (FC-GNNs), a multi-modal architecture that integrates structural and functional modalities for graph-level classification. Experiments across diverse biological datasets demonstrate significant performance improvements, highlighting the transformative potential of multi-modal approaches for advancing complex network analysis.
