The School of Computer Science is pleased to present…
Advances in Cell-Type Identification using Heterophily-Aware Graph and Hypergraph Neural Networks on Single-Cell Transcriptomics
Abstract:
Single-cell RNA sequencing (scRNA-seq) enables detailed cell-type identification, but most methods rely only on expression similarity and overlook that cells communicate through ligand–receptor (LR) interactions. LR-based graphs and hypergraphs capture this biology but are highly heterophilic, causing standard GNNs to oversmooth and lose class separability. This dissertation develops models that remain accurate under such conditions. First, we construct LR-informed graphs for Baron pancreas datasets, quantify their low homophily, and benchmark classical and heterophily-aware GNNs under a consistent evaluation protocol. Second, we introduce HeteroGraphNet, which uses adaptive random walks with restart and a gated bi-kernel design to capture informative neighborhoods while separating homophilic and heterophilic signals. Finally, we extend LR modeling to hypergraphs to represent higher-order communication events. Together, these contributions form a reproducible pipeline for biologically grounded, heterophily-aware graph and hypergraph neural network cell-type prediction.
Keywords: Single-cell RNA-seq, L-R Communication, Heterophily, Graph Neural Networks, Hypergraph Neural Networks