The School of Computer Science is pleased to present…
Graph Neural Networks for Cell Type Prediction: A Heterophily-Aware Approach in Single-Cell Analysis
PhD Dissertation Proposal by: Mahshad Hashemi
Date: Tuesday, July 29th, 2025
Time: 11:30 am - 1:00 pm
Location: Essex Hall, Room 122
Graph Neural Networks (GNNs) have shown great promise in modelling cell-cell communication for single-cell RNA-sequencing (scRNA-seq) data, particularly by capturing complex topological patterns beyond traditional similarity-based clustering. In our initial work, we demonstrated that standard GNNs—especially heterophily-aware models such as H2GCN and GBK-GNN—can effectively model intercellular signalling networks derived from ligand–receptor interactions. Studies established that cellular communication graphs are often heterophilic, where biologically dissimilar cells interact through specific signalling pathways. To better navigate this setting, we developed HeteroGraphNet, a GNN architecture that integrates adaptive random walks with a bi-kernel aggregator to preserve chain-like signalling structures and distinguish between homophilic and heterophilic interactions during message passing.
In this presentation, we will also highlight the limitations of pairwise connectivity in conventional GNNs. As a future direction, we propose the use of hypergraph neural networks, which enable modelling higher-order, multi-cell interactions and offer a more biologically grounded and expressive framework for capturing the combinatorial complexity of ligand-receptor-mediated communication.
Internal Reader: Dr. Pooya Moradian Zadeh
Internal Reader: Dr. Jianguo Lu
External Reader: Dr. Mitra Mirhassani
Advisor(s): Dr. Luis Rueda and Dr. Alioune Ngom
