Ensemble Graph Neural Spatial Clustering: A Robust Framework for Spatial Domain Discovery in Spatial Transcriptomics

Shuyan Guo, Yiting Bu

Keywords: spatial transcriptomics, ensemble learning, graph neural networks, spatial domain discovery, deep learning, tissue segmentation, spatial clustering

TL;DR: A robust ensemble graph neural network framework that integrates multiple spatial GNN models to achieve stable, accurate, and biologically coherent spatial domain discovery in transcriptomics data.

Abstract: Spatial transcriptomics integrates gene expression measurements with spatial coordinates, enabling the study of tissue organization and cellular architecture. A key analytical step is spatial domain discovery, which aims to segment tissues into meaningful biological regions based on molecular and spatial patterns. Existing clustering methods often suffer from instability due to noise, stochastic model training, and poor generalizability across datasets. To address these challenges, we propose a robust ensemble graph-neural framework that aggregates multiple spatial GNNs trained with distinct random seeds, feature bagging, and graph perturbations. Each base model jointly encodes gene expression and spatial neighborhood information using multi-view graph convolutions and a Zero-Inflated Negative Binomial (ZINB) likelihood. A consensus pipeline incorporating Hungarian label alignment, co-association matrices, and spectral clustering fuses predictions into a stable final segmentation. Across 12 human DLPFC sections and a Human Breast Cancer dataset, our method consistently achieves superior clustering accuracy, higher ARI/NMI scores, and more coherent laminar or tumor–stromal boundaries compared with state-of-the-art methods. These findings highlight ensemble GNN modeling as a powerful and generalizable strategy for spatial domain discovery in high-resolution tissue transcriptomics.

Submission Number: 67