ModelistsGCN: a multimodal graph convolutional network framework for single-cell spatial transcriptomic cell typing.

Abstract

Spatial transcriptomics technologies currently face a trade-off between spatial and molecular resolution. High-resolution single-cell methods such as MERFISH and Expansion Sequencing (ExSeq) resolve individual cells but typically profile limited gene panels, restricting the direct application of transcriptome-based cell typing strategies developed for single-cell RNA sequencing. Consequently, accurate cell-type identification in spatial single-cell data remains challenging when transcriptomic coverage per cell is sparse. Here, we introduce ModelistsGCN, a semi-supervised multimodal graph convolutional framework that integrates gene expression, spatial proximity, and quantitative cellular morphology for spatial single-cell cell typing. ModelistsGCN uses a small set of high-confidence representative cells, defined by marker-gene enrichment, to guide clustering while preserving flexibility in identifying cell types. By incorporating spatial neighborhood structure and interpretable morphological features alongside gene expression, the method compensates for limited molecular coverage and strengthens cell-type discrimination. Across a well-annotated ExSeq mouse visual cortex dataset and metastatic breast cancer tissues profiled by MERFISH and ExSeq, ModelistsGCN consistently demonstrated higher agreement with reference annotations, improved cluster separation, and stronger marker-gene coherence compared to existing spatial clustering approaches. ModelistsGCN improves cell-type inference in single-cell spatial transcriptomic datasets with limited gene coverage.