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

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics offers cellular resolution but faces a trade-off between spatial and molecular detail.
• High-resolution methods like MERFISH and Expansion Sequencing (ExSeq) profile limited gene panels, complicating cell-type identification.
• Sparse transcriptomic coverage per cell hinders accurate cell typing in spatial single-cell data.

Purpose of the Study:

• To introduce ModelistsGCN, a semi-supervised multimodal graph convolutional framework for spatial single-cell cell typing.
• To integrate gene expression, spatial proximity, and cellular morphology for improved cell-type inference.
• To address the challenge of accurate cell identification in spatial transcriptomic datasets with limited gene coverage.

Main Methods:

• Developed ModelistsGCN, a graph convolutional framework utilizing gene expression, spatial location, and cell morphology.
• Employed a semi-supervised approach using high-confidence cells to guide clustering.
• Integrated spatial neighborhood information and morphological features to compensate for sparse gene expression data.

Main Results:

• ModelistsGCN demonstrated higher agreement with reference annotations on mouse visual cortex and breast cancer datasets.
• The method showed improved cluster separation and stronger marker-gene coherence compared to existing approaches.
• Successfully enhanced cell-type inference in spatial single-cell transcriptomic data with limited gene coverage.

Conclusions:

• ModelistsGCN effectively integrates multimodal data for robust spatial single-cell cell typing.
• The framework overcomes limitations of sparse gene expression data in high-resolution spatial transcriptomics.
• Offers a significant advancement for cell-type identification in complex biological tissues.