ScGraphTrans: Pathway-Guided Graph Learning and Domain Adaptation for Cell Type Annotation in Single-Cell RNA-seq.

Abstract

The tumor microenvironment (TME) is a complex ecosystem in which intercellular communication regulates tumor progression and therapeutic response. Yet inferring cell-cell interactions from non-spatial scRNA-seq remains challenging due to incomplete ligand-receptor databases and inaccurate cell type annotations. Here, we propose scGraphTrans, a graph neural network framework that integrates functional state pseudo-labels, graph structure learning, and graph domain adaptation to improve both cell type annotation and communication inference. Pathway activity scores across 14 cancer-relevant processes (e.g., angiogenesis, apoptosis, cell cycle) are used as pseudo-labels to refine cell-cell graphs, capturing functional proximity beyond geometric similarity. A domain adaptation module further aligns embeddings across patients, enhancing cross-individual generalization. Evaluated on 38,667 cells from 15 individuals across three cancers, scGraphTrans achieved an average accuracy of 84.28%, surpassing state-of-the-art baselines while maintaining robustness across heterogeneous datasets. Statistical validation demonstrated recovery of disease-specific gene interactions (e.g., LGALS1-SUSD2 in breast invasive carcinoma and BIRC5-CASP6 in colorectal cancer) without prior ligand-receptor supervision. The source code and data used in this paper can be found in https://github.com/LiYuechao1998/scGraphTrans. Our framework thus provides an interpretable and generalizable solution for TME analysis, offering insights into biomarker discovery and therapeutic strategies.