scExGraph: Explainable graph neural network for predicting tumor environment components with single-cell sequencing data

Single-cell RNA sequencing (scRNA-seq) enables high-resolution analysis of the tumor microenvironment (TME), yet it faces key challenges in cross-patient cell type annotation and interpretable cell interaction analysis. This study introduces scExGraph, a graph neural network framework that combines adversarial graph domain adaptation and dynamic subgraph learning. Initially, a cell-cell graph is constructed via KNN, and a dual-branch graph convolutional encoder (GCN) is used to disentangle domain-specific and shared features, with adversarial training and adjacency matrix reconstruction ensuring topological consistency. Subsequently, a random attention mechanism dynamically adjusts edge weights, and KL divergence constraints generate interpretable subgraphs to identify key cell nodes. Finally, based on experimentally validated tumor-immune interaction genes, t-tests analyze these key cell nodes to identify critical cell signaling pathways affecting immune responses. Experiments across colorectal, non-small cell lung, and breast cancer datasets (88,507 cells) show scExGraph achieves an average accuracy of 0.918 in cross-patient annotation, significantly better than the benchmark GCN, and identifies immune regulatory genes like CEACAM1 and USP15. This research offers an explainable graph learning framework for decoding TME heterogeneity, balancing computational efficiency and biological significance. The source code are available at: https://github.com/an-xing456/scExGraph.