Han, ChengxuChengxuHanSong, ZhangdiZhangdiSongXu, ZimuZimuXuChen, JiaxingJiaxingChen2025-12-222025-12-222025-12-15https://boris-portal.unibe.ch/handle/20.500.12422/225841Accurately inferring cell-cell interactions from spatial transcriptomics data remains challenging due to tissue complexity and spatial heterogeneity. Recent deep learning models have started to combine different types of cellular information, such as spatial proximity, gene expression similarity, ligand-receptor signaling, and in some cases, gene regulatory networks. However, they often treat a cell's external interactions and internal gene regulation separately, merging them at the final step using simple concatenation or addition. This limits the model's ability to capture how cell-cell communication and internal molecular states are connected. In this paper, we present MAGNET (Multi-view Graph Autoencoder with Cell-Gene Attention Network), a framework that reconstructs cell-cell interaction networks by building multiple biological graphs and developing a Cell-Gene attention module to link a cell's environment with its internal gene activity in a unified representation. On benchmark datasets (seqFISH, MERFISH, STARMAP), MAGNET demonstrates superior performance in reconstructing cell-cell interaction networks, achieving an Average Precision(AP) of 0.901 on the seqFISH dataset and outperforming TENET by 0.185. The Cell-Gene Attention module is critical to MAGNET's performance, as its removal alone caused the AP on the seqFISH dataset to drop from 0.901 to 0.521. Applied to a breast cancer dataset, MAGNET found functional heterogeneity among cancer cells, distinguishing clusters with molecular signatures for either immune evasion or autonomous tumor growth.en000 - Computer science, knowledge & systemsarticle10.48620/933354139702610.1371/journal.pcbi.1013810