Background

Understanding cellular communication is crucial for constructing a cell–cell interaction network (CCI), which allows researchers to investigate the roles of different cells in biological processes and diseases. A common method for analyzing CCI involves studying the interactions between secreted ligands and their corresponding receptors (LR pairings), as these interactions are essential for signal transmission.

However, CCIs also occur through direct cell–cell contact, the extracellular matrix (ECM), and the secretion of signaling molecules [1]. Focusing solely on LR pairings overlooks the spatial context in which these interactions occur. With the advent of spatial transcriptomics (ST) data, incorporating spatial information can help address this limitation. Several methods for reconstructing CCI using ST data have been developed. For instance, Giotto [2] constructed a spatial grid based on cell coordinates to model proximal interactions but did not account for distal interactions. MISTy [3] utilizes multiple perspectives (intrinsic, local, tissue) to enhance cell knowledge for CCI inference. DeepLinc [4] employs a graph neural network (GNN) [5] to create a spatial proximity graph, using KNN [6] to identify both proximal and distal interactions. Despite these advances, modeling ST data solely at the cell level can lead to false positives and negatives due to a lack of downstream information. Incorporating downstream data, such as intracellular signaling, gene regulation, and protein changes can provide crucial insights for understanding actual CCI; CLARIFY [7] combines cellular and genetic information to build knowledge graphs for GNN inputs, effectively capturing structural features important for CCI inference. However, solely using GNNs is limiting, as they often aggregate local information, hindering the capture of global context. GNNs also struggle with similar nodes, which may lead to a loss of features’ specificity.

To overcome these challenges, we propose a novel approach called triple-enhancement-based graph neural network (TENET), designed as a comprehensive architecture for CCI reconstruction. TENET employs a graph convolution network (GCN) [8] backbone to extract global features from the data at multiple resolutions, generating robust latent feature embeddings. These embeddings then undergo a triple-enhancement mechanism that restores cell specificity that may have been lost during the GCN's global aggregation. The first enhancement mitigates the GCN's over-smoothing problem, while the secondary and tertiary enhancements refine the embeddings both locally and globally. Finally, a denoising loss function is introduced to tackle the challenges posed by noisy, low-quality data often encountered in small molecule experiments. This multi-scale enhancement strategy enables TENET to accurately and robustly reconstruct cell–cell interactions, addressing the limitations of previous methods and opening new avenues for researchers to explore the complexities of CCI.