A team led by researchers at the Princess Margaret Cancer Centre in Toronto have introduced a powerful new approach to studying how cells communicate with one another. Their deep learning framework, called GraphComm, uses single-cell RNA sequencing (RNA-seq) data to uncover the intricate web of molecular signals that allow cells to “talk” and coordinate their behavior in health and disease.

Cells constantly exchange messages through molecules like ligands and receptors, forming communication networks that regulate everything from immune responses to tissue repair. Traditional methods often miss the full picture because they treat genes as independent actors rather than as parts of larger, interconnected systems. GraphComm overcomes this limitation by using a graph-based model that integrates information about gene activity, protein interactions, and even cell location to predict how cells are interacting.

Schematic outlining the architecture of GraphComm to make CCC predictions from scRNAseq data

(A) GraphComm utilises a scRNAseq dataset with a curated ligand-receptor database to construct a directed graph reflective of the CCC ground truth. Feature Representation is implemented to extract positional information for ligands and receptors within the directed graph and scaled accordingly with protein complex and pathway information. (B) GraphComm constructs a second directed graph representing the relationship between cell groups and source/target proteins. Annotating this second directed graph with transcriptomic information, cell group information and positional features from the Feature Representation learning step to receive updated numerical node features via a Graph Attention Network. Node Features can then be used via inner product to compute communication probability for all possible ligand receptor pairs. (C) Computed communication probabilities via the Graph Attention Network can be combined with the second directed graph to construct ligand-receptor links with top-ranked CCC activity, which can be used for visualisation of activity at the ligand-receptor and cell group level. 

The tool analyzes transcriptomic data alongside a massive database of more than 30,000 protein interaction pairs, capturing both intracellular signaling patterns and extracellular communication. This allows researchers to identify not just which cells are talking, but how they are doing it and which pathways are driving their responses. GraphComm successfully detected meaningful cell–cell interactions in datasets previously validated for communication studies, as well as in datasets involving genetic or chemical perturbations and spatially mapped cells.

By bridging RNA sequencing data with systems-level biology, GraphComm gives scientists a new way to decode the molecular “language” of cells, opening possibilities for understanding disease progression, tissue development, and therapeutic responses.

Availability – The code for GraphComm and notebooks used to analyse data presented in this study are provided in the GitHub repository https://github.com/bhklab/GraphComm