First-of-its-Kind AI Tool Reveals the ‘Social Network’ of Cancer Cells

By Deborah Borfitz

May 7, 2025 | Researchers in Europe have developed a unique deep learning tool for analyzing spatial transcriptomics data to help decipher communication happening cell to cell as well as at the community level within tissues. It’s the first artificial intelligence (AI) method capable of measuring and interpreting a range of data from a cell's “social network,” including which cells are talking to each other and how they’re developing into different cellular niches, according to Sebastian Birk, a Ph.D. student at the Institute of AI for Health, Helmholtz Munich. 

The tool—called, appropriately enough, NicheCompass—is a product of the computational expertise of Birk and his colleagues at Helmholtz Munich as well as at the University of Würzburg and Wellcome Sanger Institute (UK). It made its debut in a paper that was recently published detailing how it can uncover tissue changes across breast and lung cancer patients (Nature Genetics, DOI: 10.1038/s41588-025-02120-6). 

Spatial transcriptomics data refers to information about gene expression within a tissue and where cells and their transcripts are situated, explains Birk. This enables researchers to map gene expression patterns in relation to tissue structure, revealing insights about cellular organization, interactions, and biological processes, including disease mechanisms. 

NicheCompass quantitatively characterizes niches (cellular communities) based on communication pathways and was shown to consistently outperform alternative methods. Notably, in one hour, it could identify how certain people may respond differently to treatment. 

The underlying source code is publicly available, and researchers elsewhere are already using the tool to analyze spatial data in diseased versus healthy tissues. By learning which spatial gene networks are prevalent only in diseased tissue, they can better tease out the underlying biology, which in turn could identify new targets for drug development and biomarkers to advance diagnostics.  

Since a preprint about the tool was published a year ago, it has generated considerable interest among external researchers, Birk says. “Many works are in progress at the moment that use NicheCompass, and I assume they will be published soon.”  

Cellular communication is disrupted by many types of disease, but cancers currently represent the primary area of interest, says Birk, especially how cancer cells interact with the immune system. This was also the focus of the mapping application to human cancers in the latest study. In the case of lung cancer, researchers showed five patients could be stratified by tumor microenvironment immune cell infiltration patterns that could be linked to different outcomes.  

As a computer scientist, Birk is already working on a new but similar spatial transcriptomics method focused on large patient cohorts to make predictions about clinical outcomes. The researchers constructed a whole mouse brain spatial atlas comprising 8.4 million cells in the latest paper to highlight NicheCompass’ scalability. 

Scoring ‘Communication Potential’

NicheCompass is given spatial transcriptomics data and learns the source- and target-specific “communication potential scores” quantifying communication strengths between cell pairs and aggregating them at the niche and cell type levels, Birk says. For the study, other tools were used to visualize the results indicating which cells are communicating and what spatial gene network they are using to do so. 

Transcriptomic data being probed by NicheCompass, particularly messenger RNA, is a proxy for proteins on the surface of cells, he says, and the correlation between the two has been suggested by other studies and provides “readouts for more genes.”  

Information from various cell and tissue atlases, created by single-cell and spatial genomic technologies, are the inputs used to know “who is in the neighborhood,” so to speak, what street they live on, and what influences how often they get together and [how they] behave when they do, he offers as an analogy. That data gets used by NicheCompass to learn how different cells communicate through their networks and align with similar networks of cells to create spatial niches with a common look, much like residential neighborhoods following different cultural norms or sets of neighborhood homeowners association rules. 

Collaborators on the project were brought together by the global Human Cell Atlas Initiative, which aims to create a comprehensive map of everything known about all cells in the human body. The role of the computational community is to develop tools that can help piece together this biological puzzle, which is where NicheCompass comes in, says Birk.  

The tool is currently being used mostly for basic science, but Birk says he is hopeful that people will start using NicheCompass within the next year to inform clinical drug trials. Potentially, physicians could also input data on patients to receive in-depth information to help guide their clinical decision-making—the caveat being that they would need to have some computational knowledge, he adds, “at least for now.”  

Birk says he welcomes feedback from the user community, especially the usability aspect of NicheCompass. It has an installation interface, but not yet a web interface so software is required to use the tool. He adds that he is also happy to put more effort in that direction for people with less computational background, as well as connect NicheCompass to a visualization platform so users can upload their data from a web server and immediately visualize it.