New AI Model Decodes the Signals That Guide Immune Cells Into Tumors

One of the most stubborn mysteries in cancer immunology is deceptively simple to state: why do immune cells swarm into some tumors while ignoring others that appear, on paper, to be nearly identical? For years, researchers have been able to map where immune cells sit within a tumor biopsy, but the underlying logic of that map has remained largely hidden. Now, a team at The University of Texas MD Anderson Cancer Center and collaborating institutions in Texas has built an artificial intelligence model designed to answer precisely that question, and its early findings point to signals that attract T cells into tumors as well as physical barriers that appear to lock them out.

The tool, called Spatial Analysis of Cellular Engagement and Recruitment, or SPACER, was developed by a team co-led by Tao Wang, Ph.D., associate professor of Bioinformatics and Computational Biology, and Linghua Wang, M.D., Ph.D., professor of Genomic Medicine and executive director and head of the Center for Cellular Language Intelligence at MD Anderson. The work, published in Nature Biomedical Engineering, arrives at a moment when spatial biology is one of the fastest-moving fields in biomedical science, yet interpretation of the resulting data has lagged behind the ability to generate it.

SPACER is built on spatial transcriptomics, a molecular technique that measures gene activity within intact tissue, revealing not just which genes are switched on but where in the tissue architecture that activity occurs. Traditional single-cell sequencing destroys the spatial relationships between cells, dissolving a tumor into a suspension of individual units. Spatial transcriptomics preserves those relationships, capturing the neighborhood context in which every cell lives. That context, as it turns out, is exactly where the answers to immune recruitment may be hiding.

Scientists have long compared the trafficking of immune cells to a postal system, in which molecular addresses guide each cell to its correct destination. The analogy is useful but incomplete, because until now the postal codes themselves have been poorly understood. Chemokines and other signaling molecules clearly play a role, but the full set of local cues that determine whether a T cell enters, patrols, or bypasses a given patch of tissue has resisted systematic analysis. Wang and her colleagues designed SPACER to decode those signals directly from spatial transcriptomics data rather than relying on preconceived lists of candidate molecules.

What distinguishes SPACER from many other AI models in biology is its insistence on interpretability. A common criticism of deep learning approaches is that they function as black boxes: they may predict outcomes with impressive accuracy, but they cannot explain how they arrived at a conclusion, which limits their scientific value. SPACER was constructed around the actual organizational principles of tissue. Instead of treating each cell as an isolated data point, the model examines groups of neighboring cells and how they function together, making its predictions easier to understand and, crucially, allowing researchers to extract new biological insights rather than merely generating scores.

The performance numbers are striking. The team applied SPACER across 37 datasets representing several cancer types, and in benchmarking tasks the model proved four to 21 times more effective than existing approaches at identifying biologically meaningful signals linked to cell recruitment. That range matters, because the field has no shortage of computational methods that claim to read spatial data; what it lacks is methods whose outputs survive rigorous biological scrutiny. By validating its findings across dozens of independent datasets, the team has made a case that SPACER captures something real about how tissues direct immune traffic.

Among the most consequential findings is a matter of geography. Immune cells, the analysis suggests, respond primarily to signals in their immediate surroundings rather than to signals circulating elsewhere in the body. This local-action principle has immediate therapeutic implications: systemic treatments aimed at luring T cells into tumors may fail if the immediate microenvironment of the tumor does not cooperate. The tumor itself, in other words, is not just a passive target but an active participant in deciding who gets in.

SPACER provided concrete evidence for that participation. Tumors expressing genes involved in antigen presentation, the process by which cells display protein fragments on their surface for immune cells to inspect, were more likely to attract T cells. This makes intuitive biological sense, since T cells recognize their targets through those displayed fragments, but demonstrating that antigen presentation status correlates with physical immune infiltration across real tissue data strengthens the case that local interactions between tumors and immune cells determine whether immune cells enter at all. It also suggests a measurable biomarker: antigen presentation activity within a tumor could help predict whether immunotherapy has a chance of working in a given patient.

The second major finding may prove even more provocative. The researchers uncovered evidence that mucins, a family of large, sugar-coated proteins that many cancers overproduce, may help create physical barriers that keep T cells away from tumors. Across several melanoma and gastric cancer datasets, tumors with higher levels of mucins were associated with poorer responses to immune checkpoint inhibitors, the blockbuster drugs that have transformed treatment for many patients, and with poorer patient outcomes. If mucins do indeed form a structural exclusion zone around tumor cells, then breaking down that barrier could become a therapeutic strategy in its own right, potentially sensitizing otherwise resistant tumors to existing immunotherapies.

The versatility of the tool was demonstrated beyond oncology. When the team applied SPACER to data from preclinical models of immunotherapy-related myocarditis, a dangerous inflammation of the heart that can occur as a side effect of cancer immunotherapy, the model identified regions of the heart and molecular signals associated with T cell recruitment. Notably, it suggested that CD4-positive T cells may play a larger role in driving severe heart inflammation than previously recognized, a finding that could reshape how researchers think about this serious treatment complication and how they might monitor for it.

The researchers see applications extending well past cancer and cardiac inflammation. Because cell movement and positioning are central to autoimmune disorders, tissue repair and wound healing, SPACER could eventually serve as a diagnostic tool for cancer therapies and as a general-purpose instrument for understanding diseases in which cells end up in the wrong place, or fail to arrive at the right one. The work was funded in part by the National Institutes of Health, the Cancer Prevention and Research Institute of Texas, the U.S. Department of Defense, the Adrienne Helis Malvin Medical Research Foundation, and a Baylor College of Medicine Department of Surgery Seed Grant.

For Wang and her colleagues, the larger ambition goes beyond any single disease. Their goal, as they describe it, is to understand how cells communicate and how those signals shape treatment response and patient outcomes. SPACER represents a step toward decoding that cellular language, connecting the patterns visible in tissue images to the biology that explains them, and converting those insights into testable hypotheses in the laboratory. If the tool’s early performance holds up as it moves into broader use, the invisible postal codes that govern immune traffic may finally become readable, and with them, a new generation of targets for getting immune cells exactly where patients need them most.

Subject of Research: An AI model for analyzing spatial transcriptomics data to identify signals that recruit or exclude immune cells in tumors and other tissues

Article Title: AI tool reveals what attracts immune cells to tumors and what keeps them away

Article References: AI tool reveals what attracts immune cells to tumors and what keeps them away. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: SPACER, spatial transcriptomics, artificial intelligence, tumor microenvironment, T cells, immune cell recruitment, mucins, antigen presentation, immune checkpoint inhibitors, myocarditis, MD Anderson Cancer Center, Nature Biomedical Engineering