In a laboratory in Los Angeles, researchers at Cedars-Sinai have built a tool that could transform the speed and cost of cancer treatment — by teaching an artificial intelligence system to read tumor tissue the way pathologists do, but far faster. The tool, called Path2Space, predicts gene expression patterns across a biopsy slide in minutes rather than weeks, and costs a fraction of what conventional testing requires.
Gene expression profiling tells doctors which genes are active in a tumor, a crucial piece of information for tailoring treatment to each patient's cancer. But tumors are not uniform — different regions express different genes — and mapping this "spatial" variation has always been expensive and slow. Traditional spatial gene expression profiling takes several weeks and costs thousands of dollars per sample, which has meant that personalized cancer medicine remains out of reach for many patients. Path2Space changes that equation by analyzing digital images of biopsy slides to predict spatial gene expression across nearly 5,000 genes at multiple points within a single tumor, all in minutes.
The Cedars-Sinai team, led by Dr. Eytan Ruppin, trained Path2Space on breast cancer patient data where both biopsy images and measured spatial sequencing were available, then validated it against three additional patient datasets. The results matched measured expression patterns well across all groups. Eldad Shulman, co-first author of the study published in Cell, noted the practical breakthrough: "Before we developed Path2Space, the largest cohort we could find to study the spatial organization of the tumor environment was about 30 patients. With this tool, we can study slides from thousands of patients."
That scalability opens entirely new avenues for discovery. The researchers found specific spatial patterns of gene activity within tumors that predict how patients respond to treatment — biomarkers that could guide clinical decisions and identify high-risk patients. Yet these spatial biomarkers have been nearly impossible to identify at scale, simply because the cost of generating that data has been prohibitive. Path2Space democratizes access to this information by leveraging images that already exist in pathology archives.
The tool currently analyzes groups of 10 to 20 cells together, but the team is working toward single-cell precision. The research is also expanding beyond breast cancer — collaborators are finalizing a study applying Path2Space to head and neck cancer, and the framework can be adapted to other tumor types once trained on the appropriate datasets.
Dr. Ruppin emphasizes that the path to clinical impact requires rigor. "It represents an exciting development in a growing field and has to be tested carefully," he said. "But we are hopeful that it could make an impactful contribution to science and to patient care." The next step is clinical trials, where Path2Space will be tested in real-world settings. If those trials succeed, this technology could move personalized cancer treatment from a privilege of the few to a possibility for many — simply by teaching machines to see what pathologists have always known was there.