Cancer is often described as a single disease, but under the microscope a tumor is rarely uniform. It is an evolving ecosystem made up of diverse cell populations, each carrying different genetic programs, protein networks, metabolic states, and capacities to interact with the immune system. Identifying which of these cells drive aggressive growth, treatment resistance, or metastasis has been one of the central challenges of precision oncology. A research group at the HUN-REN Biological Research Centre in Szeged, Hungary, is developing artificial intelligence-based technologies designed to address that challenge by linking the visual appearance of individual cells to their molecular behavior.
The Momentum Microscopic Image Analysis and Machine Learning Research Group, led by Péter Horváth, has combined digital pathology, machine learning, spatial omics, and automated cell isolation to examine tumors at single-cell resolution. Two recent studies report advances in this strategy. One describes a method for connecting the morphology of tumor cells with both their genetic activity and protein profiles. The other investigates how distinct tumor cell populations in a primary melanoma may be related to later metastatic lesions. Together, the findings suggest that microscopic images can serve as a molecular guide to the most consequential regions of a tumor.
Conventional molecular testing commonly analyzes tissue in bulk. A biopsy or surgical specimen is homogenized, and the resulting molecular measurements represent an average across millions of cells. That approach can identify important features of a tumor, but it may conceal the differences between neighboring cell populations. A small group of highly invasive cells can be diluted by less aggressive cells, while spatial relationships between tumor cells and surrounding tissue are lost. The Szeged team’s approach instead preserves the tissue map and uses artificial intelligence to identify specific cells or cell communities before their molecular properties are measured.
A central platform in this work is Deep Visual Proteomics, or DVP. The process begins with high-resolution histological imaging, in which tissue architecture and cellular morphology are recorded. Machine-learning algorithms analyze the images and select cells or populations according to visual characteristics associated with particular biological states. A focused laser then cuts out the selected material with high spatial precision. The isolated cells can subsequently be processed for proteomic analysis, which measures the proteins they contain. Because proteins are the active molecules that regulate cellular functions, their abundance and combinations can reveal how a cancer cell is growing, adapting to stress, communicating with neighboring cells, or resisting therapy.
The researchers extended this strategy by examining AI-selected tumor populations through both proteomics and transcriptomics. Transcriptomics measures RNA molecules and indicates which genes are actively being expressed, while proteomics measures the proteins produced as a result of cellular activity. These two layers provide complementary information: RNA can reveal the instructions being used by a cell, whereas proteins offer a closer view of the functional machinery operating inside it. Studying both may expose cases in which gene activity and protein behavior diverge, as well as biological programs that would remain invisible through morphology or a single molecular assay alone.
The method was applied to clear cell renal cell carcinoma, a kidney cancer known for substantial biological and clinical heterogeneity. The study, published in EMBO Molecular Medicine, is titled “Deep visual multi-omics profiling links morphology and molecular programs in clear cell renal cell carcinoma.” By combining visual analysis with single-cell-level genetic and protein measurements, the researchers sought to determine whether tumor regions that look different also follow distinct molecular programs. The result is a form of multi-omics mapping in which the image does not merely document the tissue; it helps direct the molecular investigation toward the cells most likely to explain disease behavior.
The same technological framework was also used to explore the possible origins of metastasis in a young patient with recurrent metastatic melanoma. Samples from the original tumor and from later lung and brain metastases were examined using AI-guided digital pathology and spatially resolved proteomics. The algorithms identified two visibly distinct tumor cell populations in the primary melanoma. When the molecular profiles were compared, cells in the later metastatic lesions most closely resembled one of those original populations. The observation suggests that a population with features associated with later spread may already have been present in the primary tumor, even before metastases became clinically apparent.
This result does not prove that the identified cells alone caused the metastases, and it does not yet provide a clinical test for predicting which tumors will spread. It does, however, illustrate the type of question that spatial single-cell technologies can address. Instead of asking only whether a tumor contains a particular mutation or protein, researchers can investigate where that feature occurs, which cells carry it, how frequently they appear, and whether the same cellular program is found in distant lesions. Such information could eventually help define the subpopulations that deserve closer monitoring or that may require treatment strategies aimed at more than the tumor’s dominant cell type.
The Szeged group developed its digital pathology and spatial omics work with collaborators in Sweden and Switzerland, including molecular pathologist Holger Moch of University Hospital Zurich and research professor György Marko-Varga of Lund University. Its automated single-cell research center is designed to isolate AI-selected cells without continuous manual intervention, potentially allowing experiments to run with consistent precision over extended periods. For cancer biology, that automation is important because large numbers of individually selected cells may be required to capture the diversity within a tumor and distinguish reproducible patterns from biological noise.
The broader significance of these studies lies in a shift in how tumors are understood. Artificial intelligence is not replacing pathologists or independently diagnosing patients in this approach. Rather, it is acting as a high-resolution instrument that connects tissue appearance with molecular function. By revealing the cellular geography of a tumor, the technology may help researchers understand why some regions become invasive, why others evade treatment, and how metastatic potential emerges. The findings are not an immediate new therapy, but they point toward a future in which cancer samples are analyzed as dynamic cellular landscapes, enabling more precise risk assessment and, ultimately, treatment decisions directed at the most dangerous populations within a tumor.
Subject of Research: AI-guided digital pathology, Deep Visual Proteomics, spatial omics, single-cell tumor analysis, cancer heterogeneity, and metastasis.
Article Title: Deep visual multi-omics profiling links morphology and molecular programs in clear cell renal cell carcinoma
Web References: https://www.brc.hu/en/research/institute-of-biochemistry/synthetic-and-systems-biology-unit/lenduelet-laboratory-of-microscopic-image-analysis-and-machine-learning; https://doi.org/10.1038/s44321-026-00484-8; https://doi.org/10.1038/s41698-026-01569-w
References: EMBO Molecular Medicine, DOI: 10.1038/s44321-026-00484-8; npj Precision Oncology, DOI: 10.1038/s41698-026-01569-w
Image Credits: András Kriston
Keywords: artificial intelligence, digital pathology, Deep Visual Proteomics, spatial omics, single-cell analysis, cancer research, clear cell renal cell carcinoma, melanoma, metastasis, precision oncology
Tags: AI-driven cancer biomarker discoveryautomated cell isolation in cancer studiescancer tumor heterogeneitydigital pathology and machine learningidentifying treatment-resistant tumor cellslinking tumor cell morphology to genetic activitymelanoma metastasis and tumor cell diversitymetastasis prediction using artificial intelligencesingle-cell resolution tumor analysisspatial omics in cancer researchtumor cell populationstumor microenvironment and immune interactions
