Researchers at the Szeged Biological Research Centre's Momentum (Lendület) Microscope Image Analysis and Machine Learning Research Group, led by Horváth Péter, have developed methods that map tumor heterogeneity at single-cell resolution in collaboration with Swedish and Swiss partners. The team published two papers recently describing how AI-based image analysis combined with spatial proteomics and transcriptomics can reveal genetic and protein-level differences between tumor cell populations.
Conventional molecular assays typically analyze bulk tissue samples, which can mask biologically distinct behavior of different tumor regions. The Szeged group aimed to create a molecular map that not only shows how a tumor looks, but also which genes and proteins are active in its individual cell populations.
International collaboration and technological platform
Key participants in the collaboration include Holger Moch, a molecular pathologist at University Hospital Zurich, and Markó-Varga György, research professor at Lund University. The technological backbone was provided by the Szeged team’s automated single-cell research centre: a worldwide-unique system that uses a robotic laser to isolate AI-identified cells from the sample with high precision and without human intervention, operating continuously if needed.
The approach is based on Deep Visual Proteomics (DVP). First, artificial intelligence analyzes histological images to detect and mark relevant cell types or groups. These marked cells are then precisely excised with a laser beam. Proteomic analysis of the extracted cells reveals which proteins are present, and protein composition directly reflects cellular functions such as growth, adaptation or emergence of resistance.
Single-cell genetic and proteomic parallels
In one of the published studies, the researchers investigated tumor clones at single-cell resolution, examining both genetic and protein composition. They focused on cell groups that, based on microscopic features, appeared potentially more aggressive. Those AI-selected groups underwent combined proteomic and transcriptomic analysis, enabling the team to determine which genes are active and how that gene activity translates into protein-level function.
Using both data types in parallel allowed a more comprehensive characterization of the tumor cell populations: gene activity and proteomics together can reveal a cell's likely biological aggressiveness, metabolism and interactions with the immune system.
Tracing the source of metastases
The group also presented a clinical case: samples from a young patient with recurrent metastatic melanoma were analyzed, including the primary tumor and later lung and brain metastases. AI-based digital pathology identified two distinct tumor cell populations in the primary lesion. Cells in the later metastases resembled primarily one of these early populations, and proteomic profiling confirmed that the molecular signature of that early group was closest to the metastases.
This finding suggests that molecular traces of future metastatic spread may already be present in the primary tumor, and certain cell populations at diagnosis can carry signatures associated with later metastatic outgrowth. Migh Ede, a participant in the research, noted that the method demonstrated not only morphological differences between tumor regions but also corresponding differences at the protein level, bringing researchers closer to identifying the cell populations responsible for metastasis.
Clinical implications and limitations
The Szeged team emphasizes that AI does not replace the pathologist in this workflow; rather, it provides much finer resolution for tumor analysis. While traditional pathology classifies larger tissue regions, AI can analyze samples cell by cell, revealing which tumor cell types are present in which patterns and ratios. The more precisely clinicians can identify the cells driving disease progression, the better they can tailor therapies to target the most dangerous tumor compartments.
The researchers also acknowledge that integrating these findings into routine clinical practice and therapeutic decision-making will take time. Nevertheless, the studies represent significant progress toward more accurate mapping of tumor biology, earlier prediction of critical risks, and more personalized treatment choices.
Tags: artificial intelligence, machine learning, Szeged, medicine, genetics, tumor, oncology, Deep Visual Proteomics, single-cell analysis
