Researchers at ETH Lausanne, in collaboration with partners including the University Hospital of Zurich (USZ) and the University of Zurich (UZH), have developed an AI model capable of analyzing complex proteome data in a targeted manner. In the long term, it is expected to help predict treatment approaches and disease progression more accurately and to advance precision oncology.
A tumor is not made up solely of cancer cells. Immune cells, blood vessels, and other components of the surrounding tissue also influence how a tumor grows and how it responds to treatment. Even tumors with a similar cellular composition can therefore respond very differently to the same treatment. The key factors are which cells are present in the tumor, where they are located, and how they interact with one another. Modern analytical methods such as spatial proteomics shed light on these complex relationships. However, the large amounts of data generated in this process are difficult to analyze, and until now, it has been possible to compare them across different studies only to a limited extent.
To address this challenge, a research group at ETH Lausanne, in collaboration with researchers from Geneva, the USZ, and UZH, has developed a foundation model capable of reconstructing three-dimensional tissues from spatial proteome data. The “Virtual Tissues” (VirTues) model can compare spatial proteome data across different studies and tumor types. It can be used to investigate biological relationships between cells, tissues, therapies, and patient treatment outcomes. The results were published in *Nature* on August 5, 2026.
Measuring Hundreds of Proteins Simultaneously
“The types of cells present in a tumor are only part of the bigger picture,” says Charlotte Bunne, head of the research group at ETH Lausanne. “We also need to know where they are located and how they interact with one another.” Answering this question across a large number of patients is as much a mathematical problem as it is a biological one.
“In oncology, the sheer volume of data generated during tissue analysis has grown exponentially in recent years,” says Andreas Wicki, deputy clinical director at the Department of Medical Oncology and Hematology at the USZ. “Computer-aided modeling is the key to making data useful for patients in a clinical setting.”
VirTues can incorporate proteins that have been measured in various studies, across different types of cancer, and using different panels. This allows each new tissue sample to be compared and analyzed within the same framework. Each new study builds on what has already been learned and continuously expands our knowledge.
“This flexibility is one of VirTues’ greatest strengths,” says Johann Wenckstern, a doctoral student in Bunne’s group and the study’s first author. “It allows us to integrate datasets into a single model, analyze tissue from different studies in the same way, and quickly interpret findings across different patient groups.”
Tissue data can provide answers to many different biological and clinical questions, but most existing computer-based models are designed to address only a single question at a time. VirTues, on the other hand, follows the logic of foundation models as they are already used in natural language processing: It is trained on a wide range of tasks rather than for a single, predefined task. Instead of learning relationships between words, the model learns relationships between proteins, cells, and their spatial context, thereby creating a unified representation that can be applied to various analyses.
Two Key Advances
“As an analytical tool, VirTues addresses questions that would otherwise require a laboratory to use separate methods to answer,” says Charlotte Bunne. “Researchers can use it to distinguish between different tumor patterns and identify biomarkers associated with disease progression and response to treatment.”
VirTues has made two key advances. First, it compiles the largest publicly available dataset of spatial proteomics measurements to date, comprising more than 12,000 images from over 5,000 patients. Second, it develops a model designed to learn and interpret complex patterns in large datasets.
“The potential of such a foundation model for precision oncology is remarkable. They allow us to evaluate many biomarkers simultaneously without having to make biologically based hierarchical assumptions about the relative importance of the markers. “This is an important step toward developing structured therapeutic prediction models,” says Andreas Wicki. “In the medium term, the further development of VirTues and other models into a medical device and their integration into our local and national tumor boards for precision oncology is a logical next step.”
Link to the publication
