VirTues turns a tumor tissue section into a learned representation that can be compared across patients and studies. © 2026 SayoStudio
EPFL researchers have developed an AI model that analyzes tumor tissues across many cancer types. In the long term, it aims at helping identify and better predict treatment response and patient outcomes.
A tumor is more than its cancer cells. Immune cells, blood vessels and other surrounding tissue influence how a cancer grows and responds to treatment. Two tumors containing the same cell types can respond very differently to the same therapy depending on how those cells are arranged. Spatial proteomics captures this organization in molecular detail, but the resulting data, which contain many different variables, are difficult to interpret and compare across studies.
To address this challenge, Charlotte Bunne's Artificial Intelligence in Molecular Medicine group at EPFL's School of Computer and Communication Sciences and School of Life Sciences, has developed Virtual Tissues (VirTues), a foundation model for tissue biology. The model learns from spatial proteomics data from different studies and cancer types, and can be used to study biology across scales, from individual cells to whole tissue sections to patient outcomes. The results have been published in Nature.
Measuring dozens of proteins at once
"Which cells are present is only part of the picture," Bunne says. "We also need to know where they are and how they interact." Answering that across many patients is as much a computational problem as a biological one.
"In oncology, the amount of data generated from analyzing tissue has grown exponentially in recent years," says Andreas Wicki, oncologist at the University of Zurich and University Hospital Zurich, who was also involved in this work. "Computational modeling is the key asset for making it actionable for patients in a clinical setting."
Because VirTues can incorporate proteins measured across different studies, cancer types, and panels, any new tissue sample can be compared and analyzed within the same framework. Each new study builds on and expands the model's learned representation rather than requiring a new model from scratch.
"This flexibility is one of the greatest strengths of VirTues," says Johann Wenckstern, a PhD candidate in Bunne's group and first author of the study. "It allows us to integrate datasets into a single atlas, analyze tissues from different studies in the same way, and quickly confirm discoveries across different groups of patients."
Tissue data can answer many different biological and clinical questions, but most computational models are designed for only one at a time. VirTues on the other hand follows the logic of foundation models used for language: it is trained broadly rather than for a single predefined task, and instead of learning relationships among words, it learns relationships among proteins, cells, and their spatial context, creating a shared representation that can be applied to different analyses.
Two central advances
"As an analysis tool, VirTues handles questions a laboratory would otherwise need separate methods to answer," Bunne says. "Researchers can use it to compare groups of patients, determine which spatial patterns distinguish them, and discover spatial biomarkers associated with disease progression and treatment response. The question changes, but the underlying model does not."
With VirTues, the team made two central advances. First, they assembled the largest open dataset of spatial proteomics measurements to date, comprising more than 12,000 images from over 5,000 patients across 31 clinical cohorts. Second, they developed a new Transformer architecture, a type of AI model designed to learn complex patterns in large datasets, that enables VirTues to learn from spatial proteomics data across studies, even when different sets of proteins are measured.
"The possibilities of such a foundation model for precision oncology are remarkable, spanning from high-precision understanding of cell types and states of a tumor, to opening the way for improved treatment selection," says Olivier Michielin, head of Precision Oncology at the Geneva University Hospital who was also involved in the study. "Incorporation of VirTues into our local and national precision oncology tumor boards should be a natural next step."
Building a virtual patient
This work is part of a larger five-year project "Virtual Patient Labs: AI-Driven Simulation and Diagnostics for Precision Oncology". The project, for which Bunne was awarded the 2026 Lopez-Loreta prize, will create a working computational model of one person's biology, built from their own data and detailed enough for clinicians to use for predicting how a disease will progress and which treatment is likely to work. VirTues provides the tissue component, which will be combined with routine pathology, genetic information, clinical data, and other patient records.
"An atlas is only clinically useful if a new patient can be placed into it. A sample taken in Geneva can be read against thousands of samples collected elsewhere. This places the individual tissue sample within a much broader biological and clinical context," says Bunne. "But describing a patient's state is only the first step. The harder step is moving from describing a tissue to predicting how it changes under a given therapy. Testing where that can improve clinical decisions is what the coming years are for."
Other contributors
University of Zurich
University Hospital Zurich
University of Geneva
Geneva University Hospital
