Researchers in Switzerland have developed an AI foundation model that can analyse the complex spatial organisation of tumour tissue across different cancer types and datasets. Called Virtual Tissues, or VirTues, the model could help researchers identify biomarkers, predict treatment response and eventually support more personalised cancer treatment.
A tumour consists of far more than cancer cells. Immune cells, blood vessels and surrounding tissue all influence how cancer develops and responds to treatment. Even tumours containing similar cell types can react differently to the same therapy depending on how those cells are organised and interact.
Spatial proteomics can capture this organisation by measuring dozens or even hundreds of proteins while preserving their location within tissue. However, studies often use different technologies, protocols and combinations of protein markers, making their datasets difficult to compare or combine.
Researchers at EPFL, ETH Zurich, the University of Geneva, Geneva University Hospitals, the University of Zurich and University Hospital Zurich developed VirTues to overcome this problem. Their results have been published in Nature.
Shared tissue model
VirTues works as a foundation model: rather than being developed for one specific analytical task, it learns broadly from large amounts of spatial proteomics data. Much as language models learn relationships between words, VirTues learns relationships between proteins, cells and their spatial context.
Which cells are present is only part of the picture, according to Charlotte Bunne, head of the Artificial Intelligence in Molecular Medicine group at EPFL, states that knowing which cells are present, is only part of the picture. Knowing where they are and how they interact is also crucial knowledge.
The researchers assembled a large collection of spatial proteomics data. Their core dataset comprised 15 cohorts involving 3,102 patients and 146 different markers. An extended dataset covered 32 cohorts, more than 5,100 patients and 239 markers across several imaging technologies.
A new Transformer architecture enables VirTues to analyse datasets even when different combinations of proteins have been measured. New samples can therefore be placed within the same computational representation without developing a separate model for every study.
The model can perform tasks including cell segmentation and classification, reconstruction of missing protein markers, identification of spatial biomarkers and patient stratification. It can also analyse previously unseen datasets without being specifically retrained for them.
Predicting treatment response
The researchers demonstrated the potential clinical relevance in several cancer datasets. In triple-negative breast cancer (TNBC), VirTues identified tissue features associated with response to chemotherapy combined with anti-PD-L1 immunotherapy. During treatment, it outperformed several existing computational approaches in predicting pathological complete response.
The model also identified spatial signatures that could stratify disease-free survival in an independent TNBC cohort. More broadly, VirTues performed strongly in tasks including lung cancer subtyping and grading and determining oestrogen receptor status in breast cancer.
Olivier Michielin, head of precision oncology at Geneva University Hospitals, sees potential for incorporating VirTues into precision oncology tumour boards. Further research will be needed, however, to determine whether its predictions can improve actual clinical decisions.
AI across cancer care
VirTues reflects a broader expansion of AI across cancer research, diagnosis and treatment. Researchers at the University of Navarra, for example, developed the open-source RNACOREX tool to analyse complex genetic regulatory networks and identify molecular patterns associated with cancer survival.
At the University of Cambridge, researchers developed SMMILe, an AI system that analyses digital pathology slides and maps different tumour regions and subtypes. The approach could ultimately help clinicians select treatments based on differences within individual tumours. Earlier research has also shown how AI can support more precise radiotherapy, including automated organ contouring, three-dimensional monitoring of tumour response and identification of treatment-resistant tumour regions.
VirTues could eventually become part of an even broader virtual representation of individual patients. Bunne is leading a five-year project that aims to combine tissue information with pathology, genetics, clinical data and other patient records.
Describing a patient's state is only the first step, according to Bunne. The harder, second step is to move from describing a tissue to predicting how it will change under any given therapy.
