EPFL · Wetenschap

AI Model Maps Tumor Tissue to Improve Cancer Care

EPFL researchers have developed Virtual Tissues (VirTues), an AI foundation model designed to analyze tumor tissues across various cancer types. This model aims to enhance the identification and prediction of treatment responses and patient outcomes by interpreting complex spatial proteomics data.

Tumor complexity extends beyond cancer cells to include immune cells, blood vessels, and surrounding tissue, all of which influence cancer growth and treatment response. Spatial proteomics captures this organization at a molecular level, but the resulting data is challenging to interpret and compare across studies. VirTues addresses this by learning from spatial proteomics data from diverse studies and cancer types, enabling analysis from individual cells to whole tissue sections and patient outcomes.

The model's ability to integrate proteins measured across different studies, cancer types, and panels allows any new tissue sample to be analyzed within a unified framework. This flexibility enables the creation of a comprehensive atlas, facilitating consistent analysis of tissues and rapid confirmation of discoveries across patient groups. VirTues is trained broadly, similar to language foundation models, learning relationships among proteins, cells, and their spatial context to create a versatile representation applicable to various analyses.

VirTues offers a unified analysis tool for multiple research questions, enabling comparisons between patient groups, identification of distinguishing spatial patterns, and discovery of spatial biomarkers. The model's development involved assembling the largest open dataset of spatial proteomics measurements to date, with over 12,000 images from 5,000 patients across 31 cohorts. A new Transformer architecture was also developed to enable VirTues to learn from spatial proteomics data even when different protein sets are measured.

This work is part of the "Virtual Patient Labs" project, which aims to create a computational model of an individual's biology using their data. VirTues will provide the tissue component, to be integrated with pathology, genetic, and clinical data to predict disease progression and treatment efficacy. The goal is to place individual tissue samples within a broader biological and clinical context and to predict how tissues change under therapy, with the aim of improving clinical decisions.

AI-samenvatting op basis van de bron.