VirTues compares tissue images across different protein panels

VirTues compares tissue images across different protein panels

On August 5, Nature published a paper on VirTues, a spatial proteomics model that analyzes proteins in a tissue section together with their locations. The model was trained on 32 clinical cohorts, data from more than 5 100 patients, and 239 proteins. The authors tested it on biopsies from patients with triple-negative breast cancer.

Tumors contain cancer cells, immune cells, and connective tissue cells in close proximity. Spatial proteomics turns a tissue section into a map that shows which proteins are present in each cell and which cells are adjacent to one another. Two samples may contain the same cell types but differ in how those cells are arranged. For example, immune T-cells may approach the tumor in one sample but remain at its periphery in another.

Laboratories measure different protein panels in tissue. One laboratory may measure 40 markers, while another measures 80. Their antibodies, instruments, and sample preparation methods may also differ. A conventional model requires a predefined list of measured proteins, so it cannot directly compare data from different cohorts when the panel changes.

“Each study uses its own set of protein markers and its own protocol. Most methods are designed for a single cohort, which makes it difficult to transfer knowledge between studies and identify robust biomarkers.”

VirTues links the signal from each measured protein to that protein’s amino acid sequence and its location in the tissue section. During training, the model masks part of the signal and reconstructs it from nearby proteins and the organization of the surrounding region. This process teaches the model to account simultaneously for an individual cell, its local environment, and the entire sample. The resulting representation can then be applied to an image generated with a different protein panel.

In a stringent test, the authors excluded the entire target cohort from training and then asked VirTues to segment the cells in its images. A cell was counted as correctly identified when the predicted boundary overlapped sufficiently with the annotated boundary. By this criterion, the system outperformed three specialized programs on eight of nine datasets.

VirTues identified four spatial signatures in pretreatment biopsies from 111 patients with triple-negative breast cancer. These signatures represented combinations of cell types and their spatial relationships. When the authors repeatedly tested the model on different subsets of these data, the combined four signatures distinguished patients who responded to chemotherapy plus PD-L1 immunotherapy from those who did not. AUROC, a measure of how well a model distinguishes two groups, was 0.823. A value of 1 on this scale represents perfect separation. The same signatures were then applied to another cohort. In the low-risk group, 3 of 33 people experienced a disease-free survival event, compared with 21 of 45 in the high-risk group.

The publicly available VirTues code and weights allow other groups to use the model to analyze their own data.