Summary
A bioRxiv preprint describes HistoProfile, an AI model that estimates immune-marker measurements from routine cancer tissue slides. The authors report AUROCs of 0.78–0.89 for distinguishing biomarker-high from biomarker-low tumours.
A preprint posted on bioRxiv on 13 October 2026 describes HistoProfile, an AI model designed to estimate tumour immune markers from routine haematoxylin-and-eosin (H&E)-stained tissue slides. The authors report that it distinguished biomarker-high from biomarker-low tumours in held-out samples, with area under the receiver operating characteristic curve (AUROC) values of 0.78–0.89 across the biomarkers studied.
The model estimates intratumoral cell densities for CD8, FOXP3, PD-1 and PD-L1, as well as the PD-L1 tumour proportion score. The work is a preprint and has not yet been peer reviewed.
Estimating immune measurements from routine slides
H&E staining is commonly used to show tissue structure and cell features under a microscope. By contrast, multiplex immunofluorescence (mIF) uses fluorescently labelled antibodies to measure several proteins in a tissue sample. The authors say mIF assays can provide standardised immune-marker measurements but are costly and not routinely used in clinical settings.
HistoProfile was trained using a pan-cancer collection of paired H&E slides and mIF measurements. It is described as weakly supervised: the model learns from examples that pair tissue images with target measurements, rather than requiring every relevant cell in every image to be individually annotated. The abstract does not detail the training cohort size or the exact labelling process.
An AUROC measures how well a model separates two groups across possible classification thresholds: 0.5 indicates chance-level ranking, while 1 indicates perfect separation. The reported 0.78–0.89 range reflects classification of biomarker-high versus biomarker-low tumours; it is not a measure of exact agreement between estimated and measured cell densities.
Prognostic patterns in survival analyses
The authors report that HistoProfile’s estimates retained prognostic stratification in downstream survival analyses. They also found cancer-specific survival associations across The Cancer Genome Atlas (TCGA) cohorts, where mIF measurements were unavailable. These results suggest that image-based estimates could help researchers study immune-marker patterns across larger collections of archived tumour slides.
The reported evidence concerns biomarker estimation and associations with survival. It does not describe prospective clinical testing of HistoProfile or its use to guide individual treatment decisions. The abstract provides no cohort count or numerical survival effect estimates, so the scale and magnitude of those prognostic findings cannot be assessed from the posted summary.