Andrology-Open Access
Open Access

Abstract

The Future of Digital Biomarker Profiling

Marija Pizurica, Kathleen Marchal*

Several molecular biomarkers have been proposed to improve prediction of risk on metastasis or on relapse after treatment for cancer patients. However, tumor heterogeneity and high costs of sequencing have obviated the clinical implementation of these molecular markers. In addition, current biomarkers are derived from bulk profiles, which contain both tumor cells and cells from the tumor microenvironment. On the one hand, this results in a confounded biomarker signal influenced by the cellular composition of the sampled tissue. On the other hand, a bulk-derived biomarker does not consider the spatial organization of cells, which has shown important prognostic and predictive potential. Resolving these shortcomings would require expensive spatial profiling, which is infeasible in clinical settings. Spatially resolved digital profiling, obtained with deep learning models from whole slide images, presents promising potential for cost-efficient exploration of biomarkers at high resolution. Also, the predicted biomarkers from these models are ideal candidates for downstream lightweight, interpretable and efficient clinical outcome prediction. Here, we highlight important guidelines for developing such WSI proxy models, in terms of dataset size and label resolution trade-off, as well as inherent limitations of predicting molecular features on WSIs. We show the added value of molecular WSI proxy models for clinical outcome prediction as opposed to training WSI models directly for outcome, in terms of interpretability, dataset size and model efficiency.

Published Date: 2025-03-09; Received Date: 2024-04-16