AI for histology-based spatial biomarker discovery

Understanding the tumor microenvironment (TME) and its spatial architecture in the context of disease is critical for improving patient outcomes. However, its study has been limited by the lack of standardized quantification methods and tools. At Owkin, we are bridging this gap by developing reproducible, AI-based biomarkers to accurately characterize patients’ TMEs.

In this presentation, we will detail our methodologies to provide a characterization of individual patient's TMEs at cellular level resolution. Our approach incorporates active learning techniques and collaboration between data scientists and pathologists to build high-quality datasets and develop robust cell-level detection and classification models.

Our models are then used to create spatially resolved characterizations of patient tumors. For instance, we show that the density of lymphocytes within the tumor core, along with metrics quantifying their spatial organization, serve as significant prognostic factors in colorectal and breast cancer.

These spatial descriptors are further integrated into our drug discovery engines to refine and optimize the selection of patient populations for therapeutic development.