Bi-level graph learning unveils prognosis-relevant tumor microenvironment patterns in breast multiplexed digital pathology.

The tumor microenvironment (TME) is widely recognized for its central role in driving cancer progression and influencing prognostic outcomes. Increasing efforts have been dedicated to characterizing it, including its analysis with modern deep learning. However, identifying generalizable biomarkers has been limited by the uninterpretable nature of their predictions. We introduce a data-driven yet interpretable approach for identifying cellular patterns in the TME associated with patient prognoses. Our method relies on constructing a bi-level graph model: a cellular graph, which models the TME, and a population graph, capturing inter-patient similarities given their respective cellular graphs. We demonstrate our approach in breast cancer, showing that the identified patterns provide a risk-stratification system with new complementary information to standard clinical subtypes, and these results are validated in two independent cohorts. Our methodology could be applied to other cancer types more generally, providing insights into the spatial cellular patterns associated with patient outcomes.