Interpretable spatial cell learning enhances the characterization of patient tissue microenvironments with highly multiplexed imaging data.

ABSTRACT

Multiplexed imaging technologies enable highly resolved spatial characterization of cellular environments. However, exploiting these rich spatial cell datasets for biological insight is a considerable analytical challenge. In particular, effective approaches to define disease-specific microenvironments on the basis of clinical outcomes is a complex problem with immediate pathological value. Here we present InterSTELLAR, a geometric deep learning framework for multiplexed imaging data, to directly link tissue subtypes with corresponding cell communities that have clinical relevance. Using a publicly available breast cancer imaging mass cytometry dataset, InterSTELLAR allows simultaneous tissue type prediction and interested community detection, with improved performance over conventional methods. Downstream analyses demonstrate InterSTELLAR is able to capture specific pathological features from different clinical cancer subtypes. The method is able to reveal potential relationships between these regions and patient prognosis. InterSTELLAR represents an application of geometric deep learning with direct benefits for extracting enhanced microenvironment characterization for multiplexed imaging of patient samples.