Abstract Background Papillary renal cell carcinoma (pRCC) is a prevalent but understudied kidney cancer pathology, constituting 15-20% of cases. While many patients achieve curative therapy using nephrectomy, others go on to develop progressive, metastatic disease with poor overall prognosis. In order to improve the current standards of care for pRCC patients, it is critical to identify new methods to identify high-risk patients likely to require more intense and risk-modulated follow-up. Our goals in this project are two-fold: (1) to identify prognostic patient clusters based on spatial architecture of tumor tissue, and (2) to identify a minimal subset of markers to accurately predict the defined clusters for clinical translation. Methods A tissue microarray of 100 pRCC patients was assayed using multiplex immunofluorescence with a 31-antibody panel with immune- and cancer-related proteins. We have developed novel algorithms to identify spatial neighborhoods and features. These features are then used to identify unsupervised clusters of patients. This new approach to classifying patients allows for a more specific and comprehensive description of each patient tumor’s spatial composition and association with diverse clinical outcomes. Next, we developed attention-based deep learning models capable of predicting these spatially-defined clusters directly from the underlying immunofluorescence images and subsequently profiled these for interpretability by retraining based on individual channels. We assessed each model iteratively on a validation set and a held-out test set for final model selection. We focussed on developing an accurate model with the minimal number of channels to enable development of a cost-efficient biomarker assay for the clinic. Results We performed a systematic analysis of the cellular and spatial phenotypes for each patient in our cohort and identified six (6) major spatial clusters. These clusters are based on the spatial neighborhood composition of the tumors (Fig 1A) and describe the composite effects of cell-cell interactions within a patient tumor. We identified one cluster in particular (cluster #6, mainly associated with spatial interaction of CD163+ macrophages) with a significantly worse prognosis than the other clusters. Using our advanced deep learning models, we have been able to achieve excellent validation accuracy in predicting a variety of clinical phenotypes and spatial features, including patient clusters and histological grade. We wished to then reduce the set of required features and began our efforts in predicting tumor vs. normal based on individual immunofluorescence channels and identified 6 channels with >75% individual predictive accuracy (Fig. 1B). Compositing all 6 markers further improves our accuracy to 90%. This success paves the way for exciting future prospects as we work to extend these efforts to identify a minimal clinically-applicable subset of protein markers. Conclusions The combination of computational spatial proteomics and deep learning models has the potential to identify new biomarkers specific to the spatial organization of patient tumors. We have identified a particular patient cluster with multiple associations with CD163+ macrophages demonstrating poor prognosis; selecting these patients may allow for tailored and more informed patient care. We are also working to enhance our deep learning models to circumvent the need for stepwise spatial analysis and, instead, allow for direct-from-image assessment of patients for particular spatial features indicative of various prognostic indicators. Our identified set of 6 markers serves as a basis for further refinements as we work to identify a minimal set of relevant markers able to predict various clinical outcomes from immunofluorescence imaging, offering a promising new avenue in the field of cancer pathology and biomarker development. DOD CDMRP Funding: yes