69 Spatial Proteomics Profiling Identifies Prognostic Biomarkers in Papillary Renal Cell Carcinoma

Abstract

Abstract Background Papillary renal cell carcinoma (pRCC) is the second most common histology of kidney cancer diagnosed in adults. However, the biologic drivers of recurrence after nephrectomy remain poorly defined, and no molecular biomarkers are routinely used in clinical practice to estimate prognosis. The primary prognostic tools used by physicians caring for patients with pRCC are clinical and histopathologic features: tumor size, lymph node status, presence of metastases, and tumor grade. The field has been unable to harness the power of cutting-edge molecular biology to inform the care of patients with pRCC tumors. Methods To explore the biology of pRCC, we assembled a tissue microarray (TMA) comprising pRCC tumors and adjacent normal kidney tissue from 102 patients treated at Vanderbilt University Medical Center. The TMA was stained with a 31-plex immunofluorescence panel (using a core set of immune and microenvironment markers) and imaged, providing single-cell resolution. Novel computational methods were developed to perform spatial analysis, identifying 22 distinct cellular neighborhoods (spatial clusters). Machine learning modeling was performed to develop predictive models which were assessed using cross-validation. Permutation models were developed to assess the empirical “null” distribution for the cross-validation accuracy. Different sets of features (cell type proportions, marker intensities and spatial clusters) were compared for predictive performance using the machine learning framework. Validation of the important features from the machine learning model was performed on the TCGA KIRP dataset. Results We systematically analyzed this dataset to identify specific markers, cell types, and spatial features associated with clinical outcome. We first identified CD8+ cytotoxic T lymphocytes (CTLs) as a population whose high abundance was strongly associated with poor clinical outcomes (p = 0.005 for overall survival; p < 0.05 for recurrence). Several markers comprising the overall CTL cell type annotation were also associated with these phenotypes, including CD3 and CD8. Additionally, CD45RO, a marker of memory T cells, was strongly associated with poor survival (p < 0.001). We further profiled these CTLs for therapeutically relevant markers and found that high PD-1 expression on these cell was even more strongly associated with recurrence (p = 0.001). We then calculated spatial clusters from cell-by-cell spatial connections and identified 22 clusters that defined the spatial associations between cells. Multivariate machine learning modeling showed that spatial clusters were more strongly predictive of clinical outcomes compared to other feature types. (Figure 1B). Figure 1A shows a heatmap of the spatial clusters in the ML model along the rows and patients along the columns (left plot). The right plot shows the cell types or markers associated with the spatial clusters. One cluster, cluster 2, included smooth muscle actin (SMA)+ cells and was the feature most strongly associated with recurrence (p < 0.005), highlighting a potential role for cancer-associated fibroblasts in aggressive pRCC tumors (example TMA core images shown in Figure 1C. A separate cluster, cluster 15, was associated with PD-1+ cells and a higher risk of recurrence (p < 0.005), providing addition support to our previous analysis of PD-1+ CTLs. By contrast, cluster 3 is associated with pan-cytokeratin+ cells and indicates positive prognosis, with lower likelihoods of recurrence (p < 0.01; example TMA core images shown in Figure 1D). Validation of the top spatial clusters, cluster 2 (SMA+) and cluster 3 (pan-cytokeratin+), in the TCGA KIRP tumors provides a basis for further biomarker development to aid clinical prediction of patients at high risk for recurrence. Conclusions We have identified specific cellular neighborhoods predictive of recurrence after radical nephrectomy and validated key features in an external RNA-seq dataset. These data provide a basis for further biomarker development to identify patients at high risk for recurrence, warranting additional scrutiny, more intensive follow-up protocols, and prioritization for adjuvant therapies. Such a limited panel of highly predictive proteins is likely compatible with immunohistochemistry (IHC) assays, which are compatible with standard clinical pathology workflows. DOD CDMRP Funding yesFigure 1 69