Automated tumor immunophenotyping and response to immunotherapy in non-small cell lung cancer using a spatial statistics approach.

Abstract

2626 Background: Based on the clinical success of immuno-oncology therapeutics, tissue-based, biomarker analyses have shifted from a focus on tumor cell phenotypes to spatial and functional analyses of the tumor immune environment. Distribution and composition of immune infiltrates have shown prognostic or predictive value in various studies; however, standardized approaches to categorize tumors into “Desert”, “Excluded” or “Inflamed” immunophenotypes based on the density and pattern of immune infiltrates are missing. This categorization is typically based on visual inspection of a stained tissue section; it is labor-intensive and associated with poor inter-observer concordance. To eliminate these barriers we developed an automated approach that relies on a set of derived spatial features and named this analysis pipeline LATIS (metric Learning based Automated Tumor Immunophenotyping with Spatial statistics). Methods: We used two clinical trials, POPLAR (phase II; n=258) and OAK (phase III; n=623) that compared anti-PD-L1 treatment vs chemotherapy in patients with advanced non-small cell lung cancer to develop and validate this approach. LATIS utilizes scans of slides stained immunohistochemically for CD8 and cytokeratin (CK); images were tiled and classified according to CK status (positive/negative). First, CD8 positive and CK density was calculated for each tile. Twenty six spatial features were then extracted for each tiled image scan. We used a labeled (manual immunophenotype calls) POPLAR dataset to learn a metric that separates classes of labeled data best with respect of spatial features provided for each data point. Then, to co-embed OAK dataset with POPLAR dataset into the same space, we used that learned metric as a measure of distance between new unlabeled points (OAK data set). K-means clustering in the embedded space was used to identify clusters of data points followed by QFMatch to assign class labels (“Inflamed”, “Excluded”, “Desert”) to each data point. Results: LATIS successfully relates spatial features to both, manual reads and patient response to therapy. In OAK, anti-PD-L1 treated patients experienced longer median overall survival (OS) when tumors showed intra-epithelial CD8+ cells (Inflamed) compared to tumors with a stromal pattern of CD8+ cells (Excluded) or a low density of CD8+ cells (Desert): 17.6 months vs 10.3 (Excluded) vs 12.1 (Desert) [p=0.0061]. Median progression-free survival (PFS) was 3.2 months (Inflamed) vs 2.6 (Excluded) vs 1.4 (Desert) [p=0.00058]. OS and PFS were not significantly different for the three categories in chemotherapy only arm. Conclusions: We suggest that tumor immunophenotype categories generated in an automated fashion by LATIS can serve as predictive biomarker for cancer immunotherapy.