ABSTRACT: Non–small cell lung cancer (NSCLC) has an increasing number of targeted and systemic therapies allowing subsets of patients to achieve long–term durable benefit. Fundamental to understanding responses to a given therapy is comprehensive molecular characterization of the underlying tumor immune microenvironment (TIME). The TIME may inform models to predict immunotherapy outcomes and features to delineate therapeutic responses and clinical endpoints. We hypothesize that an integrated multi–omics approach will uncover interactions within the NSCLC TIME and identify novel biomarkers that are predictive of immunotherapy responses, thus aiding precision oncology. To develop a spatially resolved TIME model for NSCLC immunotherapy, we utilized a multi–omics approach, combining spatial mapping of protein expression at the single–cell resolution by Phenocycler Fusion (PCF) and multi–cellular readout whole transcriptome profiling at cellular compartment resolution by Digital Spatial Profiling (DSP–GeoMx–WTA). This approach facilitated a detailed examination of the TIME in NSCLC samples from patients undergoing first–line immunotherapy. We studied two independent cohorts of advanced NSCLC tissue samples, treated with PD–1–based immunotherapies. We derived gene signatures from cell type signatures to predict treatment outcomes using a multistage Least Absolute Shrinkage and Selection Operator (LASSO) approach. Our spatial proteomic analysis identified three distinct cell types, proliferating tumor cells, granulocytes, and vessels, associated with resistance to immunotherapy. A high proportion of these cell types demonstrated a hazard ratio (HR) of 3.8 (p = 0.004) in the Yale training cohort and 1.8 (p = 0.05) in the UQ validation cohort. In the response cell–type model, higher levels of M1 macrophages, M2 macrophages, and CD4 T cells had an HR of 0.4 (p = 0.019) in the training cohort and 0.49 (p = 0.036) in the validation. In the transcriptomic analysis, gene signatures derived from these cell types predicted outcomes with high accuracy. The resistance gene model, which included 8 genes associated with epithelial–mesenchymal transition (EMT) and cell migration, showed an HR of 5.3 (p < 0.001) in the Yale training cohort, 2.2 (p = 0.036) in the UQ validation cohort, and 1.7 (p = 0.042) in the Greece validation cohort. Conversely, the response gene model, consisting of 8 genes associated with immunomodulation, had an HR of 0.22 (p = 0.005) in the Yale training cohort, 0.38 (p = 0.034) in the UQ validation cohort, and 0.56 (p = 0.041) in the Greece validation cohort. Multivariable analysis, adjusting for age, sex, disease stage, prior chemotherapy, type of immunotherapy, line of immunotherapy, histology, and smoking status, confirmed that the resistance signature remained significantly associated with worse outcomes, while the response signature remained associated with improved outcomes in both the UQ and Greece cohorts. This research highlights the potential of a multi–omics cell typing and gene expression profiling approach in advancing NSCLC treatment toward precision oncology. By offering insights into the TIME and unveiling novel biomarkers, our model seeks to define resistance and to improve prediction of response to treatment.
