Purpose of study: Non-small cell lung cancer (NSCLC) is the most common and fatal of cancers. In this work, we examine multiplexed images of NSCLC tumors to investigate both the spatial and spatiotemporal eco-evolutionary interactions between the tumor and its microenvironment, to better understand NSCLC tumor progression and therapy response. Methods: We developed a scalable, computational image analysis pipeline using cell-segmentation and quadrat-based approaches to analyze the spatial and temporal features of high-dimensional multiplexed NSCLC images. Multiplexed images enable the spatial readouts of numerous biomarkers per tissue sample and allow the interpretation of cellular states and the characterization of tumor-immune interactions across tissue ensembles. We also implement statistical approaches for ecological niche modelling combined with machine learning and deep learning models to predict disease progression and identify clinical imaging biomarkers. Data: Images were obtained from two 9-patient cohorts having advanced/metastatic NSCLC who were treated with the oral HDAC inhibitor vorinostat combined with the PD-1 inhibitor pembrolizumab. The first cohort had 4 progressors (PD) and 5 with stable disease (SD). The second cohort had 3 patients each in the PD, SD and partial response (PR) categories. Images were collected from all patients both pre- and on-treatment (during the third week). Results: Using our computational framework based on cell segments and quadrats, we confirm that different spatial neighborhoods exist that distinguish PD from SD, and that these spatial ecologies aid disease progression: PD patients have distinct ecologies with higher colocalization of PanCK+PD-1+FoxP3 indicating an immunosuppressive environment, whereas SD patients have a higher colocalization of PanCK+PD-L1 along with T cells, suggesting immunoactive tumor regions. These can be considered as potential biomarker candidates for predicting tumor progression. In an additional experiment where we include PR samples in our analyses, these distinct spatial neighborhoods are reinforced amongst PD, SD and PR patient groups corroborating the existence of spatiotemporal patterns. Further, we were able to predict treatment response with >91% accuracy given each patient’s spatial distribution of cell types from pre-treatment images. Using these predictions, we can generate risk maps at the patient level to identify areas of the tumor that are indicators of a higher probability of progression during treatment. Conclusions: We leveraged both single-cell and quadrat-resolution analysis of multiplexed imaging data and identified fundamentally distinct spatial ecologies between PD and SD patients. The ecology in PD patients appears to be primed for immune resistance even before treatment. This ecological diversity between SD and PD patients acts as a biomarker that enables accurate disease progression prediction. Our disease progression predictions can be used in conjunction with standard PD-L1 status to further strengthen personalized treatment strategies.
Citation Format: Sandhya Prabhakaran, Chandler Gatenbee, Mark Robertson-Tessi, Amer A. Beg, Jhanelle Gray, Scott Antonia, Robert A. Gatenby, Alexander R.A. Anderson. Distinct spatiotemporal tumor-immune ecologies enable disease prediction in NSCLC patients [abstract]. In: Proceedings of the AACR Special Conference on the Evolutionary Dynamics in Carcinogenesis and Response to Therapy; 2022 Mar 14-17. Philadelphia (PA): AACR; Cancer Res 2022;82(10 Suppl):Abstract nr B020.