12115 Background: Immune checkpoint inhibitors are now approved for use as therapy in advanced stage NSCLC. These drugs can decrease risk of progression by up to 60% when compared to standard chemotherapeutic regimens, but only about 20% of treated patients show significant benefit. The current gold standard for predicting response is increased tissue expression of PD-L1, but recent studies have shown this measure to be inadequate. TILs are correlated with PD-L1 levels and with antigen-induced anti-tumor immune pressure, with increased TILs associated with treatment response and longer survival. Recent work suggests that the spatial arrangement of TILs may be prognostic of outcome in several different cancer types. Here we evaluate whether computer-extracted features relating to spatial arrangement of TILs on digitized H&E images could predict response to Nivolumab in NSCLC. Methods: The study included fifty-six NSCLC patients with diagnostic tumor biopsies who were treated with Nivolumab. Responders and non-responders were classified according to clinical improvement and radiologic assessment by RECIST criteria on the first post-treatment CT scan. Two expert pathologists manually delineated tumor regions on digitized H&E images of the biopsies. Computerized algorithms automatically identified TILs within these regions, defined TIL clusters based on TIL proximity, and utilized network graph concepts to capture measurements relating to arrangement of these TIL clusters. Results: The top five features determined by a statistical feature selection method reflected the area and density of TIL clusters and the spatial proximity of the TILs to each other and to tumor cells. A machine learning classifier trained on these top 5 features had an AUC of 0.76 on the training set (n = 32) and an AUC of 0.64 on an independent validation set from another institution (n = 24). Conclusions: Computer extracted features relating to spatial arrangement of TIL clusters on digitized H&E images distinguished between patients who did and did not respond to Nivolumab. These findings need to be validated on in larger, multi-site validation sets.