Predicting anti-PD-1 immune checkpoint blockade response in melanoma patients with spatially aware machine learning models.

There is an acute need to accurately identify patients with advanced melanoma who are most likely to respond to anti-PD1 immune checkpoint blockade (ICB) therapy. While anti-PD1 therapy can be highly effective in advanced melanoma patients, only 30-40% of patients respond well. In this study, we apply single-cell spatial proteomics together with statistical and machine learning (ML) methods to successfully predict advanced melanoma patient response to anti-PD1 ICB in a cohort of 12 patients with >8 million cells. While no single molecular feature is sufficient to predict ICB response in our cohort, ML models integrating multiple molecular features accurately predict response in 11 of 12 patients. A recurrent cellular neighborhood analysis revealed a tumor-infiltrating lymphocytes niche that was present in the tumors of most responders. This neighborhood, tumor microenvironment immune cell composition, and levels of nitric oxide synthases were all important features used by our ML models to make accurate predictions. Optimal predictive performance by our ML models-a ROC AUC of 0.76-was achieved when using all molecular features, including cellular spatial relationships, but limiting our analysis to only immune-rich tissue regions. This study demonstrates the feasibility of using machine learning models to accurately predict patient response to anti-PD1 ICB therapy using spatial proteomics datasets.