Spatially resolved immunologic hallmarks of response to neoadjuvant immune checkpoint blockade in metastatic melanoma

Background: The recent groundbreaking neoadjuvant immune checkpoint blockade (NICB) clinical trials in melanoma have demonstrated decisively that the administration of ICB prior to intent-to-cure surgeries will become the new standard-of-care for metastatic melanoma and beyond. The paradigm shifts from radiographic to pathologic response assessment also presents an unparalleled “window of opportunity” for the acquisition of abundant “on-treatment” tissues to elucidate the mechanisms and biomarkers of response and for guiding post-surgery personalized therapy decisions. To date, single-cell/bulk RNA studies have demonstrated that TCF7+ stem-like CD8+ T cells and tertiary lymphoid structures (TLS) are positive predictors of ICB response. However, how these immune cells organize and communicate within the spatial context of the neoadjuvant tumor microenvironment remains poorly understood. Methods: We investigated a unique cohort of 91 FFPE biospecimens from 87 patients with stage III metastatic melanoma, including 60 treated with NICB (31 complete response, 7 partial response, 22 non-response) and 27 treatment-naïve. We deployed state-of-the-art technologies, including multiplexed error-robust fluorescent in situ hybridization (MERFISH), single cell sequencing of FFPE blocks, and multiplexed IF for this rich digital resource. Importantly, to analyze these challenging high dimensional spatial datasets, we developed 3 novel computational algorithms, including SCIRA, a scalable quantification method for spatial receptor-ligand (R-L) interactions; GC-SCAN, a graph-based clustering method to detect Germinal Center (GC)/TLS structures in single cell spatial data; and PathNet, an end-to-end AI algorithm for automated GC/TLS detection on H&E slides. Results: Here, we interrogated ~5.6 million cells and showed that increased GC/mature TLSs, TCF7+ stem-like CD8 and CD4 T-cells, exhausted CD8 T-cells, plasma cells and myeloids in spatially distinct cellular neighborhoods are significantly associated with positive response. Our spatial (R-L) analyses further identified preferential chemokine R-L interactions between GC-B cells and follicular helper T-cell, and TCF7+ stem-like T-cells with CCL19+/CCL21+ fibroblasts in organizing these immune hubs. Lastly, Our PathNet GC/TLS detection algorithm outperforms recent state-of-the-art methods in both specificity and sensitivity, proven potential utility in facilitating clinical response assessment. Conclusions: We leveraged cutting-edge spatial-omics technologies and novel computational methods to resolve the immunologic hallmarks of NICB response in metastatic melanoma. We believe our approach provides a model for how to precisely identify spatially intricate immune interactions that underline treatment response in the rapidly advancing era of standard of care neoadjuvant immunotherapy.