While immune checkpoint-based immunotherapy (ICI) shows promising clinical results in cancer patients, only a subset of patients responds favorably. Response to ICI is dictated by complex networks of cellular interactions between malignant and non-malignant cells. Although insights into the mechanisms that modulate the pivotal anti-tumoral activity of cytotoxic T-cells (Tcy) have recently been gained, much of what has been learned is based on single-cell analyses of dissociated tumor samples, resulting in a lack of critical information about the spatial distribution of relevant cell types. Here, we used multiplexed immunohistochemistry to spatially characterize the immune landscape of metastatic melanoma from responders and non-responders to ICI. Such high-dimensional pathology maps showed that Tcy gradually evolve towards an exhausted phenotype as they approach and infiltrate the tumor. Moreover, a key cellular interaction network functionally linked Tcy and PD-L1+ macrophages. Mapping the respective spatial distributions of these two cell populations predicted response to anti-PD-1 immunotherapy with high confidence. These results suggest that baseline measurements of the spatial context should be integrated in the design of predictive biomarkers to identify patients likely to benefit from ICI.