Abstract
Melanoma is the deadliest form of skin cancer. Due to its high mutation rates, melanoma is a convenient model to study antitumor immune responses. Dendritic cells (DCs) play a key role in activating cytotoxic CD8+ T lymphocytes and directing them to kill tumor cells. Although there is evidence that DCs infiltrate melanomas, information about the profile of these cells, their activity states, and potential antitumor function remains unclear, particularly for conventional DCs type 1 (cDC1). Approaches to profiling tumor-infiltrating DCs are hindered by their diversity and the high number of signals that can affect their state of activation. Multiplexed immunofluorescence (mIF) allows the simultaneous analysis of multiple markers, but image-based analysis is time-consuming and often inconsistent among analysts. In this work, we evaluated several machine learning (ML) algorithms and established a workflow of nine-parameter image analysis that allowed us to study cDC1s in a reproducible and accessible manner. Using this workflow, we compared melanoma samples between disease-free and metastatic patients at diagnosis. We observed that cDC1s are more abundant in the tumor infiltrate of the former. Furthermore, cDC1s in disease-free patients exhibit an expression profile more congruent with an activator function: CD40highPD-L1low CD86+IL-12+. Although disease-free patients were also enriched with CD40-PD-L1+ cDC1s, these cells were also more compatible with an activator phenotype. The opposite was true for metastatic patients at diagnosis who were enriched for cDC1s with a more tolerogenic phenotype (CD40lowPD-L1highCD86-IL-12-IDO+). ML-based workflows like the one developed here can be used to analyze complex phenotypes of other immune cells and can be brought to laboratories with standard expertise and computer capacity.