The human genome encodes ~1,900 secreted proteins, many of which mediate intercellular communication. Secreted proteins do not act cell-autonomously, limiting systematic approaches to characterize their functions. Here we introduce SecAct (Secreted Activity, https://secact.ccr.cancer.gov ), a computational framework that infers the signaling activities of 1,170 human secreted proteins from spatial, single-cell and bulk transcriptomic data. The inference model harnesses precomputed intercellular signaling signatures trained on 1,258 spatial transcriptomics samples spanning 37 cancer types. Transcriptomics data from antisecreted protein therapies validate SecAct's accuracy in predicting the repression of secreted protein activity following treatment. For spatial and single-cell transcriptomics data, SecAct provides interactive modules for analyzing secreted protein-mediated cell-cell communication. Applying SecAct to 54 cancer immunotherapy cohorts comprising 5,174 patients, we identified secreted proteins associated with tumor immunity. In vivo experiments validated lymphocyte antigen 86 (LY86), whose function in cancer was previously unknown, as an antitumor regulator.