Gene signatures have been shown to predict the disease outcome to immunotherapies but with only modest accuracy. The reduction in this precision might be due to devoid of spatial information, which attenuates the ability to distinguish tumor genes from TME (tumor-microenvironment) genes. Here we portray the role of spatial information in predictiveness. Digital Spatial Profiling (DSP)-GeoMX Whole Transcriptome Atlas (WTA) is used to generate a spatial transcriptomic map of 55-immunotherapy-treated melanoma samples which enables in situ hybridization against 18,190 genes in several areas of interest/compartments (i.e., CD68+ macrophages, CD45+ lymphocytes and S100B+ tumor cells) at high throughput using a sequencing readout. We developed a computational pipeline to discover compartment-specific gene signatures. The signatures were built using a split sample approach with 100 different Lasso logistic regression binomial models to estimate classification accuracy and to identify objective response, followed by deconvolution of bulk-expression into compartmentalized-expression using CIBERSORTx (in silico tissue dissection) to simulate compartment-derived gene expression data. Finally, CIBERSORT results to actual spatially collected gene signatures were compared. We achieved an AUC > 0.9 for all compartment signatures [0.94 (CD45), 0.97 (CD68), 0.93 (S100B), 0.98 (pseudo-stroma) and 0.92 (pseudo-bulk)]. Cross-testing in different compartments (i.e., CD45 signature in CD68 and S100B) showed poor performance, indicating compartment-specificity. Independent cohort (N=90) validation by deconvolving bulk into compartmentalized-gene expression showed an AUC of 0.74 (CI:0.64-0.84) for pseudo-bulk, 0.83 (CI:0.75-0.91) for CD68 and 0.66 (CI:0.54-0.77) for S100B 8-gene signatures. Poor performance with AUC:0.59 (CD45), AUC:0.56 (CD68) and AUC:0.65 (S100B) was observed when evaluating the deconvolution platform using CIBERSORTx (pseudo-compartments from pseudo bulk) with real compartment data. Spatial de novo CD45-, CD68- and S100B-signatures excel in performance with strikingly higher AUCs than computationally deconvoluted signatures. We believe that the spatially informed signatures differentially evaluate tumor vs TME and thus may be more accurate in predicting treatment outcome.