Immune checkpoint blockade (ICB) can induce durable remission in a subset of patients with advanced cancer, but accurately identifying responders before treatment is essential to maximize clinical benefit and avoid unnecessary immune-related toxicities. To address this need, we systematically analyzed bulk RNA-seq cohorts from the Tumor Immunotherapy Gene Expression Resource (TIGER) encompassing multiple tumor types treated with ICB, and developed a machine learning-based framework for response prediction in a pan-cancer, multi-cohort setting. We included cohorts of glioblastoma, head and neck squamous cell carcinoma, non-small cell lung cancer, renal cell carcinoma, gastric cancer, and several melanoma datasets, with all samples annotated as responders or non-responders. Within each cohort, we performed differential gene expression analysis between responders and non-responders and applied Hallmark gene set enrichment to define key biological pathways associated with ICB response. On this basis, we extracted transcriptomic features and trained support vector machine (SVM) classifiers. Model performance was assessed by repeated 10-fold cross-validation within cohorts and further examined using leave-one-dataset-out (LODO) validation and cross-dataset testing among melanoma cohorts. Across tumor types, responders exhibited a highly consistent immune-activation landscape, including marked up-regulation of interferon-gamma/alpha responses, allograft rejection, and TNF-NF-κB and IL6-JAK-STAT3 inflammatory signaling, together with metabolic reprogramming involving oxidative phosphorylation and cholesterol homeostasis. Non-responders more frequently showed enrichment of cell-cycle and proliferation pathways (G2M checkpoint, E2F targets) and epithelial-mesenchymal transition, with tumor type-specific patterns in melanoma and renal cell carcinoma. SVM models achieved good discrimination, with many within-cohort areas under the ROC curve (AUCs) exceeding 0.7 and some approaching 0.9, while maintaining practically useful performance in cross-cohort evaluations. Collectively, this study delineates shared immune and metabolic programs associated with ICB benefit and tumor type-dependent resistance features, and proposes a transcriptome-machine learning prediction framework applicable across multiple cancers and cohorts, providing a solid data and methodological foundation for further optimization of immunotherapy response prediction using advanced representation learning approaches.