Certain aspects of diagnosis, prognosis, and treatment of cancer patients are still important challenges to be addressed. Therefore, we propose a pipeline to uncover patterns of alternative polyadenylation (APA), a hidden complexity in cancer transcriptomes, to further accelerate efforts to discover novel cancer genes and pathways. Here, we analyzed expression data for 1045 cancer patients and found a significant shift in usage of poly(A) signals in common tumor types (breast, colon, lung, prostate, gastric, and ovarian) compared to normal tissues. Using machine-learning techniques, we further defined specific subsets of APA events to efficiently classify cancer types. Furthermore, APA patterns were associated with altered protein levels in patients, revealed by antibody-based profiling data, suggesting functional significance. Overall, our study offers a computational approach for use of APA in novel gene discovery and classification in common tumor types, with important implications in basic research, biomarker discovery, and precision medicine approaches.