MOTIVATION: A major challenge in analyzing cancer patient transcriptomes is that the tumors are inherently heterogeneous and evolving. We analyzed 214 bulk RNA samples of a longitudinal, prospective ovarian cancer cohort and found that the sample composition changes systematically due to chemotherapy and between the anatomical sites, preventing direct comparison of treatment-naive and treated samples. RESULTS: To overcome this, we developed PRISM, a latent statistical framework to simultaneously extract the sample composition and cell-type-specific whole-transcriptome profiles adapted to each individual sample. Our results indicate that the PRISM-derived composition-free transcriptomic profiles and signatures derived from them predict the patient response better than the composite raw bulk data. We validated our findings in independent ovarian cancer and melanoma cohorts, and verified that PRISM accurately estimates the composition and cell-type-specific expression through whole-genome sequencing and RNA in situ hybridization experiments. AVAILABILITYAND IMPLEMENTATION: https://bitbucket.org/anthakki/prism. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.