Abstract Background: Most tumor samples consist of a variable proportion of malignant and nonmalignant cells including epithelial cells, fibroblasts, and infiltrating immune cells, which confounds biomarker studies of response to treatment. Deconvolution approaches have been developed for transcriptomes to address this heterogeneity in tumor samples. The Cancer Genome Atlas (TCGA) project has generated high-throughput RNAseq in over 11,000 patient samples across 33 cancer types. Using these data, we systematically investigated Pan-Cancer cell-type-specific transcriptional activities using deconvolution tools. Method: We established a new deconvolution framework, DeMixT, which deconvolves high-dimensional transcriptome data from mixtures of tumor and stromal components. Besides estimating mixing proportions, DeMixT uniquely provides per-gene per-sample expression levels of each component. To further address variations observed in real data, we propose a profile likelihood-based prefiltering method to adaptively select a gene set with a high signal-to-noise ratio for proportion estimation, which sequentially improves gene expression estimation for the whole transcriptome. We applied the pre-filter-enabled DeMixT to 16 TCGA solid primary tumor types: BLCA, BRCA, COAD, HNSC, KICH, KIRC, KIRP, LIHC, LUAD, LUSC, PAAD, PRAD, READ, STAD, THCA, UCEC. In 2 cancer types: PRAD (Prostate Adenocarcinoma) and KIRC (Kidney Renal Clear Cell Carcinoma), we compared results of pathway analyses, clustering, and survival analyses from the original mixed expression data with those from the deconvolved expression data. Results: For the 16 cancer types, we obtained tumor purities and deconvolved individual-level gene expression of both tumor and stromal components. Per cancer type, the mean number of genes recovered was 13,239 (sd=886), and the mean number of tumor samples deconvolved was 371 (sd=201). In PRAD, we identified important pathways that were missed previously but are now statistically significant (FDR<0.1): Epithelial-Mesenchymal-Transition, TP53, NF-κB, Hypoxia, Estrogen Response, Apical Junction, IL2-STAT5 Signaling, and KRAS Signaling. A closer look at Urea Cycle Dysregulation (UCD) genes found augmented downregulation of ASS1 and intensified upregulation of SLC25A13 and SLC25A15 in the deconvolved tumor vs. normal comparison, consistent with the previous report in Lee et al. Cell 2018. In KIRC, hierarchical clustering of deconvolved expression data resulted in a better enrichment of a PBRM1 mutant in one cluster of patients, who had better survival outcomes. This observation is consistent with the previous report by Kapur et al. Lancet Oncol 2013. Conclusion: We provide comprehensive transcriptome deconvolution results for the TCGA Pan-Cancer datasets. These findings will enable novel studies of admixed human tumor samples and improve the understanding of tumor malignancy. Citation Format: Shaolong Cao, Rongjie Liu, Liuqing Yang, Jaeil Ahn, Jingxiao Chen, Zeya Wang, Eleni Efstathiou, Daniel E. Frigo, Hongtu Zhu, Wenyi Wang. Deconvolution reveals cell-type specific transcriptional effects across cancer types [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 4692.