Generation of multi-omic datasets using high-throughput molecular profiling of transcriptomic human data

Tumor heterogeneity significantly affects cancer progression and therapeutic response, yet quantifying it from bulk molecular data remains challenging. Deconvolution algorithms, which estimate cell-type proportions in bulk samples, offer a potential solution. However, there is no consensus on the optimal algorithm for transcriptomic or methylomic data. Here, we present an unbiased evaluation framework for the first comprehensive comparison of deconvolution algorithms across both omic types, including reference-based and -free approaches. Our evaluation covers raw performance, stability, and computational efficiency under varying conditions, such as missing or additional cell types and diverse sample compositions. We design a reproducible workflow using containerization and publicly available code to ensure transparency and re-usability. Our results highlight the strengths and limitations of various algorithms, providing practical guidance for selecting the best method based on data type and context. This benchmark sets a new standard for evaluating deconvolution methods and analyzing tumor heterogeneity. Overall design: This dataset consists of 30 mixtures (and nine pure cell types) constructed to recapitulate the heterogeneity seen in real pancreatic adenocarcinoma (following a Dirichlet distribution model). These mixes contained variable proportion of (i) human tumor cells (from 2 subtypes), (ii) cancer associated fibroblasts, (iii) human tumor derived-endothelial cells and immune cells that were FACS sorted from healthy donors (B, T4, T8 lymphocytes and neutrophils) and differentiated (M2-type macrophages). Cells were mixed and RNA/DNA were simultaneously extracted followed by RNAseq (RNA-seq poly A) and methylome (MethEPIC 850K).