Abstract
Cellular deconvolution aims to estimate cell type fractions from bulk transcriptomic and other omics data. Most existing deconvolution methods fail to account for the heterogeneity in cell type-specific (CTS) expression across bulk samples, ignore discrepancies between CTS expression in bulk and cell type reference data, and provide no guidance on cell type reference selection or integration. To address these issues, we introduce BLEND, a hierarchical Bayesian method that leverages multiple reference datasets. BLEND learns the most suitable references for each bulk sample by exploring the convex hulls of references and employs a "bag-of-words" representation for bulk count data for deconvolution. To speed up the computation, we provide an efficient EM algorithm for parameter estimation. Notably, BLEND requires no data transformation, normalization, cell type marker gene selection, or reference quality evaluation. Benchmarking studies on both simulated and real human brain data highlight BLEND's superior performance in various scenarios. The analysis of Alzheimer's disease data illustrates BLEND's application in real data and reference resource integration.