BLEND: probabilistic cellular deconvolution with individualized single-cell reference integration

Abstract

Cellular deconvolution estimates cell-type fractions from bulk transcriptomic data, but current methods often overlook cell type-specific expression varying across samples, discrepancies between bulk and single-cell data, or lack guidance on reference data selection and integration. Therefore, we present BLEND, a hierarchical Bayesian method that leverages multiple single-cell reference datasets to perform cellular deconvolution. BLEND estimates cellular fractions accurately by learning the most suitable reference for each bulk sample, accounting for the aforementioned issues. BLEND outperforms state-of-the-art methods in comprehensive benchmarking studies using human brain cortex data and provides reliable insights into Alzheimer’s disease progression.