Deconvolution of single-cell and bulk transcriptomes reveals key cell types and regulatory networks in cattle

Understanding the cellular heterogeneity of mammary gland is crucial for elucidating genetic basis of milk production. While bulk RNA-seq lacks cellular resolution and single-cell RNA-seq is cost-prohibitive for large studies, computational deconvolution offers an alternative. Here we present an optimized deconvolution pipeline based on a bovine mammary single-cell atlas of 198,538 cells and 12 cell types, through evaluation of 14 cell-type proportion estimation methods and three cell-type-specific expression inference methods on simulated and paired real data. This pipeline employs an SCTransform-normalized single-cell reference, estimates cell-type proportions with BayesPrism, and infers cell-type-specific expression profiles using bMIND. Applying it to 188 bulk transcriptomes, we identify luminal secretory (LumSec) and myoepithelial cells as drivers of lactation and involution, respectively. We discover a potential BHLHA15-RHOF-TMEM120B regulatory axis potentially coordinating vesicular transport in LumSec_lac cells, and a putative ESR1-SERPINE1-PLAU circuit governing tissue remodeling in myoepithelial cells during involution. This framework enables cost-effective cell-type-resolution analysis of transcriptomes and provides mechanistic targets for improving milk production traits.