bioLUCID: quantifying batch and biological effect in single-cell transcriptomics by hypothesis-driven variance decomposition

Abstract Batch-correcting single-cell gene expression data enables multi-sample analyses but risks erasing sample-specific biology along with batch effects. To address this, we introduce a novel heuristic to decompose gene expression variation into biological and batch effects. This decomposition allows (i) integrating cell embeddings across samples without erasing sample-specific biology, and (ii) flagging samples with excessive batch effects, with benefits for multiple applications such as exploring transcriptional heterogeneity and multi-omics integration.