Condition-Associated Pattern Extraction and Recovery From Multi-Condition Single-Cell RNA-seq Data With CAPER

A central challenge in multi-condition single-cell RNA sequencing (scRNA-seq) data analysis is the disentanglement of true biological signals from unwanted variations in complex experimental designs. Current statistical and machine learning-based methods struggle with this task, often providing only visualizable embeddings, over-correcting and discarding biological signal, or failing to resolve cell-type-specific responses. Here, we present CAPER, a matrix factorization framework that explicitly disentangles shared biological states from condition-specific variations. CAPER directly outputs an interpretable, batch-corrected expression matrix in which the signal of interest is preserved and isolated. The performance of CAPER is validated using extensive simulations, followed by three real-world multi-condition scRNA-seq data applications, representing distinct signal-to-noise ratio (SNR) scenarios: a controlled immune stimulation in PBMCs with high SNR, a tumor-microenvironment dataset from LUAD with confounded SNR, and a complex autoimmune disease dataset from T1D with low SNR. Across these settings, CAPER yields interpretable latent factors linked to relevant biology, accurately recovers key differentially expressed genes, and correctly identifies the most responsive cell populations. CAPER is a robust and interpretable tool for recovering biological signals from multi-condition single-cell RNA-seq data, enabling reliable discovery in disease research and functional genomics.