Patches: A Representation Learning Framework for Decoding Shared and Condition-Specific Transcriptional Programs in...

Single-cell genomics enables the study of cell states and cell state transitions across biological conditions like aging, drug treatment, or injury. However, existing computational methods often struggle to simultaneously disentangle shared and condition-specific transcriptional patterns, particularly in experimental designs with missing data, unmatched cell populations, or complex attribute combinations. To address these challenges, Patches identifies universal transcriptomic features alongside condition-dependent variations in scRNA-seq data. Using conditional subspace learning, Patches enables robust integration, cross-condition prediction, and biologically interpretable representations of gene expression. Unlike prior methods, Patches excels in experimental designs with multiple attributes, such as age, treatment, and temporal dynamics, distinguishing general cellular mechanisms from condition-dependent changes. We applied Patches to both simulated data and real transcriptomic datasets from skin injury models, focusing on the effects of aging and drug treatment. Patches revealed shared wound healing patterns and condition-specific changes in cell behavior and extracellular matrix remodeling. These insights deepen our understanding of tissue repair and can identify potential biomarkers for therapeutic interventions, particularly in contexts where the experimental design is complicated by missing or difficult-to-collect data.