Effectively Preserving Biological Variations in Multi-Batch and Multi-Condition Single-Cell Data Integration

Abstract

Understanding phenotypic differences at the cell level is critical for comprehending the underlying pathogenesis of related complex diseases. However, the biological variations are obscured by batch effects, posing a challenge for integrating multi-batch and multi-condition single-cell datasets. Here, we present scFLASH, a deep learning-based model specially designed to explore single-cell biological variations while correcting undesired batch effects. scFLASH employs a conditional variational autoencoder with adversarial training to separate biological variations from technical noise and introduces a penalized condition classifier to preserve condition-specific biological signals. Through comprehensive benchmarking evaluations, scFLASH shows superior integration performances compared to other state-of-the-art methods. Applied to datasets such as Alzheimer’s disease, COVID-19, and diabetes, we demonstrate that scFLASH is applicable to various scenarios, effectively integrating datasets with two or more conditions and different batch sources. scFLASH can enhance the gene expression profiles and identify the condition-related cell subpopulations, facilitating downstream analyses and offering biological insights into the cellular mechanisms of disease pathology.