RBET guides case-specific choice of batch correction methods for single-cell transcriptomic data with overcorrection awareness and large batch effect robustness

Abstract Integrating single-cell RNA-sequencing datasets from different sources is a common practice to empower in-depth interrogation for biological insights, where batch effect correction (BEC) is of vital importance. However, an inappropriate BEC may lead to overcorrection and report misleading results on downstream analyses including cell annotation, trajectory inference and cell-cell communication. Hence, we develop the Reference-based Batch Effect Testing (RBET), a novel statistical framework for evaluating the performance of different BEC methods by leveraging housekeeping-gene inspired reference genes and MAC statistics for distribution comparison. In addition to higher batch effect detection power on simulated and real datasets than the existing tools, RBET is overcorrection sensitive, computationally efficient, and robust to large batch effect sizes. Furthermore, extensive multi-scenario real examples show that RBET selects optimal BEC tools for consistent downstream analysis results, which confirm prior biological knowledge. This comprehensive BEC decision-making tool is available as an R package.