Batch alignment of single-cell transcriptomics data using deep metric learning.

scRNA-seq has uncovered previously unappreciated levels of heterogeneity. With the increasing scale of scRNA-seq studies, the major challenge is correcting batch effect and accurately detecting the number of cell types, which is inevitable in human studies. The majority of scRNA-seq algorithms have been specifically designed to remove batch effect firstly and then conduct clustering, which may miss some rare cell types. Here we develop scDML, a deep metric learning model to remove batch effect in scRNA-seq data, guided by the initial clusters and the nearest neighbor information intra and inter batches. Comprehensive evaluations spanning different species and tissues demonstrated that scDML can remove batch effect, improve clustering performance, accurately recover true cell types and consistently outperform popular methods such as Seurat 3, scVI, Scanorama, BBKNN, Harmony et al. Most importantly, scDML preserves subtle cell types in raw data and enables discovery of new cell subtypes that are hard to extract by analyzing each batch individually. We also show that scDML is scalable to large datasets with lower peak memory usage, and we believe that scDML offers a valuable tool to study complex cellular heterogeneity.