11To fully utilize the power of single-cell RNA sequencing (scRNA-seq) technologies for generating cell 12 lineages and identifying bona fide transcriptional signals, it is necessary to combine data from multiple 13 experiments. We present BERMUDA (Batch-Effect ReMoval Using Deep Autoencoders), a novel transfer-14 learning-based method for batch-effect correction in scRNA-seq data. BERMUDA can be effectively applied 15 to batches with vastly different cell population compositions, and can properly combine different batches 16 while transferring biological information from one batch to amplify the corresponding signals in other 17 batches. We demonstrate that BERMUDA outperforms multiple existing methods in both removing batch 18 effects and separating different cell types.
Background 21Understanding cellular transcriptional dynamics at the single-cell resolution is crucial to identify disease 22 mechanisms and understand normal cell physiology in heterogeneous tissue settings. Up until recently, the 23 45 project cells from different experiments to a common bias-reduced low-dimensional representation.
46However, this type of correction does not account for the variations in cellular heterogeneity among studies, 47 such as different cell types or different numbers cells within each cell type. Alternatively, to account for this 48 heterogeneity, mnnCorrect utilizes mutual nearest neighbors (MNN) between batches to identify matching 49 cell types through MNN pairs [22]. By identifying the corresponding cells, a cell-specific correction can be 50 learned for each MNN pair. As a result of performing batch-correction locally, mnnCorrect avoids the 51 3 assumption of similar cell population compositions between batches required by previous methods.
52Following mnnCorrect, a series of new methods have been developed to integrate scRNA-seq data from 53 different experiments [23][24][25][26][27]. For example, in order to account for the cellular heterogeneity between 54 studies, Seurat v3 [23] utilizes MNN pairs between the reference batch and query batches to detect 55 "anchors" in the reference batch. "Anchors" represents cells in a shared biological state across batches and 56 are further used to guide the batch correction process through CCA. BBKNN [24] takes advantage of a 57 more efficient process for creating neighborhood graphs, which are further incorporated into visualization 58 and clustering of the cells. However, most existing batch correction methods for scRNA-seq data rely on 59 generating neighborhood graphs based on pairwise similarities between cells, which do not fully utilize the 60 clustering structures of different cell populations to identify the optimal batch-corrected subspace.
61In this paper, by considering scRNA-seq data from different batches as different domains, we take 62 advantage of the domain adaptation framework in deep transfer learning to properly remove batch effects 63 by finding a low-dimensional representation of the data. Different from existing methods, the proposed 64 method, B...