Two-Stage Alignment Framework for Batch Effect Removal and Cell Type Classification in ScRNA-seq

Cell type annotation plays a crucial role in scRNA-seq data analysis, but traditional manual annotation approach is time-consuming and susceptible to subjective factors and errors. While supervised classification methods alleviate this problem, there are risks associated with this label transfer strategy. For example, the sensitivity of batch effects may affect the classification results of the target data. To address this issue, we propose a two-stage alignment framework(TAF) based on domain adaptation, which aims to address batch effects in scRNA-seq data and perform cell type classification. First, the nonlinear features with the highest correlation between each source domain reference dataset and the target domain query dataset are extracted using deep generalized canonical correlation analysis (DGCCA). Second, the first stage of alignment is achieved by domain-specific feature extraction, in which the distribution of each source domain reference dataset and the query dataset in the target domain are aligned, and the domain-invariant feature representation is learned. Then, the second stage of alignment is performed by domain-specific labeled classifiers, and the output probability matrices of the query datasets are aligned using virtual adversarial training (VAT) and inconsistent loss constraint, which adjusts the prediction results of target samples near the category boundaries to achieve aligning the outputs of different classifiers. Finally, weights are assigned to the classifiers’ output probability matrices based on the reference dataset’s sample size to obtain more accurate prediction labels. Experimental results on four human pancreatic datasets demonstrate that the TAF method outperforms some state-of-the-art methods.