Optimization and redevelopment of single-cell data analysis workflow based on deep generative models

The present single-cell RNA sequencing(scRNA-seq) analysis pipelines require a combination of appropriate normalization, dimension reduction, clustering, and specific-gene analysis algorithms, but the rationale for the choice of these algorithms is relatively subjective because of the lack of ground truth assessment conclusions. As the number of captured single-cells increases, the number of different types of noise cells also increases, which can strongly affect the analysis efficiency. For scRNA-seq, a technology that generates data through multi-process operations, the deep generative model should be a good choice for this type of data analysis, allowing simultaneous estimation of multiple unobservable parameters assumed in the data generation process. Hence, in our study, we sequenced a pool of pre-labeled single cells to obtain a batch of scRNA-seq data with main and fine labels, which was then used to evaluate the clustering and specific-gene analysis methods. Afterward, we applied two deep generative models to infer the probabilities of pseudo and impurity cells. And by stepwise removing the inferred noise cells, the clustering performance and the consistency of different specific-gene analysis methods are both greatly improved. After that, we applied Deep-LDA (a latent Dirichlet allocation-based deep generative model) to scRNA-seq data analysis. And this model takes the count matrix as input, and makes the classification and specific gene optimization process mutually dependent, which has more practical sense and simplifies the analysis workflow. At last, we successfully implemented the model with transferred knowledge to make single-cell annotation and verified its superior performance.