Research On Cell Type Identification From Single-cell RNA Sequencing Data

The rise of single-cell transcriptome sequencing technology has highlighted an elaborate resolution,which is cell.Different types of cells contribute to the heterogeneity of human tissues.Now,single cell transcriptome sequencing data helps us understand and explain the attribute of a single cell and the heterogeneity of tissues from the perspective of gene expression level.The current method of cell type prediction based on literature researching after unsupervised clustering takes a lot of time when dealing with the increasing data size.With the development of single-cell sequencing technology,a key analytical challenge is how to make use of the reference map composed of the existing information to better study the identification of cell types for new sequencing data.Here,we propose a new transfer learning algorithm,which constructs neural networks for both to accomplish dimensionality reduction and label transfer.For the reference,construct a three-layer nonlinear neural network with decreasing dimensions based on deep classification model to obtain the coding features of the reference which is used to train the classifier.For the new sequencing data,an autoencoder is constructed in order to ensure accurate classification on the coding feature while reserving the original structure.Supposing that the distribution of samples from the same type should be identity can transfer the cell-type information of reference to new sequencing data.Through experiments on the human pancreas dataset and the human peripheral blood monocyte dataset,we demonstrate that the iteration of our method has a result of batch effect correction.We also demonstrate that the low-dimensional representation obtained by our method is friendly to the tasks like clustering and classification.Facing with the imbalanced distribution of datasets,the comparation with the state-of-art methods indicate that our method has the highest accuracy of classification and even performs better on the cell-type prediction of rare cell types.In summary,our method is a transfer learning method for single-cell RNA sequencing data that can achieve dimensionality reduction,batch effect correction,and cell type prediction.