A Single-Cell Imputation Method Based on Mixture Models and Neural Networks

In the field of single-cell sequencing, single-cell imputation has always been an important research direction. In the single-cell matrix, there are not only zeros that are truly expressed, but also many non-zero items whose expression values are too low to become zeros. The latter phenomenon is called ‘Dropout’. In the past, old imputation models often considered the connection between cells and ignored the role of highly expressed genes. To address this difficulty, we introduced a new imputation method that uses a mixture model to Identify highly expressed genes, which are then fed into the neural network imputation model. The high-expressed genes are used as datasets to learn and train the model, and the predicted results are outputted and interpolated into the corresponding low-expressed models. The comparative evaluation of human datasets and mouse datasets shows that this method can effectively identify dropout in cells. Value, strengthen the clustering between cell populations, and improve the ability of differential expression analysis.