Fig. 1

The overall framework of scMTD. a scMTD reorders the cells based on the pseudo-time inferred by TSCAN to construct the cell-state specific space. b According to the local cell neighbors in each space, scMTD predicts gene expressions of each cell and denotes the cell-level imputation for gene \(i\) and cell \(j\) as \(c_{ij}\). c scMTD constructs the specific gene co-expression network in each space to predict expressions of each gene and denotes the gene-level imputation for gene \(i\) and cell \(j\) as \(g_{ij}\). d scMTD utilizes a decreasing logistic model to estimate the dropout probability for expression entries in each space and denotes the dropout probability for gene \(i\) and cell \(j\) as \(p_{ij}\). e scMTD combines \(c_{ij}\),\(g_{ij}\), and \(p_{ij}\) to impute raw expression entries, where \(x_{ij}\),\(\hat{x}_{ij}\) represent the raw expression and imputed expression for gene \(i\) and cell \(j\), and \(\alpha ,\beta\) are the standard deviation of cell-level and gene-level predictions, respectively