Fig. 1. Overview of SciBet algorithm.

a Training set Pre-process by calculating the mean gene expression form the original expression matrix. Here we use marker genes G₁, G₂, and G₃ along with a non-marker gene G₄ as examples. b Using E-test to select cell type-specific genes for the downstream classification. Genes with total entropy difference larger than the predefined threshold will be kept. Genes selected by E-test are used for the model training and prediction. c Training SciBet model by obtaining the parameters for the multinomial models of each cell type. For each cell type, the sum of all parameters belonging to different genes equals to 1, which represent the expression probability of different genes. d Calculating the likelihood function of a test cell using the trained SciBet model and annotating cell type for the test cell with maximum likelihood estimation. Each cell in the test set is independently annotated.