Figure 1.

The pipeline of CellTICS. (A) Marker genes are prioritized by considering their cell-type-specific high and low expression measures. (B) The Reactome pathways are used to construct the network hierarchy. (C) A neural network constructed by the pathway hierarchy is trained and then used to predict cell types of the test set. The predicted results from all predictive layers are integrated by voting. (D) Based on the predicted results of the first stage, several sub-neural networks are trained for sub-cell-type prediction. Important pathways defining a cell type (or a sub-cell type) can be obtained by comparing their activation values across different cell types.