Concurrent de novo inference of cell types and gene regulatory networks with NetMap
Existing methods for GRN inference based on single-cell RNA-sequencing (scRNA-seq) data typically require prior cell type calling and then infer GRNs at cell type resolution. This risks propagation of errors from cell annotation to GRN inference. Moreover, it makes existing methods blind to regulatory programs that are shared between different cell types, as well as to subtle differences in regulatory activity between cells of the same type. To address these limitations, we present NetMap, a pipeline that jointly infers cell-level GRNs and clusters cells based on shared regulatory programs, combining state-of-the-art single-cell autoencoders and explainable artificial intelligence methods. NetMap outperforms existing methods on synthetic data with a known ground-truth GRN and is able to reliably detect biologically plausible regulons on real data. Moreover, requiring only a small set of user-provided marker genes, NetMap can be used for GRN structure-informed probabilistic cell type assignment. We demonstrate the utility of this approach using PBMC data, where we achieve near perfect clustering and cell type recovery performance compared with manually created cell type labels. Finally, we applied NetMap to a data set with naive, activated and exhausted T-cells, and were able to verify that NetMap retrieves relevant T-cell state signatures from the data, with two approaches. NetMap can be installed as a python package from https://github.com/bionetslab/netmap and is compatible with the scverse system. Scripts to reproduce our results can be found at https://github.com/bionetslab/netmap-evaluation.