CeSpGRN: Inferring cell-specific gene regulatory networks from single cell multi-omics and spatial data

Abstract

Single cell gene expression datasets have been used to uncover differences between single cells, leading to discoveries of new cell types and cell identities, which are usually defined by the transcriptome profiles of cells. Biological networks, in particular, gene regulatory networks (GRNs), can be viewed as another feature of the cells, that contributes to the uniqueness of each single cell. However, methods that reconstruct cell-specific GRNs are still missing. We propose CeSpGRN (Cell Specific GRN), which infers cell-specific GRNs from single cell gene expression data. CeSpGRN uses a Gaussian weighted kernel which allows the GRN of a given cell to be learned from the gene expression profile of this cell and cells that are upstream and downstream of this cell in the developmental process. CeSpGRN can be applied to infer cell-specific GRNs in cell populations of any trajectory or cluster structure, and it does not require time information of cells as additional input. We compared the performance of CeSpGRN and baseline methods on simulated datasets obtained under various settings. CeSpGRN showed superior performance in reconstructing the GRN for each cell, as well as in detecting the regulatory interactions that differ between cells. We also applied CeSpGRN to real datasets including THP-1 human myeloid monocytic leukemia cells and mouse embryonic stem cells, where CeSpGRN suggested a set of interactions between genes that rewire during the differentiation process in these cells.

Learning cell-specific networks from dynamics and geometry of single cells

“…Without trying to address this problem in generality, cell-specific networks allow us to disentangle interactions for a certain subclass of higher-order interactions: those which can be understood to be locally first-order, conditional on some latent variable that is encoded in the cellular context. Finally, since existing cell-specific network inference methods [13, 14] infer undirected edges only, such methods are unable to distinguish the direction of information flow in this example and thus would predict identical networks for either branch.…”

“…tr( X T LX ) is large for rapidly fluctuating X . The term associated to λ 2 is a Lasso [6] term that encourages X to be sparse, reflecting our knowledge of biological networks [37, 13]. Together, the objective (9) encourages both parsimony and sharing of information along the cell state manifold.…”

“…Comparison to other methods To date, literature for the problem of inferring cell-specific GRNs is limited. CeSpGRN [13] is perhaps the method most similar to locaTE in terms of the problem it seeks to address, namely inferring a N ˆG ˆG array of cell-specific gene-gene interaction scores. It is limited to undirected interactions due to use of the (Gaussian) Graphical Lasso.…”

“…The target rank was set to k = 8, L was taken to be the normalised k-NN graph Laplacian, and K 1 = K 2 = I . We chose α = 10, β = 0, λ 1 = λ 2 = μ 1 = μ 2 = 1. For comparison, PIDC [4], GRISLI [19], TENET [16], SCODE [17] and CeSpGRN [13] were applied using the same log-transformed expression values and pseudotime estimates. For TENET a history length of 1 was used, following the usage guidelines in its documentation.…”

“…For comparison, PIDC [4], GRISLI [19], TENET [16], SCODE [17] and CeSpGRN [13] were applied using the same log-transformed expression values and pseudotime estimates. For TENET a history length of 1 was used, following the usage guidelines in its documentation.…”