SCENIC: single-cell regulatory network inference and clustering.

TLDR: On a compendium of single-cell data from tumors and brain, it is demonstrated that cis-regulatory analysis can be exploited to guide the identification of transcription factors and cell states.

Abstract: We present SCENIC, a computational method for simultaneous gene regulatory network reconstruction and cell-state identification from single-cell RNA-seq data (http://scenicaertslaborg) On a compendium of single-cell data from tumors and brain, we demonstrate that cis-regulatory analysis can be exploited to guide the identification of transcription factors and cell states SCENIC provides critical biological insights into the mechanisms driving cellular heterogeneity

Figures

Figure 4. SCENIC overcomes tumor batch effect and recovers relevant cell types and GRNs in oligodendroglioma. (a) Comparison of batch-effect removal methods on an oligodendroglioma dataset. t-SNEs and diffusion plots on the raw expression matrix (first row), after correcting by tumor of origin with Combat or Limma (rows 2-3), or on the binary activity matrix from SCENIC (row 4). The cells are coloured based on the tumor of origin or GRN activity (red: astrocyte-like regulons, green: oligodendrocyte-like regulons, blue: regulons related to cell cycle or stemness). (b) Simplified binary regulon activity matrix (output of SCENIC) for the oligodendroglioma dataset. Highlighted regulons (colored TF names) are known to be characteristic in oligodendrocytes or astrocytes, respectively.

Figure 5. SCENIC reveals melanoma heterogeneity. (a) t-SNE on the raw melanoma expression matrix colored by tumor of origin. (b-f) t-SNE on the binary activity matrix after applying SCENIC. (b: cells colored by tumor of origin, c: density plot, d: cells colored by the AUC of the MITF regulon, e,f: same for NFATC2 and E2F1). (g) Binary regulon activity matrix for this dataset. The color bar above the heatmap indicates the tumor of origin; regulons associated to the cell cycle (green), invasive (pink) and proliferative (blue) states are zoomed in. (h) Three most dominant networks: MITFhigh, usually known as proliferative, (MITF/STAT); MITFlow, invasive state, (NFATC2/NFIB); and cell cycle (E2F). Confirmed by ChIP-seq: a tick indicates that the regulon presents enrichment of targets in a ChIP-seq dataset for the same transcription factor. (i-j) Aggregation plots for MITF and STAT1 ChIP-seq signal on the predicted target regions of MITF, STAT, and NFATC2, showing high specificity of the predictions. (k) Immunohistochemistry (IHC) on 25 human melanomas using

Figure 1. The SCENIC workflow. (a) In the first step, co-expression modules between transcription factors and candidate target genes are inferred with GENIE3 or GRNboost. Each module consists of a transcription factor together with its predicted targets, purely based on coexpression. (b) In the second step, each co-expressed module is analyzed with RcisTarget to identify enriched motifs; only modules and targets for which the motif of the TF is enriched are retained. Each TF together with its potential direct targets is a regulon. (c) In the third step, the activity of each regulon in each cell is evaluated using AUCell, which calculates the Area Under the recovery Curve. The AUCell scores are used to generate the Regulon Activity Matrix. This matrix can be binarized by setting an AUC threshold for each regulon, which will determine in which cells the regulon is “active”. (d) The Regulon Activity Matrix can be used to cluster the cells (e.g. t-SNE) and, thereby, identify cell types and states based on the shared activity of a regulatory subnetwork.

Figure 2. SCENIC analysis of the mouse brain. (a) AUC histograms for a few key regulons. The AUC allows to split the populations of cells with high versus low activity of a regulon. (b) The clustered regulon activity matrix reveals that the known cell types have distinct regulatory networks. The color bar above the heatmap indicates the cell type assigned by the authors of the dataset. Key regulons (rows) are magnified and colored according to the cell type in which they are active. The two annotation columns on the right indicate whether the TF is known to be relevant (A: for the cell type, manually curated from literature; B: brain-related TFs annotated by MGI), and the main DNA motifs. (c) t-SNE on the binary regulon activity matrix. Each cell is assigned the color of the most active regulons. (d) Microglia gene regulatory network. The regulons associated to microglia can be summarized based on the binding motif of the associated TF (network built in iRegulon). The genes that are included in a previously published microglia signature (Lavin et al. 105) are indicated by a larger font size; the color of the node indicates the number of regulators (lighter:

Figure 3. Cross-species comparison of networks and cell types. (a) Reciprocal activity of human and mouse Dlx1/2 regulons on mouse and human single-cell data. In both cases, the Dlx network is mainly active in interneurons, indicating strong agreement between the species. (b) Shared targets between DLX1/2 regulons inferred from mouse and human. The genes highlighted in red also have associations with Dlx1/2 in GeneMANIA 106 (protein-protein interactions, genetic interactions, co-expression, or literature co-mentioning). (c) Clustering based on GRN activity allows combining cells from human and mouse: The clustering is driven by the cell type, rather than by the organism (unlike the clustering resulting from the normalized expression matrix, Figure S9). Colored TF names show conserved regulons between human and mouse.

Frequently asked questions

Q1. What have the authors contributed in "Scenic: single-cell regulatory network inference and clustering" ?

Here the authors describe a computational resource, called SCENIC ( Single Cell rEgulatory Network Inference and Clustering ), for the simultaneous reconstruction of gene regulatory networks ( GRNs ) and the identification of stable cell states, using single-cell RNA-seq data. Importantly, the authors show that cell state identification based on GRNs is robust towards batch-effects and technical-biases. The authors applied SCENIC to a compendium of single-cell data from the mouse and human brain and demonstrate that the proper combinations of transcription factors, target genes, enhancers, and cell types can be identified. The authors further validated these predictions by showing that two transcription factors are predominantly expressed in early metastatic sentinel lymph nodes. As scalable alternative to GENIE3, the authors also provide GRNboost, paving the way towards the network analysis across millions of single cells. Not peer-reviewed ) is the author/funder. 

Q2. What is the reason for the differences between the tumors?

The apparent differences between the tumors at single-cell level may be due to differences in copy number profiles, which are unique for each tumor and can have a strong impact on the gene expression profile 55,57,75. 

Q3. What are the databases used for the analyses presented in this paper?

The databases used for the analyses presented in this paper are the "18k motif collection" from iRegulon (genebased motif rankings) for human and mouse. 

Q4. How many genes were run on the unlogged matrix?

GiniClust 20 was run on the unlogged TPM matrix with the default parameters, which resulted in a matrix with 17843 genes and one single cluster. 

Q5. What methods have been developed for the analysis of single-cell RNA-seq data?

methods that exploit co-expression or networks for the analysis of single-cell RNA-seq data such as “network synthesis toolkit” 24, Pina’s approach 25, PAGODA 13, and SINCERA 26 have tentatively been developed. 

Q6. What were the non-tumoral cells removed from the expression matrix?

The authors removed these non-tumoral cells from the expression matrix using hierarchical clustering based on the markers cited in the article (mature oligodendrocytes and microglia, respectively). 

Q7. How did the authors test the performance of SCENIC?

To test SCENIC performances the authors applied it to a scRNA-seq data set with well-known cell types from the adult mouse brain previously described in Zeisel et al. 

Q8. What type of interneuron was identified in the human data set?

SCENIC identified an "interneuron-like" and a "excitatory neuron-like" subpopulation within the fetal quiescent cells in the human data set, expressing DLX1,2,5 and MAF, and NEUROD1. 

Q9. How does SCENIC compare to other clustering methods?

In conclusion, SCENIC competes with the best clustering methods to discovering cell types and correctly assigning cells to each cell type; but SCENIC goes beyond existing methods by reducing data dimensionality using TF regulons rather than principal components, thereby accounting for noise and removing technical biases, and uncovering master regulators and gene regulatory networks for each cell type. 

Q10. What is the preferred expression value for the first network-inference step?

note that the first network-inference step is based on co-expression, and some authors recommend avoiding within sample normalizations (i.e. TPM) for this task because they may induce artificial co-variation 82. 

Q11. What are the limitations to using transcription factor motifs to filter and prune co-expression modules?

There are still limitations to using transcription factor motifs to filter and prune co-expression modules, the most obvious being that not for all transcription factors motifs are available, that some factors have motifs with higher information content than others, and that not all transcription factors are co-expressed with their target genes. 

Q12. What is the main cell type in hippocampus and somatosensory cortex?

This data set has been used extensively for benchmarking purposes 13,14,20,27–31 and contains the main cell types in hippocampus and somatosensory cortex, namely neurons (pyramidal excitatory neurons, and interneurons), glia (astrocytes, oligodendrocytes, microglia), and endothelial cells.