Table 1 Summary of single-cell network modeling approaches
CategoryExample methodsUnderlying biological assumptionAlgorithmic basisAdvantagesLimitations
Dynamic network (extensively reviewed in refs [5355])SCNS [81]Single-gene changes between cell transition states can inform on gene regulatory relationsBooleanDoes not rely on prior knowledge. Has a web UI. Resulting models are executable and can be used to make predictionsNeed data discretization; limit to small numbers of genes; regulatory relations need to follow Boolean rules
SCODE [82]TF expression dynamics (pseudo-time) and TF regulatory relations (GENEI3)ODE; Bayesian model selectionEstimate relational expression efficiently using linear regression; reduction of time complexity; fast algorithmNeed dimension reduction first for computing speed and memory feasibility; assumes that all cells are on the same trajectory; optimization is computationally intractable
GRISLI [83]Variability in scRNAseq data caused by cell cycle, states, etc. allows the inference of pseudo-time associated with each individual cellODEMakes no restrictive assumption on the gene network structure; can consider multiple trajectories; fast algorithmHas to estimate the velocity of each individual cell using information from neighbors
SINCERITIES [84]Changes in the expression of a TF will alter the expression of target genesRidge regression and partial correlation analysisLow computational complexity and able to handle large-scale dataRequires scRNAseq data at multiple time points. Restricted to TFs and their targets to infer edges
Scribe [85]Cell ordering can be improved with time-series or cell velocity estimationsRDIOutperforms other pseudo-time methods given time-series data. Can be applied to any data type if the data structure is appropriateRequires time-ordered gene expression profiles or velocity estimation from introns and exons
AR1MA1-VBEM [40]The cell differentiation process or response to external stimulus reveals the hierarchical structure of the transcriptomeFirst-order autoregressive moving-average and variational Bayesian expectation-maximizationWeighted interactions between genes along psuedotime. Model used accounts for noisy dataData are expressed as fold changes between timepoints/conditions or scaled by housekeeping genes
SCINGE [86]Learned target regulator genes can be used to assign each cell to their progress along a trajectoryGranger causalitySmooths irregular pseudo-times and missing expression valuesNear random performance for predicting targets of individual regulators
SoptSC [87]Similarities between whole transcriptomes of single cells can be used to order themCells ordered by minimum paths on weighted cluster-to-cluster graph derived from cell similarity matrixIncludes comprehensive single-cell workflow; leverages information from other parts of the workflow to improve performanceCannot be run with other tools, have run the full workflow to get pseudo-time inference
Within-cell or cell population networkSCENIC [88]TF target-based regulationCombining TF regulatory relations (GENIE3) with TF-binding motif analysisRobust against dropouts, get a TF score for individual cells (no averaging of cells).Limited to TF-based relations
Pina et al. [89]TFs drive lineage commitmentOdds ratio for on/off gene associations and spearmen correlation for expression levels associationsRobust to dropoutsBased on single-cell multiplex qRT-PCR, may be difficult to extend the method to sparse single-cell data (selected 44 genes to test)
Iacono et al. [90]Coexpression is regulated by TFs, cofactors, and signaling molecules which can be captured with gene–gene correlationsPearson correlation using z-score-transformed countsCan compute correlations at the single-cell level and it is robust to dropouts and noise inherent to single-cell dataNetworks are very dense (some have millions of significant edges)
PIDC [39,91]Gene regulatory information reflected in dependencies in the expression patterns of genesPartial information decomposition using gene triosCompared with correlation, captures more complicated gene dependenciesNetworks are influenced by data discretization, choice of mutual information estimator, method developed for sc-qPCR data, may not be extendable to higher throughput and sparser scRNAseq data
Jackson et al. [92]Deletion of TFs combined with experimental conditions allows for the inference of gene relationshipsMTL to leverage cross-dataset commonalities and incorporate prior knowledgeDoes not require sophisticated normalization of single-cell data or imputation. Able to combine multiple conditions/datasets for more accurate inference. TF deletions give strong causal link to affected genesRequires single-cell data with TF deletions and/or environmental perturbations
Wang et al. [93]Gene perturbations allow for inference of causal relationshipsScoring of conditional independence test to identify optimal DAGGives causal relationships between genesRequires interventional data. No loops allowed in DAG
ACTION [94]Functional identity of cells is determined by a weak, but specifically expressed set of genes which are mediated by TFsKernel-based cell similarity and geometric approach to identify primary functionsRobust to dropout and does not require averaging. Identifies functions unique to cell typesRequires TFs and their targets. Only provides TF-driven networks
SINCERA [95]TF target-based regulationFirst-order conditional dependence on gene expression to construct a DAGKey TFs identified using multiple importance metricsOnly considers TFs and their targets. Requires genes/TFs to be DEGs or expressed in >80% of cells
Cell–cell communication networkiTALK [96]Ligand–receptor interactionsThreshold ranked list of genes from two cell types for ligand–receptor pairsAllows for the inference of directionality of interactionRequires curation of ligand–receptor interactions (not all interactions are known). Average expression at the cell-type level (no longer single cell). Cannot reveal novel interactions beyond known ligand–receptor knowledge
Zhou et al. [97]Ligand–receptor interactionsExpression of ligand and corresponding receptor more than three standard deviations greater than the meanAllows for the inference of directionality of interactionRequires curation of ligand–receptor interactions (not all interactions are known). Average expression at the cell-type level (no longer single cell)
Kumar et al. [98]Ligand–receptor interactionsProduct of the average expression of ligand and corresponding receptorAllows for the inference of directionality of interaction. Interaction score gives the strength of interaction (rather than just significance)Requires curation of ligand–receptor interactions (not all interactions are known). Average expression at the cell-type level (no longer single cell)
Arneson et al. [99]Ligand to downstream signalingCoexpression of ligand genes in source cells with other genes in target cellsUse secreted ligands as a guidance for directional inference between cell populationsGene expression is summarized to the cell population level and coexpression is at the sample level, requiring large sample sizes
SoptSC [87]Ligand–receptor interactionsLikelihood estimate of the interaction between two cells based on expression of the ligand, receptor, and downstream pathway target genes (including expression direction). Consensus signaling network derived from all cells in each clusterIncorporates target genes of pathways and their directionality. Computes interaction likelihood at the single-cell level and summarizes across all cells in the cluster for higher confidenceRequires curation of ligand–receptor interactions and their downstream pathways
scTensor [100]Ligand–receptor interactionsTensor decomposition with cell–cell interactions as hypergraphsAllows L–R pairs to function across multiple cell-type pairs (not restricted to a single-cell-type pair), which is more reflective of underlying biologyRequires curation of ligand–receptor interactions. Averages single cells to the cell-type level