Summary of algorithms used in the BEELINE framework that were leveraged in our proposed consensus network approach
| Algorithm Name . | Description . |
|---|---|
| GENIE3 [17] | Tree-based ensemble methods such as Random Forest that predict the expression profile of target genes from all other genes. Interaction weights stem from how important an input gene on a target’s expression data. GENIE3 has been demonstrated to be a consistent performer across various datasets [2]. |
| PPCOR [8] | Computes partial and semi-partial correlation coefficients for every pair of genes. Ranks are scaled between -1 and 1, supporting inhibitory and activating network inference. PPCOR results in an undirected network. The negative and positive weights (-1 to 1) are meant to signify an activating or inhibitory interaction. PPCOR was demonstrated to be a consistent performer across various types of datasets [2]. |
| LEAP [7] | Lag-based expression association for pseudotime series (LEAP). Calculates Pearson correlation of normalized mapped read counts; final weightage per edge is the maximum Pearson correlation across all lag values. LEAP also contains a permutation based test that assists in decreasing false discovery rates. LEAP outputs a directed network. |
| SCODE | Implements ODEs for regulatory network representation from gene expression dynamics. Combines linear regression and dimension reduction to improve algorithm efficiency. |
| PIDC [18] | Partial Information Decomposition and Context (PIDC). Computes pairwise mutual information between two genes. From here, PIDC calculates per-gene thresholds that identify the most important interactions for each gene. PIDC outputs an undirected network. |
| SINCERITIES [19] | SINgle CEll Regularized Inference using TIme-stamped Expression profileS (SINCERITIES). Linear regression-based model to recover directed regulatory relationships between genes. SINCERITIES uses Granger casualty, which infers the relationship between change in gene expression of TFs from one window of time and its target genes in another window of time. The edges are inferred through partial correlation analyses. |
| GRNVBEM [20] | GRN Variational Bayesian Expectation-Maximization (GRNVBEM). Implements a Bayesian network model using a first-order autoregressive system to estimate gene fold change at specific times. From here, GRNVBEM uses a Bayesian framework and produces a directed graph with associated signs. |
| SCRIBE [21] | Uses Restricted Directed Information (cRDI) to measure mutual information between the past state and current state of a target gene based on time-stamped single-cell gene expression data. SCRIBE is further made efficient for larger datasets by using a context likelihood of relatedness algorithm, which removes edges that do not correspond to direct effects from a TF and a target. |
| GRNBoost2 [22] | Based on GENIE3 framework, improves efficiency through stochastic Gradient Boosting Machine regression. GRNBOOST2 trains a regression model to infer its edges for each gene in the dataset. GRNBOOST2 has been demonstrated to have consistent performance across various types of datasets [2]. |
| GRISLI [23] | Gene Regulation Inference for Single-cell with Linear differential equations and velocity interference (GRISLI). Estimates cell velocity based on changing gene expression data. From here, GRISLI computes the GRN by solving a sparse regression problem that relates the gene expression of each cell. |
| SINGE [24] | Single-cell Inference of Network using Granger Ensembles (SINGE). Uses kernel-based Granger Causality regression to solve irregularities in time-stamped single-cell genomics data. The inspiration from SINGE came from the fact that pseudotime data for each cell do not take into account the over cell system’s dynamic processes. |
| Algorithm Name . | Description . |
|---|---|
| GENIE3 [17] | Tree-based ensemble methods such as Random Forest that predict the expression profile of target genes from all other genes. Interaction weights stem from how important an input gene on a target’s expression data. GENIE3 has been demonstrated to be a consistent performer across various datasets [2]. |
| PPCOR [8] | Computes partial and semi-partial correlation coefficients for every pair of genes. Ranks are scaled between -1 and 1, supporting inhibitory and activating network inference. PPCOR results in an undirected network. The negative and positive weights (-1 to 1) are meant to signify an activating or inhibitory interaction. PPCOR was demonstrated to be a consistent performer across various types of datasets [2]. |
| LEAP [7] | Lag-based expression association for pseudotime series (LEAP). Calculates Pearson correlation of normalized mapped read counts; final weightage per edge is the maximum Pearson correlation across all lag values. LEAP also contains a permutation based test that assists in decreasing false discovery rates. LEAP outputs a directed network. |
| SCODE | Implements ODEs for regulatory network representation from gene expression dynamics. Combines linear regression and dimension reduction to improve algorithm efficiency. |
| PIDC [18] | Partial Information Decomposition and Context (PIDC). Computes pairwise mutual information between two genes. From here, PIDC calculates per-gene thresholds that identify the most important interactions for each gene. PIDC outputs an undirected network. |
| SINCERITIES [19] | SINgle CEll Regularized Inference using TIme-stamped Expression profileS (SINCERITIES). Linear regression-based model to recover directed regulatory relationships between genes. SINCERITIES uses Granger casualty, which infers the relationship between change in gene expression of TFs from one window of time and its target genes in another window of time. The edges are inferred through partial correlation analyses. |
| GRNVBEM [20] | GRN Variational Bayesian Expectation-Maximization (GRNVBEM). Implements a Bayesian network model using a first-order autoregressive system to estimate gene fold change at specific times. From here, GRNVBEM uses a Bayesian framework and produces a directed graph with associated signs. |
| SCRIBE [21] | Uses Restricted Directed Information (cRDI) to measure mutual information between the past state and current state of a target gene based on time-stamped single-cell gene expression data. SCRIBE is further made efficient for larger datasets by using a context likelihood of relatedness algorithm, which removes edges that do not correspond to direct effects from a TF and a target. |
| GRNBoost2 [22] | Based on GENIE3 framework, improves efficiency through stochastic Gradient Boosting Machine regression. GRNBOOST2 trains a regression model to infer its edges for each gene in the dataset. GRNBOOST2 has been demonstrated to have consistent performance across various types of datasets [2]. |
| GRISLI [23] | Gene Regulation Inference for Single-cell with Linear differential equations and velocity interference (GRISLI). Estimates cell velocity based on changing gene expression data. From here, GRISLI computes the GRN by solving a sparse regression problem that relates the gene expression of each cell. |
| SINGE [24] | Single-cell Inference of Network using Granger Ensembles (SINGE). Uses kernel-based Granger Causality regression to solve irregularities in time-stamped single-cell genomics data. The inspiration from SINGE came from the fact that pseudotime data for each cell do not take into account the over cell system’s dynamic processes. |
Summary of algorithms used in the BEELINE framework that were leveraged in our proposed consensus network approach
| Algorithm Name . | Description . |
|---|---|
| GENIE3 [17] | Tree-based ensemble methods such as Random Forest that predict the expression profile of target genes from all other genes. Interaction weights stem from how important an input gene on a target’s expression data. GENIE3 has been demonstrated to be a consistent performer across various datasets [2]. |
| PPCOR [8] | Computes partial and semi-partial correlation coefficients for every pair of genes. Ranks are scaled between -1 and 1, supporting inhibitory and activating network inference. PPCOR results in an undirected network. The negative and positive weights (-1 to 1) are meant to signify an activating or inhibitory interaction. PPCOR was demonstrated to be a consistent performer across various types of datasets [2]. |
| LEAP [7] | Lag-based expression association for pseudotime series (LEAP). Calculates Pearson correlation of normalized mapped read counts; final weightage per edge is the maximum Pearson correlation across all lag values. LEAP also contains a permutation based test that assists in decreasing false discovery rates. LEAP outputs a directed network. |
| SCODE | Implements ODEs for regulatory network representation from gene expression dynamics. Combines linear regression and dimension reduction to improve algorithm efficiency. |
| PIDC [18] | Partial Information Decomposition and Context (PIDC). Computes pairwise mutual information between two genes. From here, PIDC calculates per-gene thresholds that identify the most important interactions for each gene. PIDC outputs an undirected network. |
| SINCERITIES [19] | SINgle CEll Regularized Inference using TIme-stamped Expression profileS (SINCERITIES). Linear regression-based model to recover directed regulatory relationships between genes. SINCERITIES uses Granger casualty, which infers the relationship between change in gene expression of TFs from one window of time and its target genes in another window of time. The edges are inferred through partial correlation analyses. |
| GRNVBEM [20] | GRN Variational Bayesian Expectation-Maximization (GRNVBEM). Implements a Bayesian network model using a first-order autoregressive system to estimate gene fold change at specific times. From here, GRNVBEM uses a Bayesian framework and produces a directed graph with associated signs. |
| SCRIBE [21] | Uses Restricted Directed Information (cRDI) to measure mutual information between the past state and current state of a target gene based on time-stamped single-cell gene expression data. SCRIBE is further made efficient for larger datasets by using a context likelihood of relatedness algorithm, which removes edges that do not correspond to direct effects from a TF and a target. |
| GRNBoost2 [22] | Based on GENIE3 framework, improves efficiency through stochastic Gradient Boosting Machine regression. GRNBOOST2 trains a regression model to infer its edges for each gene in the dataset. GRNBOOST2 has been demonstrated to have consistent performance across various types of datasets [2]. |
| GRISLI [23] | Gene Regulation Inference for Single-cell with Linear differential equations and velocity interference (GRISLI). Estimates cell velocity based on changing gene expression data. From here, GRISLI computes the GRN by solving a sparse regression problem that relates the gene expression of each cell. |
| SINGE [24] | Single-cell Inference of Network using Granger Ensembles (SINGE). Uses kernel-based Granger Causality regression to solve irregularities in time-stamped single-cell genomics data. The inspiration from SINGE came from the fact that pseudotime data for each cell do not take into account the over cell system’s dynamic processes. |
| Algorithm Name . | Description . |
|---|---|
| GENIE3 [17] | Tree-based ensemble methods such as Random Forest that predict the expression profile of target genes from all other genes. Interaction weights stem from how important an input gene on a target’s expression data. GENIE3 has been demonstrated to be a consistent performer across various datasets [2]. |
| PPCOR [8] | Computes partial and semi-partial correlation coefficients for every pair of genes. Ranks are scaled between -1 and 1, supporting inhibitory and activating network inference. PPCOR results in an undirected network. The negative and positive weights (-1 to 1) are meant to signify an activating or inhibitory interaction. PPCOR was demonstrated to be a consistent performer across various types of datasets [2]. |
| LEAP [7] | Lag-based expression association for pseudotime series (LEAP). Calculates Pearson correlation of normalized mapped read counts; final weightage per edge is the maximum Pearson correlation across all lag values. LEAP also contains a permutation based test that assists in decreasing false discovery rates. LEAP outputs a directed network. |
| SCODE | Implements ODEs for regulatory network representation from gene expression dynamics. Combines linear regression and dimension reduction to improve algorithm efficiency. |
| PIDC [18] | Partial Information Decomposition and Context (PIDC). Computes pairwise mutual information between two genes. From here, PIDC calculates per-gene thresholds that identify the most important interactions for each gene. PIDC outputs an undirected network. |
| SINCERITIES [19] | SINgle CEll Regularized Inference using TIme-stamped Expression profileS (SINCERITIES). Linear regression-based model to recover directed regulatory relationships between genes. SINCERITIES uses Granger casualty, which infers the relationship between change in gene expression of TFs from one window of time and its target genes in another window of time. The edges are inferred through partial correlation analyses. |
| GRNVBEM [20] | GRN Variational Bayesian Expectation-Maximization (GRNVBEM). Implements a Bayesian network model using a first-order autoregressive system to estimate gene fold change at specific times. From here, GRNVBEM uses a Bayesian framework and produces a directed graph with associated signs. |
| SCRIBE [21] | Uses Restricted Directed Information (cRDI) to measure mutual information between the past state and current state of a target gene based on time-stamped single-cell gene expression data. SCRIBE is further made efficient for larger datasets by using a context likelihood of relatedness algorithm, which removes edges that do not correspond to direct effects from a TF and a target. |
| GRNBoost2 [22] | Based on GENIE3 framework, improves efficiency through stochastic Gradient Boosting Machine regression. GRNBOOST2 trains a regression model to infer its edges for each gene in the dataset. GRNBOOST2 has been demonstrated to have consistent performance across various types of datasets [2]. |
| GRISLI [23] | Gene Regulation Inference for Single-cell with Linear differential equations and velocity interference (GRISLI). Estimates cell velocity based on changing gene expression data. From here, GRISLI computes the GRN by solving a sparse regression problem that relates the gene expression of each cell. |
| SINGE [24] | Single-cell Inference of Network using Granger Ensembles (SINGE). Uses kernel-based Granger Causality regression to solve irregularities in time-stamped single-cell genomics data. The inspiration from SINGE came from the fact that pseudotime data for each cell do not take into account the over cell system’s dynamic processes. |
This PDF is available to Subscribers Only
View Article Abstract & Purchase OptionsFor full access to this pdf, sign in to an existing account, or purchase an annual subscription.
