PANDA is a method for estimating bipartite gene regulatory networks (GRNs) consisting of two types of nodes; transcription factors (TFs) and genes. An edge between TF i and gene j indicates that gene j is regulated by TF i. The edge weight represents the strength of evidence for this regulatory relationship obtained by integrating three types of biological data: gene expression data, protein-protein interaction (PPI) data, and transcription factor binding motif (TFBM) data. PANDA is an iterative approach that begins with a seed GRN estimated from TFBMs and uses message passing between data types to refine the seed network to a final GRN that is consistent with the information contained in gene expression, PPI, and TFBM data.
PUMA extends the PANDA framework to model how microRNAs (miRNAs) participate in gene regulatory networks. PUMA networks are bipartite networks that consist of a regulatory layer and a layer of genes being regulated, similar to PANDA networks. While the regulatory layer of PANDA networks consists only of transcription factors (TFs), the regulatory layer of PUMA networks consists of both TFs and miRNAs. A PUMA network is seeded using a combination of input data sources such as motif scans or ChIP-seq data (for TF-gene edges) and an miRNA target prediction tool such as TargetScan or miRanda (for miRNA-gene edges). PUMA uses a message passing framework similar to PANDA to integrate this prior information with gene-gene coexpression and protein-protein interactions to estimate a final regulatory network incorporating miRNAs. Kuijjer and colleagues [7] apply PUMA to 38 GTEx tissues and demonstrate that PUMA can identify important patterns in tissue-specific regulation of genes by miRNA.
OTTER is a GRN inference method based on the idea that observed biological data (PPI data and gene co-expression data) are projections of a bipartite GRN between TFs and genes. Specifically, PPI data represent the projection of the GRN onto the TF-TF space and gene co-expression data represent the projection of the GRN onto the gene-gene space. OTTER reframes the problem of GRN inference as a problem of relaxed graph matching and finds a GRN that has optimal agreement with the observed PPI and coexpression data. The OTTER objective function is tunable in two ways: first, one can prioritize matching the PPI data or the coexpression data more heavily depending on one's confidence in the data source; second, there is a regularization parameter that can be applied to induce sparsity on the estimated GRN. The OTTER objective function can be solved using spectral decomposition techniques and gradient descent; the latter is shown to be closely related to the PANDA message-passing approach (Glass et al. 2013).
SPIDER extends the PANDA framework by incorporating DNase-Seq data to account for chromatin state for the prediction of TF binding sites. The method consists of processing DNase-Seq data to find open chromatin regions and build a “mask” matrix that is then overlaid on the TF-gene motif network to select binding sites that are available fro TF binding. This method can be applied for various biological contexts such as cell lines and tissues. Sonawane and colleagues have employed their method to model cell- type specific GRNs using DNase-Seq data from ENCODE and showed that integrating epigenetic data in SPIDER networks allows building more accurate networks.
EGRET incorporates genetic variants as a fourth data type in the PANDA message-passing framework, enabling the estimation of genotype-specific GRNs. Genetic variants can alter transcription factor binding by affecting the composition of motif sites on the DNA. Not every genetic variant has such an affect; EGRET incorporates only genetic variants which have (1) been shown to be associated with gene expression (expression quantitative trait loci, or eQTL), and (2) are predicted to affect transcription factor binding based on a tool called QBiC (Martin et al. 2019). This information is used in combination with TFBM predictions as input to the PANDA message-passing framework. The resulting EGRET network is a genotype-specific bipartite GRN that is similar to a PANDA network but incorporates the information contained by individual genetic variation.