Table 1 Overview of methods for estimating regulatory activity from transcriptome data comparing input data, modelling, computational aspects and outcome variables
Method | Input | Model | Computation | Output |
|---|---|---|---|---|
Approach by Schacht et al. | - mRNA expression data - TF binding information | Linear model \( \widehat{g_{i, s}}= c+{\displaystyle \sum_t}{\beta}_t{b}_{t, i}\left({\theta}_{a, t} ac{t}_{t, s}+{\theta}_{g, t}{g}_{t, s}\right) \) with \( a c{t}_{t, s}=\frac{{\displaystyle {\sum}_i}{b}_{t, i}{g}_{i, s}}{{\displaystyle {\sum}_i}{b}_{t, i}},\ {\theta}_{a, t}+{\theta}_{g, t}=1,{\theta}_{a, t},{\theta}_{g, t}\in \left\{0,1\right\} \) | - Optimization criterion: minimize sum of absolute errors - Mixed-integer linear programming - Optimization via Gurobi 5.5 | - parameter for each TF: β t - decision for each TF if θ a,t or θ g,t was chosen |
RACER | - mRNA expression data - copy number variation - DNA methylation - miRNA expression signals - TF binding information - miRNA target site info (c) | Linear models: 1) \( \widehat{g_{i, s}}= c+{\theta}_{CNV, s} C N{V}_{i, s}+{\theta}_{DM, s} D{M}_{i, s}+{\displaystyle \sum_{t\ }}{\beta}_{t, s}\ {b}_{t, i} + {\displaystyle \sum_{mi\ }}{\beta}_{mi, s}\ {c}_{i, mi} miRN{A}_{mi, s} \) 2) \( \widehat{g_{i, s}}=\tilde{c}+{\tilde{\theta}}_{i, CNV} C N{V}_{i, s}+{\tilde{\theta}}_{i, DM} D{M}_{i, s}+{\displaystyle \sum_{t\ }}{\gamma}_{i, t}\ {\beta}_{t, s} + {\displaystyle \sum_{mi\ }}{\gamma}_{i, mi}\ {\beta}_{mi,\mathrm{s}} \) | - Optimization criterion: minimize sum of squared errors with L1 norm penalty on linear coefficients - Elastic-net regularized generalized linear models and LASSO | 1) sample-specific TF and miRNA activities β t,s and β mi,s 2) TF-gene γ i,t and miRNA-gene γ i,mi interactions across all samples |
RABIT | - differential mRNA expression data - somatic mutations - DNA methylation - copy number variation - TF binding info - recognition motifs for RNA-binding protein (RBP) | Linear model: \( \widehat{g_t} = {\displaystyle \sum_f}{\theta}_f{B}_{f, i} + {\displaystyle \sum_t}{\beta}_t{b}_{t, i} \) With B: background factors (gene CNA, promoter DNA methylation, promoter degree promoter CpG content) | - Frisch-Waugh-Lovell method, select subset of significant TFs via model selection procedure and remove TFs with insignificant correlation across tumors | - regulatory activity score for each TF (t value of linear regression coefficient of t-test) |
ISMARA | - gene expression or chromatin state measurements - annotation of promoters (number of predicted sites for motifs) - transcripts and associated promoters - miRNA target site predictions | Linear model \( \widehat{g_{p, s}}={c}_p+{c}_s+{\displaystyle \sum_m}{N}_{p, m}\ {\beta}_{m, s} \) | - Optimization criterion: minimize sum of errors - Bayesian procedure, ridge regression - Gaussian prior for β m,s to avoid overfitting | - inferred motif activity profiles β m,s with set of TFs and miRNAs binding to sites of these motifs (= key regulators) - predicted target promoters, associated transcripts and genes - Network of known interactions between predicted targets and predicted regulatory interactions - enriched ontology categories |
biRte | - mRNA differential expression - miRNA, TF measurements, CNV (optionally) - regulator (R) – target network | Likelihood model: \( {L}_{D,\theta}(R)= p\left( D\Big| R,\theta \right)={\displaystyle \prod_{\widehat{D}} p}\left(\widehat{D}\Big| R,\theta \right) = {\displaystyle \prod_{\widehat{D}}{\displaystyle \prod_c{\displaystyle \prod_i p\left({\widehat{D}}_{i c}\Big|{R}_c,\theta \right)}}} \) | - data specific marginal likelihoods using estimation of hidden state variables with via MCMC - Nested effects model structure Learning to reconstruct transcriptional network | - Estimation of active regulators - Estimation of associated transcriptional network |
ARACNE | - microarray expression profiles | none | - local estimation of pairwise gene expression profile mutual information | - Reconstruction of gene regulatory network |