Fig 1.
Integrative inference of regulatory networks controlling host response to influenza infection.
The main components of our approach are: (1) MERLIN to learn regulatory networks and modules from genome-wide host mRNA profiles from multiple independent virus infections, (2) Multi-Task Group LASSO (MTG-LASSO) approach to predict protein-level regulators (dark blue squares) of mRNA-based modules, (3) construction of active network components for different virus strains and time points, (4) siRNA-based validation of predicted regulators, (5) predicting upstream signaling networks for each module by identifying minimal physical subnetworks that connect module regulators (dark blue squares and light blue ellipses) with a small number of intermediate nodes (yellow diamonds).
Fig 2.
Overview of human influenza response module expression patterns.
Shown are 41 human Calu-3 modules with at least 10 genes. The red-blue heat map shows mean expression of all genes in each module. The more red an entry the higher the expression in infected versus mock, while the more blue the entry, the higher the expression in mock compared to the infected sample. Under “Viruses and time points (h)”, each virus treatment's time series is shown separately, and viruses are ordered from low to high pathogenicity. Low-pathogenicity samples labeled H1N1 used A/CA/04/09 H1N1 unless labeled '(NL)' for A/Netherlands/602/09. Samples labeled H5N1 used A/VN/1203/04 H5N1 and laboratory mutants thereof. The “Genes” column shows the size of each module; larger values are shown in darker blue. Under “Enrichment”, a red box indicates module enrichment with any MSigDB motif, any MSigDB gene set, any Gene ontology process, any influenza screen set, or any immune response gene set (Methods).
Fig 3.
Overview of human influenza response modules and top predicted regulators.
A. A selection of sixteen human modules (grey, center), with their top predicted regulators (shapes on left, connected to modules), and enriched gene sets relevant to immune response or viral life cycle (colored boxes, right). Modules were included in this figure if they were enriched for one of the following categories: (a) influenza host genes (pink) identified by RNAi, protein-protein interactions, expert curation (Zhang et al. 2009), or enrichment for other virus life cycle gene sets (from GO or KEGG); (b) motifs (green) for transcriptional regulators whose annotations indicate relevance to the innate immune response; (c) innate immune response gene sets (blue) from experimental results or manual curation. One module, Cluster 1593, is also included because its consensus regulator set (but not targets) was enriched for host proteins that are involved in the influenza life cycle from Watanabe et al (2014). Also shown are seventeen (ellipses and diamonds) predicted mRNA-level regulators, and ten predicted protein regulators (hexagons), that are associated with these modules. Regulators with evidence for influenza relevance are shaded in pink. Host genes that significantly impact viral replication identified by this study (APP, FGFR3, HCLS1, HOXA7, SERPINA3) are indicated with yellow stars. B. Comparison of the MERLIN inferred mouse and human influenza response networks ("Influenza (this study, other system)") to each other and to other regulatory networks. Fold enrichment of shared edges over expected fraction is visualized in the heat map, with significant comparisons marked with asterisks (hypergeometric p-value < 0.05).
Table 1.
Catalog of MERLIN modules for human and mouse systems.
Fig 4.
Conservation between human and mouse subnetworks.
A. Modules conserved between human and mouse. There were fifteen (red shading) significantly overlapping module pairs (hypergeometric p-value < 0.0001), including 12 human modules and 13 mouse modules. B. A core set of interactions conserved between the human and mouse consensus regulatory networks. Shown are the 48 of 96 shared interactions that belong to subnetworks with at least three nodes. Shading indicates Calu-3 ISGs (violet) and nodes identified both as known host genes and ISGs (magenta). Star indicates a gene that significantly impacted viral replication on knockdown as identified in this study.
Table 2.
Summary of results of siRNA validation study of twenty regulators predicted by MERLIN.
Fig 5.
Multi-task Group LASSO (MTG-LASSO) structured-sparsity approach for integration of protein data with expression-based regulatory module networks.
A. Illustration of MTG-LASSO framework for predicting protein regulators for one module (Methods). Horizontal separations in the X and Y data boxes represent different virus time course. Rows of X and Y represent time points. Columns of X correspond to proteins and columns of Y correspond to mRNA levels of genes in a module. B. Comparison of number of nonzero regression weights identified by MTG-LASSO and LASSO. Each dot (MTG-LASSO) or plus sign (LASSO) represents the number of non-zero regression weights for one setting of λ (the sparsity parameter) for one module. Number of nonzero weights is averaged over 10 folds of cross-validation. C. Comparison of cross-validation predictive quality between MTG-LASSO and LASSO. Results are shown for λ = {0.10, 0.75}; results for other settings are in S4 Fig. In each scatterplot, there is one point per module. Left two plots compare methods based on Pearson correlation (ρ) of predicted to actual expression values; right plots compare on the basis of root mean squared error (RMSE). Inset ρ gives Pearson correlation between MTG-LASSO and LASSO scores. Diagonal line is shown for comparison. D/E. Examples of curves used to select λ for individual modules; human (D) and mouse (E). Y-axis gives Pearson correlation (cross-validation predictive quality); X-axis gives the average number of nonzero regression weights for that module; this value is higher than the final number of high-confidence, high-weight regulators. Stars indicate the chosen value of λ for the example modules.
Table 3.
Consensus human protein regulators identified by MTG-LASSO.
Table 4.
Top consensus mouse protein regulators identified by MTG-LASSO.
Table 5.
Summary of results of siRNA validation study of seven human protein regulators predicted by MTG-LASSO.
Fig 6.
Time-point and virus-specific active subnetwork clusters.
A. Clusters identified using hierarchical clustering of edges based on the presence/absence pattern of edges across time and virus strain. Each row represents an edge. All clusters are uniformly scaled to the same height, but comprise varying numbers of edges (shown on left). B. Active subnetworks defined by clusters in A for each virus. The subnetwork for each virus is defined by taking the union of all edges present at any time point for that virus. Dark nodes/edges are active in both the cluster and any time-point for the virus; light nodes/edges are active in all other viruses and part of this cluster.
Table 6.
Characterization of time- and virus-specific subnetworks.
Fig 7.
Integration of expression- and protein-based regulators into a physical subnetwork for module 1549.
A. Candidate subnetwork for module 1549 given as input to the Integer Linear Programming (ILP) approach to find a minimal network. The candidate subnetwork is extracted from the background interaction network by connecting MERLIN (mRNA) and MTG-LASSO (protein) regulators through at most one intermediate node and mRNA signaling proteins to TFs (Materials and Methods). Nodes with dark outlines are input to the ILP-based subnetwork inference method: MERLIN regulators (diamonds and ellipses), protein regulators (octagons), and host genes for influenza identified by Watanabe et al (2014) (boxes). Nodes without borders are candidate intermediates between two regulators and were not given special treatment by the method. Node color indicates host genes involved in influenza obtained from siRNA or protein interaction studies and Calu-3 ISGs (as in Fig 3); this information is for visualization purposes and is not used in the subnetwork inference method. Only edges between intermediate nodes and input nodes are shown. Additional available protein-protein interactions between intermediate nodes are not shown for visual clarity. B. Minimal subnetwork for Cluster 1549's regulators inferred using ILP. Yellow stars indicate host genes for which siRNA knockdown significantly impacted viral replication. Also shown are interactions between Watanabe host genes and viral proteins (yellow hexagons). Grey edges are input to the method but not selected by the approach. All selected edges and all nodes other than YWHAZ, TP53, SNW1 (brown boxes) had FDR ≤ 0.10. FDR was assessed based on permutation tests.
Fig 8.
Target and regulator expression profile of human Module 1549.
In the “Regs” columns, consensus regulators for each gene are marked with purple boxes. In “Motifs”, genes containing MSigDB regulatory motifs, including miRNA motifs, are marked in green. Motifs shown are only those that are enriched in the module. Next, colored boxes indicate host genes identified from screening studies (salmon) and immune response gene sets (violet). Below the rows for module genes are rows for MERLIN-predicted mRNA regulators and MTGLASSO-predicted protein regulators. Gene expression values are scaled (-2,2); protein values are scaled (-1,1) to improve visibility. Time courses entirely missing at the protein level are indicated as “No protein data”.
Fig 9.
Target and regulator expression profile and the physical regulatory program of Module 1540.
A. Heatmap visualization of module 1540 genes and regulators. In the “Regs” columns, consensus MERLIN regulators for each gene are marked with purple boxes. Next, genes with SREBP binding motif are marked in green. Below the heatmap for module genes are rows for MERLIN-predicted mRNA regulators and MTGLASSO-predicted protein regulators. Gene expression values are scaled (-2,2); protein values are scaled (-1,1) to improve visibility. B. Original subnetwork for Cluster 1540. Black edges link regulators to intermediates; these are provided as input to the ILP method. Additional interactions from the background network between regulators and intermediates are shown in grey. See Fig 7B for additional legend details. C. Minimal subnetwork for Cluster 1540, output from the ILP-based subnetwork inference method. Compared to the original subnetwork, the minimal subnetwork prunes away one protein: HIST1H3A. See Fig 7B for legend.