Comparing signaling networks between normal and transformed hepatocytes using discrete logical models

Abstract

Substantial effort in recent years has been devoted to constructing and analyzing large-scale gene and protein networks on the basis of "omic" data and literature mining. These interaction graphs provide valuable insight into the topologies of complex biological networks but are rarely context specific and cannot be used to predict the responses of cell signaling proteins to specific ligands or drugs. Conversely, traditional approaches to analyzing cell signaling are narrow in scope and cannot easily make use of network-level data. Here, we combine network analysis and functional experimentation by using a hybrid approach in which graphs are converted into simple mathematical models that can be trained against biochemical data. Specifically, we created Boolean logic models of immediate-early signaling in liver cells by training a literature-based prior knowledge network against biochemical data obtained from primary human hepatocytes and 4 hepatocellular carcinoma cell lines exposed to combinations of cytokines and small-molecule kinase inhibitors. Distinct families of models were recovered for each cell type, and these families clustered topologically into normal and diseased sets.

Figure 1. Experimental design and primary data

(a) Outline of the experimental approach showing five ligand cues and their receptors, 16 intracellular “signals” assayed using phospho-specific antibodies and xMAP technology (the relevant phosphorylated residues are indicated, as are the likely upstream kinases) and three small molecule kinase inhibitors (b) Dataset from human hepatocytes. Rows represent intracellular signals assayed immediately prior to ligand addition and 25 minutes thereafter, and columns represent different ligand combinations (see legend). For each combination of ligands, one of three kinase inhibitors was applied, as well as all possible combinations of them. Data are color-coded to highlight induction (relative to basal activity). (c) Data from HepG2 cells following the same design as in (a). The intensity of background color for each box (yellow to red) indicates how the proportion of tumor cell lines that also responded to a particular drug-ligand combination; complete data can be found in Fig. S1. Data was processed using DataRail software (18).

Figure 2. General workflow and goodness of fit of the cell-specific models

Schematic of workflow of model assembly, training, validation and extension. A PKN derived from interaction data is imported into CellNOpt and converted into an assembly of all possible logical interactions. The assembly is then trained against experimental data, generating cell-specific models having lower MSE error. In general, this involves removing interactions present in the PKN that are not supported by the data. Interactions absent from the PKN can then be added to see if they reduce residual error (green line). Once a satisfactory model is found, is is analyzed to identify differences among cell lines.

Figure 3. Residual error (MSE deviation, see methods) of the models

(a) MSE for the ensemble of all possible models compatible with the PKN and pairwise combinations of trained, cell-specific models with various data sets. The magnitude of the error is mapped to the intensity of the background (with red highest). Trained models fit corresponding data best, but all trained models exhibit lower error with all data sets than the starting ensemble. (b) Average MSE error of models across all cell types for specific experimental conditions and biochemical assays. For a single experiment, the error varies between 0 and 1; the maximal error was observed for phospho-STAT3 in cells treated with IL6 (MSE ~0.4).

Figure 4. Inferred models of immediate-early signaling downstream in primary and transformed hepatocytes

A graph representing nodes and edges present in a set of trained Boolean models deviating by 1% in MSE and considered to be indistinguishable. The graph was created by exporting a GraphML file from from ProMoT (20) and modifying it with GraphViz (48). Line thickness denotes the frequency OR-gated edges found in trained models of primary hepatocytes (blue) or an average of the four HCC models (red); solid black edges were found in both primary and HCC cell lines. Thin grey edges were absent from all models, and dashed grey edges removed in the network preprocessing step (1). Green hexagons denote stimuli, red rectangles targets of kinase inhibitors, and blue rectangles readouts. Rectangles with magenta fill were both measured and subjected to inhibition using small molecule drugs. Rectangles with white fill were compressed during graph processing. An AND gate is present upstream of IKK only in models of primary hepatocytes (see text), no other AND gate was consistently identified. The dashed black line denotes an interaction missing from the starting PKN but whose inclusion reduced the MSE of both hepatocyte and HCC models. This interaction implies inhibition of Stat3 phosphorylation by IKK in IL6-stimulated cells (see text).

Figure 5. Topological clustering of logic-based models of primary and transformed hepatocytes

(a) Distances between primary hepatocytes and the four HCC cell lines (Huh7, Focus, Hep3B, and HepG2) and a ‘average HCC cell line’ (AvgHCC) based on the mean difference in the topologies of Boolean models as defined by the distance d² (see methods) and plotted using the edge-weighted spring embedded algorithm in Cytoscape (49). The full range of distances is listed in the supplementary materials. (b) Hierarchical clustering of primary and transformed hepatocytes based on the logical models, and a comparison to previous published clusters based on gene expression (the dendogram was manually redrawn from reference (26) and distances are approximate). A classification of cell lines as epithelial or mesenchymal (25) is also shown (see Fig. S8).

Figure 6. Key differences uncovered by logic-based modeling

Schematic of inferred differences between HCC cell lines and hepatocytes. The callouts show data associated with each topological difference with data from primary hepatocytes as blue bars and HCC as overlaid red bars. Fold-increase in actual values are shown, except in the case of IκB, where normalized data are compared to predictions from Boolean models (for clarity). Numbered green circles label differences between primary and HCC, cells.

Figure 7. Probing the biochemistry of an inferred IKK-Stat3 interaction

For in vivo experiments, cells were pretreated with 2 to 20 µM TPCA-1or BMS345541. The concentration range of TPCA-1 was 10 fold lower than BMS345541 to match published IC50 values for inhibition of Iκb phosphorylation. Cells were pretreated with drug for one hour and then stimulated with IL6; the phosphorylation of Stat3Y705 was measured 25 minutes thereafter. In vitro experiments were performed with recombinant kinases and phosphorylation levels measured using a fluorescence assay.