Central Role for MCP-1/CCL2 in Injury-Induced Inflammation Revealed by In Vitro, In Silico, and Clinical Studies

The translation of in vitro findings to clinical outcomes is often elusive. Trauma/hemorrhagic shock (T/HS) results in hepatic hypoxia that drives inflammation. We hypothesize that in silico methods would help bridge in vitro hepatocyte data and clinical T/HS, in which the liver is a primary site of inflammation. Primary mouse hepatocytes were cultured under hypoxia (1% O2) or normoxia (21% O2) for 1–72 h, and both the cell supernatants and protein lysates were assayed for 18 inflammatory mediators by Luminex™ technology. Statistical analysis and data-driven modeling were employed to characterize the main components of the cellular response. Statistical analyses, hierarchical and k-means clustering, Principal Component Analysis, and Dynamic Network Analysis suggested MCP-1/CCL2 and IL-1α as central coordinators of hepatocyte-mediated inflammation in C57BL/6 mouse hepatocytes. Hepatocytes from MCP-1-null mice had altered dynamic inflammatory networks. Circulating MCP-1 levels segregated human T/HS survivors from non-survivors. Furthermore, T/HS survivors with elevated early levels of plasma MCP-1 post-injury had longer total lengths of stay, longer intensive care unit lengths of stay, and prolonged requirement for mechanical ventilation vs. those with low plasma MCP-1. This study identifies MCP-1 as a main driver of the response of hepatocytes in vitro and as a biomarker for clinical outcomes in T/HS, and suggests an experimental and computational framework for discovery of novel clinical biomarkers in inflammatory diseases.


Introduction
Among many other functions, the liver plays a critical role in inflammation and innate immunity, processes that are controlled by multiple cell types including hepatocytes, Kupffer cells, and other non-parenchymal cells. Although at least 15 different cell types can be found in normal liver [1], hepatocytes constitute the largest pool of parenchymal cells, comprising approximately 60-80% of the total liver cells [1,2]. Inflammatory conditions such as ischemia/reperfusion (I/R) and post-trauma hemorrhagic shock (T/HS) are associated with liver hypoxia [3,4]. It is now accepted that hypoxia is not merely an outcome of the inflammatory response, but rather is a key driver of the development of inflammation through the regulation of O 2 -dependent signal transduction and gene expression [5,6].
Mathematical and computational (in silico) methods have emerged as adjuncts to in vitro and in vivo studies of acute inflammation [7]. For example, we have recently applied both mechanistic and data-driven computational modeling to help define the dynamic, multi-dimensional inflammatory response to T/HS in vivo [8][9][10][11][12]. The goal of the present study was to determine if combined in vitro/in silico studies could help elucidate key hepatic inflammatory mediators relevant to human T/HS. This study identifies the chemokine Monocyte Chemoattractant Protein-1 (MCP-1/CCL2) as a main driver of the response of hepatocytes in vitro and as a biomarker for organ damage in clinical settings of T/HS, and, more generally, suggests a pathway for combined experimental and computational studies to facilitate the discovery of novel clinical biomarkers of inflammation.

MCP-1 is a central component of the dynamic, multidimensional response of hepatocytes to cell stress
To assess the response of hepatocytes to hypoxia, primary wildtype mouse hepatocytes were subjected to 1% O 2 for 1-72 h, and 18 mouse cytokines were measured in both the supernatant and whole-cell lysate. Hepatocytes cultured under normoxic (21% O 2 ) conditions served as controls. One-way ANOVA showed that, in normoxic hepatocytes, MCP-1, KC, and IP-10 (in lysates) and MCP-1, KC, and MIG (in supernatants) were altered significantly ( Table 1). In hypoxic hepatocytes, the significantly altered mediators were MCP-1, MIG, IL-1a, IL-1b, IL-10, and IL-13 (in lysates) and MCP-1, IP-10, IL-1a, and VEGF (in supernatants). Thus, MCP-1 was the only mediator that exhibited significant changes in all four conditions examined, as shown in Fig. 1A.
Hierarchical Clustering and Principal Component Analysis reveal key differences in inflammatory mediator production/release in the hepatocyte response to cell stress In order to address the question of which groups of mediators exhibit similar dynamics in normoxic and hypoxic hepatocytes, we performed hierarchical clustering on fold changes for each mediator at each pair of consecutive time points. We examined both raw values and fold change values for all mediators over each time interval. Each inflammatory mediator's values over time made up its dynamic pattern of expression and release. To identify those inflammatory mediators that showed similar production or secretion behavior, these patterns were then compared and grouped using hierarchical clustering, with inflammatory mediators in the resulting dendrogram ordered according to their membership in the clusters that emerged (Fig. 1B).
We also sought to determine the subset of mediators that evolve dynamically in a manner that is most characteristic of the response to normoxia or hypoxia, respectively. Inflammatory mediators that have similar patterns of covariance have similar loadings onto principal components, and so would be expected to project in a similar direction in principal component space. We therefore used the angle of an inflammatory mediator's projection into principal component space as the basis of a distance metric in k-means clustering (Figures S1, S2, S3, S4). This analysis suggested that, despite an overlap in common mediators, k-means clustering could differentiate normoxia from hypoxia and lysates from supernatants. Furthermore, MCP-1 and KC were generally grouped together and were the predominant mediators along the first principal component.
A predominant role for MCP-1, KC, and IL-1a in the hepatocyte response to stress inferred from Dynamic Network Analysis We sought to determine the consensus of the three aforementioned, independent clustering experiments. If the cluster memberships were similar, then those mediators that did not segregate consistently across different clusterings could be treated as noise. In contrast, those inflammatory mediators that consistently segregated together were considered as mediators exhibiting similar or correlated dynamic trajectories. We first calculated an Adjusted Rand Index (ARI) [13][14][15][16] to quantify the degree to which any two clustering results were in agreement. This comparison (see Figures S9, S10, S11, S12) resulted in a ''consensus clustering'' showing the inflammatory mediators that exhibited the same segregation pattern across all analysis methods, suggesting that MCP-1 and KC segregate together across all experimental conditions (Fig. 1C).
This meta-clustering analysis suggested groups of inflammatory mediators with similar dynamic patterns. We next performed Dynamic Network Analysis (DyNA) [12] to suggest mediators that appear to be ''hubs'' (or most connected to other mediators). (Figs. 2A-B; Figure S17). The total number of dynamic network connections is shown in Fig. 2C. Based on DyNA, MCP-1 and KC were the most connected mediators across all time points in normoxia supernatants, and at 24-72 h in normoxia lysates ( Fig. 2A). In hypoxia, the most connected inflammatory mediators were IL-1a and VEGF in supernatants and IL-1a and MIP-1a in lysates (Fig. 2B). Overall, the most connected inflammatory mediators were MCP-1 and IL-1a, with MCP-1 being the only highly-connected node present in all experimental conditions examined (Fig. 2C).
Dynamic Network Analysis suggests altered networks of inflammatory mediators in MCP-1 2/2 hepatocytes We next carried out DyNA of lysates and supernatants from normoxic and hypoxic MCP-1 2/2 hepatocytes (Figs. 4A-C; Figure S18). This analysis suggested a significant decrease in the number of connections (Figs. 4A-B) in both lysates and supernatants of normoxic cells, and in hypoxia supernatants, as compared to wild-type hepatocytes (see Figs. 2A-B). In supernatants of normoxic MCP-1 2/2 hepatocytes, KC was the most significantly changed mediator (though with nearly zero connections across all time points; Fig. 4A), whereas in supernatants of hypoxic MCP-1 2/2 hepatocytes this chemokine was a central node with one connection only at 24-72 h (Fig. 4B). In lysates from normoxic MCP-1 2/2 hepatocytes, IL-6 was the most significantly changed mediator, but with zero connections (Fig. 4A). However, this cytokine was the most connected mediator (along with IFN-c) in hypoxia lysates at all-time intervals after 6 h (Fig. 4B). DyNA suggested a clear reduction in the total number of dynamic connections among inflammatory mediators in MCP-1 2/2 cells, with IL-6 being the most connected mediator (but only in hypoxia lysates; Fig. 4C).

IL-6 expression and release is reduced in MCP-1 2/2 hepatocytes
Because one-way ANOVA and DyNA revealed that the most significant changes in inflammatory mediator production and release were observed for MCP-1 and IL-6, we hypothesized that MCP-1 would affect IL-6 expression and release in stressed hepatocytes. IL-6 levels were significantly lower in both normoxia lysates and supernatants of MCP-1 2/2 hepatocytes as compared to normoxic (Fig. 5A-B) and hypoxic ( Fig. 5C-D) wild-type controls. These results suggest that the differential expression and release of IL-6 in mouse hepatocytes depends, at least in part, on MCP-1.
In situ hepatocyte expression of MCP-1 and IL-6 is attenuated in MCP-1 2/2 cells The differential expression of MCP-1 and IL-6 in both wildtype and MCP-1 2/2 cells was confirmed using confocal immunofluorescence (Fig. 6A). Quantitative analysis of the images (Fig. 6B) revealed that MCP-1 is indeed elevated in wild-type hepatocytes as compared to MCP-1 2/2 cells, with much higher levels in normoxia vs. hypoxia, confirming the results obtained by Luminex measurements (see Fig. 1A). Similarly, cellular IL-6 levels were lower in the MCP-1 2/2 cells as compared to wild-type hepatocytes, especially under hypoxic conditions (Fig. 6B).

Elevated plasma MCP-1 levels as biomarker for mortality and morbidity in human trauma/hemorrhage
Given the central role of the liver in the in vivo response to T/HS [17], we hypothesized that plasma MCP-1 levels would serve as an indicator of mortality in human T/HS. Indeed, an analysis of a cohort of 30 trauma patients revealed that patients could be segregated between survivors and non-survivors based on plasma MCP-1 levels within the first 24 h (Fig. 7A). Similar segregation could be made using plasma IL-6 levels for the same groups of patients: survivors (193.4636.7 pg/ml) vs. non survivors (1050.16522.9 pg/ml) (P,0.001, analyzed by Mann-Whitney U test).
MCP-1 also appeared to coordinate stress-induced production of mediators including IL-6 (see above), and IL-6 has been identified consistently as a marker of negative outcomes in T/HS [18,19]. Accordingly, we next evaluated the potential of plasma MCP-1 levels as a biomarker for morbidity in a separate cohort of survivors of blunt trauma. A cohort of 31 hypotensive patients was matched with 31 normotensive patients (all survivors). Plotting IL-6 vs. MCP-1 plasma concentrations for the trauma patient cohort as a whole (n = 62) suggested that circulating IL-6 levels in these patients could be segregated into two groups based on circulating MCP-1 levels: low MCP-1 (,1000 pg/ml) and high MCP-1 (.1500 pg/ml, Figs. 7B-C). Furthermore, analysis of the data showed a positive correlation between MCP-1 and IL-6 in both normotensive (Pearson Product Moment correlation coefficient = 0.59, Fig. 7B) and hypotensive (0.41, Fig. 7C) blunt trauma patients. The P values associated with these correlation coefficients were highly significant (5.16610 217 and 0.000047 for normotensive and hypotensive blunt trauma patients, respectively).
We then performed a different grouping irrespective of blood pressure status, based solely on MCP-1 concentrations (,1000 pg/ml vs. .1500 pg/ml). This grouping reduced the study cohort to 46 T/HS patients. T/HS patients with lower circulating MCP-1 levels ( Fig. 7D;   Patients with higher circulating MCP-1 levels had worse outcomes, as indicated by a longer total length of stay (LOS), longer intensive care unit (ICU) LOS, and greater requirement for mechanical ventilation (Fig. 7F). Interestingly, statistical analysis using two-way ANOVA showed no significant differences between MCP-1 levels in the initial cohort of normotensive vs. hypotensive patients (n = 31 each, Figs. 8A-B), and analysis of the clinical outcomes of the patients segregated based on their normotensive/ hypotensive status revealed no differences between normotensive (n = 27) and hypotensive patients (n = 19) (Fig. 8C).

Discussion
Liver ischemia and subsequent reperfusion, which can occur in T/HS and resuscitation, lead to inflammatory liver injury as well as to the systemic elaboration of inflammatory mediators and subsequent multiple organ dysfunction [17,20,21]. Liver hypoxia is a primary feature of HS in experimental models [3,4]. Hypoxic stress, however, is but one of a variety of stresses (e.g. endoplasmic reticulum stress and autophagy) encountered by hepatocytes in settings such as T/HS [22] and I/R [23]. We therefore reasoned that the analysis of synthesis and secretion of inflammatory mediators by stressed hepatocytes would help in our understanding of the pathophysiology of T/HS and other conditions characterized, at least in part, by hepatic stress.
Our key finding is that a combined in vitro/in silico analysis of the dynamics of synthesis and secretion of multiple inflammatory mediators points to the chemokine MCP-1 as a central coordinator of the inflammatory response of stressed hepatocytes. Pursuing this in vitro/in silico-derived hypothesis allowed us to segregate blunt trauma patients based on circulating levels of MCP-1. Importantly, this segregation of patients would have proven ineffective if performed based solely on the patients' normotensive/hypotensive status post-injury.
Chemokines are small cytokine-like, heparin-binding molecules help orchestrate immune and stem cell infiltration into the liver in response to both acute and chronic injuries [20,24]. Hepatic chemokines are released during inflammatory injury (infection, drug intoxication), and play a key role in the activation and proliferation of liver-resident cells (including hepatocytes, stellate cells, and endothelial cells). In our studies, multiple chemokines were synthesized with diverse time courses by stressed hepatocytes. Though chemokines are often thought to be highly redundant in their actions, MCP-1 was the only mediator that exhibited significant changes in both normoxia and hypoxia, and in both cell lysates (denoting cellular expression) and supernatants (denoting secretion by cells). Our findings of perturbed dynamic inflammatory networks in MCP-1 2/2 mice supports the hypothesis that this chemokine plays a central, driving role in the hepatocyte response to stress. The apparently incomplete elimination of MCP-1 (or presence of a protein highly-homologous to MCP-1) in MCP-1 2/2 cells raises the possibility that the actual effects of MCP-1 on coordinating inflammation might be even more pronounced than those we observed in these studies.
MCP-1 (CCL2) was first identified as a monocyte-specific chemoattractant, and was later shown to attract memory T lymphocytes and NK cells. Binding to its seven-transmembrane Gprotein-coupled receptor CCR2 (CC chemokine receptor 2) results in signaling events that lead to the recruitment of monocytes in a variety of in vivo experimental models of infection and injury [25,26]. In addition to recruiting monocytes, MCP-1 can also activate macrophages and endothelial cells and has been implicated in the attraction of neutrophils and the generation of neutrophil-dependent tissue damage [26]. Lu et al. showed that despite normal numbers of circulating leukocytes and resident macrophages, MCP-1 2/2 mice were specifically unable to recruit monocytes 72 h after intraperitoneal thioglycollate administration, indicating that MCP-1 is essential for monocyte recruitment and the expression of cytokines related to T helper responses in vivo [27].
MCP-1 can be upregulated in the liver in vivo by I/R [28] as well as by bacterial endotoxin [29], and in isolated mouse hepatocytes by hypoxia [30] and the present study). The expression of MCP-1 in normoxic, primary cultured hepatocytes was thought to be due to the stress of cell isolation [31], and we likewise found elevated MCP-1 in normoxic hepatocytes. The effects of MCP-1 in the liver range from protective (e.g. in experimental acetaminopheninduced injury) to detrimental (e.g. during experimental I/R injury) [30].
In addition to MCP-1, our studies suggest that IL-1a, IL-6, KC, and MIG exhibit the most significant dynamic changes in stressed hepatocytes. Our results suggest that MCP-1 regulates the production of these other inflammatory mediators. This hypothesis is supported by prior studies, in which the expression of hepatic pro-inflammatory cytokines (including IL-6 and KC) was induced in alcohol-fed wild-type, but inhibited in MCP-1 2/2 animals [32]. In support of this central role for KC, Frink et al showed that neutralization of this chemokine ameliorated liver damage after T/HS in mice. These authors also showed that MCP-1 causes organ damage via upregulation of KC, supporting our findings of a network of hepatic inflammation that involves these two chemokines. However, the exact source of KC remained unknown [33]; our present study suggests that hepatocytes are a key source of this chemokine.
Multiple studies have shown that IL-6 is a biomarker for adverse outcomes in settings such as trauma/hemorrhage and sepsis and specifically of poor outcome in trauma [18,19]. This was reflected in our own clinical cohort, since we could segregate survivors from non-survivors based on plasma IL-6 levels. Importantly, our studies suggest that IL-6 is highly correlated with MCP-1 in T/HS patients, and that the production of IL-6 by hypoxic hepatocytes was essentially abolished in the absence of endogenous MCP-1. Taken together, our studies suggest that MCP-1 drives IL-6 production by hepatocytes. Interestingly, a recent study suggests the presence of a positive feedback loop involving IL-6 and MCP-1 in a setting of vascular inflammation [34].
The use of applied mathematical models and algorithms constitute a novel tool for interpreting complex biological data. Primary hepatocytes from wild-type mice were cultured under normoxic or hypoxic conditions (1-72 h) followed by measurement of inflammatory mediators in both lysates and supernatants and lysates as described in the Materials and Methods. After data normalization, DyNA during each of the following five time frames: 1-3 h, 3-6 h, 6-24 h, 24-48 h, and 48-72 h was performed for both lysates and supernatants as indicated. Panels show a summary of the DyNAs representing the most connected inflammatory mediator ''nodes'' for both normoxia (A) and hypoxia (B) lysates and supernatants. Each mediator's node size is proportional to the number of connections it has in a given time interval. (C) Stacked bars representing the total number of connections for each inflammatory mediator over all time intervals. doi:10.1371/journal.pone.0079804.g002 In 2005, a pioneering study in this field reported on a systems model comprising a large number of intracellular signaling events directly linked to outputs associated with apoptosis, created in order to study how molecular information is processed as a network [35]. Prior studies have also utilized advanced computational analyses to discern key features of the inflammatory biology of hepatocytes. For example, Alexopoulos et al studied 26,000 protein state measurements from isolated primary human hepatocytes and HepG2 liver cancer cells exposed to growth factors or inflammatory mediators, yielding interaction graphs using multilinear regression [36]. The authors suggested major differences between primary and transformed hepatocytes with respect to Toll-like receptor-4 signaling and NF-kB-dependent secretion of chemokines and cytokines, and suggested, as we do here, that combined in vitro/in silico analyses can yield novel insights into hepatocyte biology [36]. More recently, Boolean logic models of immediate-early signaling in liver cells were created by training a literature-based prior knowledge network against biochemical data obtained from primary human hepatocytes as well as four hepatocellular carcinoma cell lines exposed to combinations of cytokines and small-molecule kinase inhibitors [37]. Our present study extends this paradigm to define key nodes in the stress-induced inflammatory networks induced in hepatocytes, as well as suggesting higher-order, clinically-applicable insights. Indeed, this study points to various similarities between hepatocytes in vitro and T/HS in vivo. For example, as in the present study, we found that, together with MIG, IL-6 was highly elevated but disconnected from other inflammatory mediators comprising a dynamic network of inflammation induced by experimental T/HS in mice [12]. Our present study raises the possibility that MCP-1 is the stimulus for post-HS MIG and IL-6.
In the present study, we developed a novel algorithm for consensus clustering of multiple data-driven analyses of dynamic responses. We hypothesized that those inflammatory mediators exhibiting the most coordination across experimental conditions could be important drivers or indicators of that process. The strategy we employed was to extract patterns from dynamic data, and to assess the measure of similarity across multiple dynamic patterns. To distill the most important information from three independent clustering results over 18 mediators for each of 4 experimental conditions, we filtered the results by discarding inflammatory mediators that showed inconsistent segregation patterns across the three analyses. A consensus clustering containing only the mediators whose coordination patterns in which we could be most confident was the basis of comparison across experimental conditions. This method identified MCP-1 as the most relevant mediator in our studies.
This novel approach to data-driven modeling helped formulate a key hypothesis of our study, namely that circulating levels of MCP-1 could serve as a biomarker for mortality in human trauma/hemorrhage, and that the outcomes of trauma survivors would differ in high-vs. low-producers of MCP-1. Importantly, our results suggest that circulating levels of MCP-1 are better discriminators of outcome in our moderately-injured blunt trauma survivors than hypotension. This, despite the fact that even a brief period of hypotension was suggested to elevate post-trauma morbidity and mortality in injured patients [38]. Collectively, our results provide a systems view of the hepatocyte inflammatory response to cell stress, with implications for ischemia and hemorrhage, and point to hepatocytes as a likely cellular source of cytokines/chemokines in the inflammatory response. Furthermore, these results highlight the capacity of data-driven analyses for suggesting novel, clinically-relevant targets.

Mouse hepatocyte isolation and culture
All procedures involving animals were approved by the Animal Care and Use Committee of the University of Pittsburgh.  Hepatocytes were harvested from wild-type C57BL/6 mice (n = 8 animals from Charles River Laboratories, Wilmington, MA) or MCP-1/CCL2-null (MCP-1 2/2 ) mice (n = 4 animals from The Jackson Laboratory, Bar Harbor, ME) [27] and plated as previously described [39]. After overnight incubation, the medium was removed and cells were further incubated with fresh media containing 5% heat-inactivated calf serum. Hypoxic conditions were obtained by placing the cells into a modular incubator chamber (Billups-Rothenburg, Del Mar, CA) flushed with a hypoxic gas mixture containing 1% O 2 , 5% CO 2 and 94% N 2 . Normoxic hepatocytes served as control.

Protein isolation, sample preparation and Western blot analysis
At the end of each experiment, the supernatants were stored at 220uC until further analysis and the cell monolayers were washed twice with ice-cold PBS and resuspended in ice-cold lysis buffer (Cell Lysis Buffer 106 from Cell Signaling Technology, Danvers, MA) containing the protease inhibitors leupeptin (0.1 mg/ml) and phenylmethylsulfonyl fluoride (1 mM) followed by total protein isolation as previously described [39]. Protein samples (50 mg) were separated on 15% SDS-polyacrylamide gels followed by electroblotting onto PVDF nitrocellulose membranes. Immunodetection of MCP-1 was done using a mouse specific rabbit polyclonal anti-MCP-1 antibody (Cell Signaling, Danvers, MA) at 1:1000 dilution, and the immunoreactive bands were visualized after incubation with the SuperSignal West Dura Extended Duration Substrate mouse kit (Thermo Scientific, Rockford, IL). For normalization, the membranes were stripped and re-probed with an anti-b-actin antibody from Abcam (Cambridge, MA).

Statistical analysis and data-driven modeling
Before the statistical analyses and in order to account for experimental variability in cell number and protein concentration between individual experiments, the final mediator concentrations adjusted for protein content (CF) were obtained from the Luminex analysis for each individual mediator measurement (C) and its respective lysate protein concentration (as determined by BCA, see above) using the following formulas: and for cell lysates CFC Experimental data are presented as mean 6 SEM. Analysis was performed by ttest, One-way or Two-way Analysis of Variance (ANOVA) followed by the Holm-Sidak post-hoc test as indicated. Significance of differences in clinical outcomes was determined by Mann-Whitney U test using SigmaPlot (Systat Software, San Jose, CA).
Before any dynamic analysis, the final mediator concentrations (CF) were first normalized for each inflammatory mediator, so that all mediator levels were converted into the same scale (from 0 to 1). All measurements of a given inflammatory mediator from a single animal were normalized as a group, by dividing each value by the Euclidean norm of the group. In this way, any artifactual effects on variance due to the different ranges of concentration observed for different mediators were eliminated.

Hierarchical clustering
Fold change values were calculated by dividing the difference between the later (x 2 ) and earlier (x 1 ) mediator values by the earlier value: (x 2 -x 1 )/x 1 ; we avoided division by 0 by replacing x 1 with a very small number (10 25 ) when necessary. The columns of the dendrogram contained the mean of fold changes across all animals. We employed several MatLabH functions to carry out this analysis: linkage, with average link and Pearson correlation distance, and cluster was used to find 2-4 clusters in both fold change and raw data (mediator concentrations adjusted for protein content vs. time, unaveraged, all animals included).

Principal Component Analysis and k-means clustering
Normalized inflammatory mediator data (each measurement taken to be a single point in 18-dimensional inflammatory mediator space) were transformed into principal component space using the MatLabR function, princomp. We then examined projections of inflammatory mediators into principal component  space by using the score coefficients for the first 2, 3, or 4 principal components. We employed k-means clustering using the MatLabH function, kmeans, using as a distance metric the cosine of the angle between a point and the origin. We computed clusters starting with k = 2, 3, or 4 random centroids. At each level of k, we repeated the clustering 10 times, resetting new centroids each time and taking the lowest mean distance between points and their nearest centroids as the best result.   Fig. 7 Panels B-C). Panel C shows the overall demographics of the normotensive (n = 27) and hypotensive patients (n = 19) segregated according to plasma MCP-1 levels (,1000 pg/ml and .1500 pg/ml) and blood pressure status. Results represent the mean 6 SEM. doi:10.1371/journal.pone.0079804.g008

Meta-clustering
The Adjusted Rand Index (ARI) was first calculated pairwise between clustering results including 2, 3, or 4 principal components, choosing the result that captured the most information with the fewest principal components. We then used ARI to compare hierarchical clustering results with k = 2, 3, and 4 to k-means clustering results with k = 2, 3 and 4 in a pairwise manner. The maximum Rand Index dictated the k of the hierarchical and kmeans clustering results that were moved forward into further analysis. We compared PCA k-means clustering to each of fold change hierarchical clustering and raw data hierarchical clustering individually. The k was chosen for k-means clustering that maximized both ARI values. We then filtered the results by discarding inflammatory mediators that showed inconsistent segregation patterns across the three analyses. The basis of comparison across different experiments was a consensus clustering containing only the mediators for whose coordination patterns we had confidence, as shown in detail in Figures S9, S10, S11, S12, S13, S14, S15, S16. Briefly, a block of data was constructed with one mediator in each row and its corresponding cluster label for each of the three results in the columns using Microsoft ExcelH. Sorting this block of data by each of the cluster label columns successively allowed us to compare the memberships of clusters across the three results and to identify corresponding clusters between results. These were clusters that contained a common set of mediators across all three results. We used the following heuristic to determine consensus: once a cluster from one result was associated with clusters from the other results, (e.g. Cluster A from Raw Data was associated with Cluster 2 from Fold Change and Cluster IV from PCA clustering) it could not be associated with any other clusters. For example Cluster 1 could not associate with both Cluster A and Cluster B. Mediators that did not segregate into corresponding clusters across all results were deemed inconsistent and omitted from the consensus clustering result. This consensus outcome became the basis of comparison among experimental conditions (see Figures S9, S10, S11, S12, S13, S14, S15, S16).

Dynamic Network Analysis (DyNA)
Inflammatory mediator networks were created over time periods between consecutive measurements (1-3 h, 3-6 h, 6-24 h, 24-48 h, and 48-72 h) using MatlabH and InkscapeH software (http://inkscape.org/). In order to be included in the DyNA, a given mediator had to be statistically significantly different from its baseline value (no treatment [time = 1 h]; P,0.05 by Student's t-test). Connections in the network were created if the correlation coefficient between two nodes (inflammatory mediators) was greater or equal to a threshold of 0.7 (based on a total of 12 samples with 10 degrees of freedom, P,0.05).

Human trauma patient selection and analysis
All human sampling was done following approval by the University of Pittsburgh Institutional Review Board and written informed consent was obtained from each patient or next of kin as per Institutional Review Board regulations. From a cohort of 493 blunt trauma patients admitted to Presbyterian University Hospital (part of the University of Pittsburgh Medical Center), we selected two cohorts of blunt trauma patients (all $18 years old). Exclusion criteria included: life expectancy ,24 h, penetrating trauma only, and traumatic brain injury. The first cohort included 30 Figure S6 for detailed description). (TIF) Figure S9 Meta-clustering process: from three independent clustering results to a consensus clustering. Each method is colored one hue (blue = hierarchical clustering of raw data, orange = hierarchical clustering on fold changes, green = kmeans clustering in PCA space). Clusters within each method are demarcated by shades and tones of each hue (Step 1). Each column was sorted alphabetically by mediator name (Step 2), and the clustering results were combined into a single matrix (Step 3). The rows of this matrix were sorted by cluster labels, for each method sequentially (Step 4). This sorting allowed visual identification of clusters that were associated across methods. From here, a consensus (Step 6) was determined by identifying associated clusters and removing inconsistent mediators (Step 5) that fell outside of the associated clusters. Associated clusters are those that contain the same mediators across all three clustering methods. If two or more clusters from one method were associated with one cluster from each of the other methods, they were considered to be part of the same cluster, provided neither cluster was associated with any other clusters. Analysis for normoxia supernatants (wild-type): In Step 4 for this matrix, the rows were first sorted according to fold change cluster labels (orange), then by raw data cluster labels (blue), and finally by PCA clusters (green). In Step 5, all mediators from Cluster A (light blue) were also found in Cluster 2 (bright orange) and Cluster IV (dark green). Likewise, all mediators from Cluster 1 (light orange) were found in Clusters B (bright blue) and II (light green). Members of those clusters that were not found in Cluster 1 were discarded. Step 4 for this matrix, the rows were first sorted according to raw data cluster labels (blue), then by fold change cluster labels (orange), and finally by PCA clusters (green). In

Supporting Information
Step 5, all mediators from Cluster D (dark blue) were also found in Cluster 2 (bright orange) and Cluster II (light green) and therefore, these three clusters were considered to be the same. Any mediators associated with Cluster II, but not with both Cluster 2 and Cluster D were then marked as inconsistent and discarded from the analysis. Clusters I (grey), III (medium green), and IV (dark green) were all associated with Cluster B (bright blue) and Cluster 2 (bright orange). Step 4 for this matrix, the rows were first sorted according to fold change cluster labels (orange), then by raw data cluster labels (blue), and finally by PCA clusters (green). In Step 5, all mediators from Cluster C (medium blue) were also found in Cluster 1 (light orange) and Cluster I (grey) and therefore, these three clusters were considered to be the same. Any mediators associated with Cluster C, but not with both Cluster 1 and Cluster I were then marked as inconsistent and discarded from the analysis. Clusters B (bright blue) and D (dark blue) were both associated with Cluster II (light green) and Cluster 2 (bright orange). (TIF) Step 4 for this matrix, the rows were first sorted according to fold change cluster labels (orange), then by raw data cluster labels (blue), and finally by PCA clusters (green). In Step 5, most mediators from Cluster A (light blue) were also found in Cluster 2 (medium orange) and Cluster I (grey) and therefore, these three clusters were considered to be the same. Any mediators associated with Cluster A, but not with both Cluster 2 and Cluster I were then marked as inconsistent and discarded from the analysis. Clusters C (medium blue) and D (dark blue) were both associated with Cluster 4 (dark orange) and Cluster III (medium green). (TIF) Step 4 for this matrix, the rows were first sorted according to fold change cluster labels (orange), then by raw data cluster labels (blue), and finally by PCA clusters (green). In Step 5, a majority of mediators from Cluster II (light green) were also found in Cluster B (bright blue) and Cluster 1 (light orange) and therefore, these three clusters were considered to be the same. Any mediators associated with Cluster II, but not with both Cluster B and Cluster 1 were marked as inconsistent and discarded from the analysis. Likewise, Clusters D (dark blue), 2 (bright orange), and III (medium green) were associated. (TIF) Figure S14 Meta-clustering analysis (see legend of Fig.  S9) for hypoxia supernatants (MCP-1 2/2 ): In Step 4 for this matrix, the rows were first sorted according to fold change cluster labels (orange), then by raw data cluster labels (blue), and finally by PCA clusters (green). In Step 5, all mediators from Cluster A (light blue) were also found in Cluster 2 (bright orange) and Cluster III (medium green) and therefore, these three clusters were considered to be the same. Any mediators associated with Cluster A, but not with both Cluster 2 and Cluster III were marked as inconsistent and discarded from the analysis. Step 4 for this matrix, the rows were first sorted according to fold change cluster labels (orange), then by raw data cluster labels (blue), and finally by PCA clusters (green). In Step 5, all mediators from Cluster 1 (light orange) were also found in Cluster B (bright blue) and Cluster II (light green) and therefore, these three clusters were considered to be the same. Any mediators associated with Cluster 1, but not with both Cluster B and Cluster II were marked as inconsistent and discarded from the analysis. Clusters III (medium green) and IV (dark green) were both associated with Cluster B (bright blue) and Cluster 2 (bright orange). Step 4 for this matrix, the rows were first sorted according to raw data cluster labels (blue), then by fold change cluster labels (orange), and finally by PCA clusters (green). In Step 5, all mediators from Cluster B (medium blue) were also found in Cluster 3 (bright orange) and Cluster III (medium green) and therefore, these three clusters were considered to be the same. Any mediators associated with Cluster B, but not with both Cluster 3 and Cluster III were marked as inconsistent and discarded from the analysis. Clusters I (grey), II (light green), and IV (dark green) were all associated with Cluster A (light blue) and Cluster 1 (light orange). (TIF) Figure S17 Dynamic Network Analysis (DyNA) of inflammatory mediators produced by normoxic and hypoxic mouse hepatocytes. Primary hepatocytes from wild-type mice were cultured under normoxic or hypoxic conditions (1-72 h) followed by measurement of cytokines/ chemokines in both lysates and supernatants and lysates as described in the Materials and Methods. Gray boxes indicate that the mediator is statistically significantly different from its baseline value (no treatment [time = 1 h]; P,0.05) and the digits represent the number of connections resulting from the DyNAs during each of the following five time frames: 1-3 h, 3-6 h, 6-24 h, 24-48 h, and 48-72 h for both lysates and supernatants. (TIF) Figure S18 Dynamic Network Analysis (DyNA) of inflammatory mediators produced by normoxic and hypoxic mouse hepatocytes from MCP-1 2/2 mice. Primary hepatocytes from MCP-1 2/2 mice were cultured under normoxic or hypoxic conditions (1-72 h) followed by measurement of cytokines/chemokines in both lysates and supernatants and lysates as described in the Materials and Methods. Gray boxes indicate that the mediator is statistically significantly different from its baseline value (no treatment [time = 1 h]; P,0.05) and the digits represent the number of connections resulting from the DyNAs during each of the following five time frames: 1-3 h, 3-6 h, 6-24 h, 24-48 h, and 48-72 h for both lysates and supernatants. (TIF)