Conceived and designed the experiments: RZ YV. Performed the experiments: AT RN. Analyzed the data: QM GC CZ AS TRB JCP YV. Contributed reagents/materials/analysis tools: QM GC CZ AS. Wrote the paper: QM GC CZ TRB RZ YV. Contributed to the making of the hemorrhagic shock apparatus: TB.
Current address: Department of Internal Medicine, Hurley Medical Center, Michigan State University Collage of Human Medicine, Flint, Michigan, United States of America
The authors have declared that no competing interests exist.
Complex biological processes such as acute inflammation induced by trauma/hemorrhagic shock/ (T/HS) are dynamic and multi-dimensional. We utilized multiplexing cytokine analysis coupled with data-driven modeling to gain a systems perspective into T/HS.
Mice were subjected to surgical cannulation trauma (ST) ± hemorrhagic shock (HS; 25 mmHg), and followed for 1, 2, 3, or 4 h in each case. Serum was assayed for 20 cytokines and NO2−/NO3−. These data were analyzed using four data-driven methods (Hierarchical Clustering Analysis [HCA], multivariate analysis [MA], Principal Component Analysis [PCA], and Dynamic Network Analysis [DyNA]). Using HCA, animals subjected to ST vs. ST + HS could be partially segregated based on inflammatory mediator profiles, despite a large overlap. Based on MA, interleukin [IL]-12p40/p70 (IL-12.total), monokine induced by interferon-γ (CXCL-9) [MIG], and IP-10 were the best discriminators between ST and ST/HS. PCA suggested that the inflammatory mediators found in the three main principal components in animals subjected to ST were IL-6, IL-10, and IL-13, while the three principal components in ST + HS included a large number of cytokines including IL-6, IL-10, keratinocyte-derived cytokine (CXCL-1) [KC], and tumor necrosis factor-α [TNF-α]. DyNA suggested that the circulating mediators produced in response to ST were characterized by a high degree of interconnection/complexity at all time points; the response to ST + HS consisted of different central nodes, and exhibited zero network density over the first 2 h with lesser connectivity vs. ST at all time points. DyNA also helped link the conclusions from MA and PCA, in that central nodes consisting of IP-10 and IL-12 were seen in ST, while MIG and IL-6 were central nodes in ST + HS.
These studies help elucidate the dynamics of T/HS-induced inflammation, complementing other forms of dynamic mechanistic modeling. These methods should be applicable to the analysis of other complex biological processes.
The advent of multi-dimensional datasets derived from dynamic experiments on complex biological systems has resulted in a deluge of data, but this massive increase in data has not necessarily translated to enhanced mechanistic understanding
Traumatic injury, often accompanied by hemorrhage, represents the most common cause of death for young people, as well as a significant source of morbidity and mortality for all ages
The complex nature of the response to T/HS, with its many redundant and overlapping pathways and mediators
Herein, we applied a set of novel, data-driven methods to dynamic, multi-dimensional data derived from a highly-precise, survivable mouse model of T/HS in order to discern novel mechanistic interactions directly from data. These studies demonstrate that survivable trauma elicits an inflammatory response as early as 1 h post-injury. Our results also suggest that the response to low-level trauma is driven by particular cytokines in a complex and well-ordered manner, while the addition of survivable HS leads to the elaboration of distinct inflammatory mediators as part of a much less complex and less organized response.
This study was approved by the Institutional Animal Care and Use Committee of the University of Pittsburgh (protocol No. 1003645) and was conducted in accordance with the National Institutes of Health Guidelines for the Care and Treatment of Small Laboratory Animals. All studies were initiated only following a two-week acclimatization period at the University of Pittsburgh, Biomedical Science Tower Animal Facility, with access to food and water
A central goal of this study was to assess the dynamics of several key inflammatory analytes, which are representative of the acute inflammatory response and which have been shown to be modulated in humans that have undergone trauma/HS
The following analyses were carried out in an attempt to discern differences in, and derive mechanistic insights from, changes in inflammatory mediators across experimental procedures. The null hypothesis for all of these studies was that inflammatory mediators could not segregate ST from ST + HS. The schematic of the analyses and their respective goals is depicted in
Mice were subjected to ST ± HS followed by measurement of cytokines, chemokines, and NO2−/NO3− as described in the
Univariate analysis explores individual variable in a data set. It describes the pattern of response to the variable. Our response variables are the 21 inflammatory mediators described above, and
IL-12.Total | IL-6 | IP-10 | KC | MIG | |
|
2.43 | 1.76 | 3.31 | 2.20 | 2.07 |
|
2.76 | 2.18 | 1.35 | 0.99 | 3.13 |
|
6.88 | 2.24 | 2.22 | 2.96 | 6.35 |
|
3.04 | 3.08 | 2.17 | 4.84 | 1.46 |
|
4.75 | 2.65 | 3.00 | 3.31 | 4.42 |
Mice were subjected to ST ± HS followed by measurement of cytokines, chemokines, and NO2−/NO3− as described in the
In the sections below, we describe the approach utilized to determine if inflammatory mediators could predict the
The goal of this analysis was to highlight the natural variability, as well as any overlap, in inflammatory mediators from animals subjected to ST or ST + HS. Hierarchical clustering is a simple and unbiased clustering method which aims to build a hierarchy of clusters. The limitation is the cluster must be built pairwise; since it is purely based on the similarity between the data, the cluster may lack biological relevance
The goal of this analysis was to determine which inflammatory mediators reach levels sufficiently different following each insult so as to discriminate between ST and ST + HS. To do so, a multivariate statistical model was developed that takes as input the cytokine readings in mice and yields as output the probability that the mouse in question belongs to a specific group: ST + HS or ST only. The model uses an additive, main effects only design. The experimental procedures ST and ST + HS represent a binary response. Specifically, if
where the β's are unknown parameters subject to estimation, and the X's represent the predictor variables (selected inflammatory mediators).
Several predictive classes of models were investigated, and the logistic family was found to be the best suited for this task. The individual predictive ability of each mediator was ranked by using the corresponding p-values derived through the logistic model fit involving that sole mediator as input variable. In addition, a predictive model involving just two cytokines,
The goal of this analysis was to identify the subsets of mediators (in the form of orthogonal normalized linear combinations of the original mediator variables, called principal components) that are most strongly correlated with a given experimental procedure (ST or ST+HS), and that thereby might be considered principal drivers of each response. PCA is a non-parametric statistical method of reducing a multidimensional dataset to a few principal components
We initially examined the levels of inflammatory analytes in the serum of C57Bl/6 that were subjected to ST ± HS, to confirm prior studies that have demonstrated elevations in circulating inflammatory analytes (e.g. TNF-α, IL-6, IL-10, NO2−/NO3−) following T/HS in mice
Despite these dynamic changes in inflammation biomarkers as a function of time, we sought to determine if a significant proportion of these 21 mediators were altered as a function of time. This question is especially important for any conclusions that might be drawn regarding principal drivers or dominant networks.
Mice were subjected to ST ± HS followed by measurement of cytokines, chemokines, and NO2−/NO3− as described in the
To gain a systems perspective on these complex, time-dependent responses to ST ± HS, we carried out univariate analysis, multivariate analysis (MA), hierarchical clustering analysis (
Mice were subjected to ST ± HS followed by measurement of cytokines, chemokines, and NO2−/NO3− as described in the
Mice were subjected to ST ± HS followed by measurement of cytokines, chemokines, and NO2−/NO3− as described in the
Mice were subjected to ST ± HS followed by measurement of cytokines, chemokines, and NO2−/NO3− as described in the
Despite this overlap, we hypothesized that data-driven analyses would uncover distinct features of inflammation in ST vs. ST + HS. We initially employed both univariate and multivariate analyses.
We first focused on the time-dependent differences in individual mediators by performing independent univariate analyses. The means of the 21 inflammatory mediators induced in response to ST ± HS in the present study are depicted in
We next carried out an ANOVA for the five responses deemed most significant from the initial univariate analysis (IL-12, IL-6, IP-10, KC, and MIG); the results of this analysis are summarized in
A more refined model clarifies exactly which interactions between
with e denoting a Gaussian random variable with 0 mean, and Time.L, signifying the linear effect of time. All three effects are statistically significant at a level of 0.007. The quadratic time effect Time.Q and the interaction Procedure*Time.Q are not statistically significant (the p-values are 0.31 and 0.28, respectively). This means that IL-12.total grows linearly only (not quadratically) with time, and that this linear time growth depends on the Procedure (explaining the existing interaction). A comparison of the actual Total IL-12 data to the fitted value produced by this model is found in the
The five cytokine responses of interest are correlated across all data. The correlation matrix appears in
A major goal of this analysis is to determine a model that uses mediators as predictors for the experimental
With the relevant mediators for the
with the estimated coefficients carrying p-values of 0.0005, 0.0019, 0.0082, respectively.
As depicted in
We next attempted to leverage the insights gained from statistical analyses into mechanistic insights regarding the dynamics of inflammation following T/HS. We initially utilized PCA in order to identify the subsets of mediators that are most strongly correlated with ST or ST + HS, and that thereby might be considered principal drivers of each response. Importantly, PCA is based on time-dependent changes in variance, and therefore we hypothesized that this analysis would yield insights into the dynamic responses of the various inflammatory mediators.
Finally, we wished to expand our mechanistic analysis further by examining the time-dependent evolution of cytokine networks inferred from correlated changes in circulating inflammatory mediators; we refer to this process as Dynamic Network Analysis (DyNA). We wished not only to determine which networks were present at various time intervals, but also to assess the total degree of connectivity at each of these intervals.
Finally, we wished to go beyond an examination of inflammatory mediators and assess the global state of inflammatory networks, by quantifying the degree of network connectivity as a function of time following ST ± HS (
Detailed cellular and molecular analyses explored in isolation have provided valuable insights into the pathobiology of sepsis and T/HS, but have often been limited in their global applicability
In the studies described herein, mice were subjected to highly precise and reproducible experimental T/HS (bleeding down to 25 mmHg without resuscitation) for 1–4 h using a computerized hemorrhage system described previously
We hypothesized that the data regarding the dynamic evolution of these 21 mediators/biomarkers could be analyzed using data-driven modeling approaches, following the framework depicted in
In the first category (across experimental procedures), we employed two distinct methods. Hierarchical Clustering Analysis was used to examine both the natural variability of and the overlap in circulating inflammatory mediators in animals subjected to ST or ST + HS. This analysis highlighted the relatively high degree of overlap between ST and ST + HS. A prior study from our group had also described this large overlap in the pathways induced by ST and ST + HS, though this prior study only examined TNF-α, IL-6, IL-10, and NO2−/NO3− as well as changes in the liver transcriptome
Recent studies have reported on the use of multiplexed cytokine analysis coupled with multivariate regression modeling in mouse models of inflammation, e.g. colitis
In the second category (within a given experimental procedure), PCA was employed in order to discern the main drivers of inflammation and DyNA was utilized in order to define the principal (most connected) nodes being elaborated dynamically as a function of pro-inflammatory insult. The hypothesis underlying the use of PCA was that such main drivers might act “behind the scenes”, and be discerned as those mediators exhibiting the greatest, insult-specific, time-dependent variance. Thus, these principal mediators are hypothesized to define a given experimental procedure across the entire time range studied. It is therefore entirely possible that principal mediators defined thus may not reach statistical significance, since they may carry out their function for a limited period of time and drive the production of other mediators that would in fact remain statistically elevated to a degree sufficient to be detected by MA. Though utilized in a manner somewhat similar to PCA, DyNA was used to gain insights into dynamic changes in network connectivity of the inflammatory response to ST and ST + HS over time, allowing for insights that are difficult, if not impossible, to gain from any of the other data-driven analyses utilized in this study.
We gained several insights from our network analysis. For example, the earliest pro-inflammatory mediators in our mechanistic mathematical models of post-T/HS inflammation is TNF-α, with IL-6 elaborated fairly soon afterwards
Beyond such mediator-focused insights, the DyNA studies also uncovered an additional dimension of information about the connectivity of the early inflammatory response to T/HS, namely that the response to a minor trauma (ST) appeared well-ordered and was driven by defined networks orchestrated by chemokines and cytokines. In contrast, the response to that same minor trauma in the presence of HS (ST + HS) was characterized by a complete lack of connectivity among mediators in the first 2 h. Though the degree of connectivity appeared to recover, the networks involved in this attempt at recovery were distinct from those present in the mice not subjected to combined T/HS. Intriguingly, a comparison of network density / complexity over time suggested a “mirror image” pattern when comparing ST vs. ST + HS. While we do not wish to over-interpret this aspect of our data, such a pattern may imply that baseline inflammatory connectivity is initially perturbed upwards (more complexity) by ST, while the addition of HS perturbs baseline connectivity downward (lower complexity) to approximately the same degree. Over time, both responses appear to return towards baseline connectivity, with inflammatory connectivity in ST still remaining higher than ST + HS. We hypothesize that this difference is due to the presence of HS and not to the animals' being near death, since our prior experience
Each of the analyses we performed served a distinct purpose, and therefore these analyses were expected to provide complementary, rather than identical, results. We also expected to find some concordance with our prior mechanistic mathematical modeling of T/HS in mice. Importantly, using the above-described methods, the difference between ST and ST + HS in this experimental model could clearly be distinguished over time, based on certain inflammatory mediators (as well as the mediators that correlated highly with these distinguishing mediators, namely IL12-total, MIG, KC, IL6, and IP10. In addition, the finding by MA that total IL-12 was a good discriminator of ST vs. ST + HS at 3 h is in accord with the DyNA results, which suggests an interaction of IL-12 with NO2−/NO3− between 2 and 3 h post-ST. Interestingly, Diefenbach
We suggest that data at the mRNA and protein levels, combined with data-driven methods such as those described in this study, may facilitate further mechanistic modeling of the dynamics of acute inflammation as well as driving clinically-relevant advances
Several limitations are associated with our study. One central limitation may revolve around the possible confounding role of anesthesia in our analyses. Prior analyses have suggested that anesthesia may affect inflammatory and related physiological responses
Thus, at least some of the inflammatory response associated with either ST or ST + HS may be due to (or modulated by) the anesthesia used for both procedures. Another limitation concerns the lack of certain key mediators and biomarkers in T/HS, e.g. DAMP's such as HMGB1 or soluble TNF-α receptors. Another limitation of the interpretation of our study is that insult-specific mediators defined by MA may reach statistically different levels not because they are necessarily primary drivers of ST or ST + HS, but perhaps because they are induced to the greatest degree or for the longest duration. Despite these limitations, we suggest that mechanism-focused data-driven analyses based on time-varying, high-content datasets will serve to generate hypotheses regarding the induction and propagation of inflammation, and eventually yield insights into novel therapies.
(PPT)
(PPT)
(PPT)
(PPT)
(DOC)
(DOC)
(DOC)
(DOC)
(DOC)
(DOC)
(DOC)
(DOC)
(RAR)
(RAR)