Cytokine Expression Patterns Associated with Systemic Adverse Events following Smallpox Immunization

Abstract

Vaccinia virus is reactogenic in a significant number of vaccinees, with the most common adverse events being fever, lymphadenopathy, and rash. Although the inoculation is given in the skin, these adverse events suggest a robust systemic inflammatory response. To elucidate the cytokine response signature of systemic adverse events, we used a protein microarray technique to precisely quantitate 108 serum cytokines and chemokines in vaccine recipients before and 1 week after primary immunization with Aventis Pasteur smallpox vaccine. We studied 74 individuals after vaccination, of whom 22 experienced a systemic adverse event and 52 did not. The soluble factors most associated with adverse events were selected on the basis of voting among a committee of machine-learning methods and statistical procedures, and the selected cytokines were used to build a final decision-tree model. On the basis of changes in protein expression, we identified 6 cytokines that accurately discriminate between individuals on the basis of adverse event status: granulocyte colony–stimulating factor, stem cell factor, monokine induced by interferon-γ (CXCL9), intercellular adhesion molecule–1, eotaxin, and tissue inhibitor of metalloproteinases–2. This cytokine signature is characteristic of particular inflammatory response pathways and suggests that the secretion of cytokines by fibroblasts plays a central role in systemic adverse events.

Vaccination of healthy adults with vaccinia virus induces a protective response in the majority of individuals who are immunized, which is indicated by a significant amount of vaccinia virus–neutralizing antibodies in the serum. Recent studies have demonstrated that vaccinia-specific T lymphocytes secrete interferon-γ after immunization and that these cells may be long lived [1-3]. In a previous study, we investigated the effect of the Aventis Pasteur smallpox vaccine (APSV) on a limited panel of systemic cytokine concentrations in a cohort of individuals who were previously vaccinia naive [4]. Systemic cytokines representing lymphocyte functional subsets of Th1 cells (interferon-γ, tumor necrosis factor [TNF]–α, and interleukin [IL]-2) and Th2 cells (IL-4, IL-5, and IL-10) were measured using a sensitive flow cytometric bead array assay that allowed multiple cytokine analyses from a single sample [5]. Most notably, we found that smallpox immunization induces an interferon-γ–dominant response in the systemic compartment 1 week after immunization, with concentrations returning to baseline levels during convalescence. However, the concentration of systemic interferon-γ did not differ between subjects who experienced an adverse event, compared with subjects who did not.

To determine the molecular and cellular basis for systemic adverse events, we precisely quantitated 108 serum cytokines and chemokines using rolling-circle amplification technology [6-12] prior to and 1 week after immunization with APSV. We studied 74 individuals after primary vaccination, of whom 22 experienced a systemic adverse event and 52 experienced no adverse event. To limit the number of false discoveries while maintaining statistical power, we implemented an unweighted voting strategy among a committee of machine- and statistical-learning methods and procedures to analyze the responses. We used the support vector machines (SVM) and the nearest shrunken centroids (NSC) learning methods and the false discovery rate (FDR) corrected Wilcoxon rank-sum test to select the soluble factors most associated with adverse events. We then used a decision tree to model the functional relationship between the selected cytokines and systemic adverse events. In this analysis, we found systemic cytokine patterns characteristic of inflammation marked by the prominent induction of IL-17– and interferon-γ–related cytokines, as well as patterns characteristic of tissue inflammation and moderate destruction.

METHODS

Subjects

Healthy adult subjects 18–32 years old were enrolled in a multicenter study of primary immunization against smallpox using the APSV at National Institutes of Health (NIH) Vaccine and Treatment Evaluation Units. At the Vanderbilt University Medical Center site, 148 volunteers were enrolled in this NIH-sponsored APSV immunization trial (NIH–Division of Microbiology and Infectious Diseases Protocol 02-054). Vaccines, study subjects, and study design have been described elsewhere [13]. All subjects participating in the main smallpox immunization study at the Vanderbilt University Medical Center were invited to participate in this cytokine assessment sub-study. Serum samples for cytokine analysis were obtained from 107 of the 148 subjects vaccinated in this study at Vanderbilt after they provided informed consent, under approval from the Vanderbilt University Institutional Review Board. All 22 subjects who experienced systemic adverse events among the 107 who donated serum samples were studied for cytokine and chemokine response. A random selection of 52 subjects without adverse affects was studied for comparison.

Clinical assessments

Trained physicians and nurse providers examined the subjects by history and physical examination for vaccine take and adverse events at 5 visits in the first month after immunization (1 visit during days 3–5, days 6–8, days 9–11, days 12–15, and days 26–30). For the purpose of the current study, we considered the occurrence of the following 3 systemic adverse events: generalized rash, fever, and lymphadenopathy. Fever was defined as an oral temperature of >38.3°C. A generalized rash was defined as skin eruptions in regions not contiguous with the site of vaccination. The frequent acneiform rashes observed in this trial have been described elsewhere [14]. Lymphadenopathy was defined as tenderness or enlargement of regional lymph nodes associated with vaccination.

Sample collection

Prevaccination serum samples (baseline) were collected during a screening visit or just prior to vaccination, and postvaccination samples were obtained 6–9 days after vaccination (acute). Serum samples were collected in 5-mL Vacutainer serum separator tubes (Becton Dickinson) and were centrifuged at 700 g for 10 min, after which the serum was collected, aliquoted into cryovials (Sarstedt), and stored at −80°C until assayed for cytokine concentrations. Rolling-circle amplification technology was used to measure 108 serum cytokines and chemokines for all 22 subjects who experienced a systemic adverse event and for the 52 subjects who did not. In this study, we focused on systemic adverse events, which we expected to be associated more strongly with serum cytokine expression than we did a local adverse event.

Cytokine/chemokine assay

The expression levels of 108 protein analytes were measured in 100-μL serum aliquots of the patient samples using custom dual antibody sandwich immunoassay arrays, as described elsewhere [6-12]. The list of analytes is shown in table 1. Briefly, monoclonal capture antibodies specific for each analyte were fixed to glass slides, with 12 replicate spots for each analyte. Duplicate samples of serum were incubated for 2 h and then were washed. Slides then were incubated with secondary biotinylated polyclonal antibodies, and signals were amplified using a “rolling circle” method [9]. Quality control measures included optimization of antibody pairs, the use of internal controls to minimize array-to-array variation, and standardized procedures of chip manufacturing [9]. Arrays were scanned using a Tecan LS200 unit (Tecan Group), and mean fluorescence intensities were generated with customized software. To ensure a dynamic working range for each assay, 15 serial dilutions of recombinant analytes at known concentrations (studied in parallel on each slide) were used to develop best-fit equations for each analyte, and the upper and lower limits of quantitation were defined. Because of the broad range of systemic cytokine expression found for each patient before and after immunization, changes in serum cytokine concentrations that occurred during the early postimmunization phase were calculated as the percentage of the corresponding subject's baseline expression at the prevaccination visit.

Table 1.

Gene names and symbols of 108 protein analytes measured in 100-μL serum aliquots from patient samples using custom dual-antibody sandwich immunoassay arrays.

Gene symbol	Gene name
CSF3 (G-CSF)	Colony-stimulating factor 3 (granulocyte)
IL-10	Interleukin 10
IFNG	Interferon-γ
ALCAM	Activated leukocyte cell adhesion molecule
ANGPT4	Angiopoietin 4
BDNF	Brain-derived neurotrophic factor
CXCL13 (BLC)	Chemokine (C-X-C motif) ligand 13 (B-cell chemoattractant)
CCL28 (MEC)	Chemokine (C-C motif) ligand 28
TNFSF7 (CD27)	Tumor necrosis factor (ligand) superfamily, member 7
TNFSF8 (CD30)	Tumor necrosis factor (ligand) superfamily, member 8
CCL27 (CTACK)	Chemokine (C-C motif) ligand 27
TNFRSF21 (DR6)	Tumor necrosis factor receptor superfamily, member 21
EGF	Epidermal growth factor (β-urogastrone)
CXCL5 (ENA78)	Chemokine (C-X-C motif) ligand 5
CCL11 (Eot)	Chemokine (C-C motif) ligand 11
CCL26 (Eot3)	Chemokine (C-C motif) ligand 26
CCL24 (Eot2)	Chemokine (C-C motif) ligand 24
FGF4	Fibroblast growth factor 4 (heparin secretory transforming protein 1, Kaposi sarcoma oncogene)
FGF7	Fibroblast growth factor 7 (keratinocyte growth factor)
FGF9	Fibroblast growth factor 9 (glia-activating factor)
FGF2 (FGFB)	Fibroblast growth factor 2 (basic)
FGF1	Fibroblast growth factor 1 (acidic)
FAS	Fas (tumor necrosis factor receptor superfamily, member 6)
FASLG	Fas ligand (tumor necrosis factor superfamily, member 6)
FLT3LG	Fms-related tyrosine kinase 3 ligand
FST	Follistatin
CX3CL1	Chemokine (C-X3-C motif) ligand 1 (fractalkine, neurotactin)
CXCL6 (GCP2)	Chemokine (C-X-C motif) ligand 6 (granulocyte chemotactic protein 2)
GDNF	Glial cell–derived neurotrophic factor
CSF2 (GMCSF)	Colony-stimulating factor 2 (granulocyte-macrophage)
CXCL3 (GRO3)	Chemokine (C-X-C motif) ligand 3
CXCL2 (GRO2)	Chemokine (C-X-C motif) ligand 2
CCL14 (HCC1)	Chemokine (C-C motif) ligand 14
CCL16 (HCC4)	Chemokine (C-C motif) ligand 16
HGF	Hepatocyte growth factor (hepapoietin A; scatter factor)
TNFRSF14 (HVEM)	Tumor necrosis factor receptor superfamily, member 14 (herpesvirus entry mediator)
CCL1 (I309)	Chemokine (C-C motif) ligand 1
CXCL11 (ITAC)	Chemokine (C-X-C motif) ligand 11
ICAM1	Intercellular adhesion molecule 1 (CD54), human rhinovirus receptor
ICAM3	Intercellular adhesion molecule 3
IGF2	Insulin-like growth factor 2 (somatomedin A)
IGF1R	Insulin-like growth factor 1 receptor
IGFBP1	Insulin-like growth factor binding protein 1
IGFBP3	Insulin-like growth factor binding protein 3
IGFBP4	Insulin-like growth factor binding protein 4
IGFBP2	Insulin-like growth factor binding protein 2, 36 kDa
IL10RB	Interleukin 10 receptor, β
IL-13	Interleukin 13
IL-15	Interleukin 15
IL-17	Interleukin 17 (cytotoxic T-lymphocyte–associated serine esterase 8)
IL-1A	Interleukin 1, α
IL-1B	Interleukin 1, β
IL1-RN	Interleukin 1 receptor antagonist
IL1RL2	Interleukin 1 receptor–like 2
IL-2	Interleukin 2
IL2RB	Interleukin 2 receptor, β
IL2RA	Interleukin 2 receptor, α
IL-3	Interleukin 3 (colony-stimulating factor, multiple)
IL-4	Interleukin 4
IL-5	Interleukin 5 (colony-stimulating factor, eosinophil)
IL-6	Interleukin 6 (interferon, β 2)
IL-7	Interleukin 7
IL-8	Interleukin 8
IL2RG	Interleukin 2 receptor, γ (severe combined immunodeficiency)
IL5RA	Interleukin 5 receptor, α
IL-9	Interleukin 9
SELL	Selectin L (lymphocyte adhesion molecule 1)
CSF1	Colony-stimulating factor 1 (macrophage)
CSF1R	Colony-stimulating factor 1 receptor
CCL2 (MCP1)	Chemokine (C-C motif) ligand 2
CCL8 (MCP2)	Chemokine (C-C motif) ligand 8
CCL7 (MCP3)	Chemokine (C-C motif) ligand 7
CCL13 (MCP4)	Chemokine (C-C motif) ligand 13
CXCL9 (MIG)	Chemokine (C-X-C motif) ligand 9 (monokine induced by interferon γ)
CCL3 (MIP1A)	Chemokine (C-C motif) ligand 3
CCL4 (MIP1B)	Chemokine (C-C motif) ligand 4
CCL5 (MIP1D)	Chemokine (C-C motif) ligand 5
CCL20 (MIP3A)	Chemokine (C-C motif) ligand 20
CCL19 (MIP3B)	Chemokine (C-C motif) ligand 19
MMP7	Matrix metalloproteinase 7 (matrilysin, uterine)
MMP9	Matrix metalloproteinase 9 (gelatinase B, 92 kDa gelatinase, 92 kDa type IV collagenase)
CCL23 (MPIF1)	Chemokine (C-C motif) ligand 23
NTF3	Neurotrophin 3
NTF5	Neurotrophin 5 (neurotrophin 4/5)
OSM	Oncostatin M
PARC	P53-associated parkin-like cytoplasmic protein
PDGFRA	Platelet-derived growth factor receptor, α polypeptide
PECAM1	Platelet/endothelial cell adhesion molecule (CD31 antigen)
PGF	Placental growth factor, vascular endothelial growth factor–related protein
TNFRSF11A	Tumor necrosis factor receptor superfamily, member 11a, activator of NFKB
CCL5 (RANTES)	Chemokine (C-C motif) ligand 5
KITLG (SCF)	KIT ligand
KIT (SCFR)	V-kit Hardy-Zuckerman 4 feline sarcoma viral oncogene homolog
CXCL12 (SDF1)	Chemokine (C-X-C motif) ligand 12 (stromal cell–derived factor 1)
IL1RL1	Interleukin 1 receptor–like 1
CCL17 (TARC)	Chemokine (C-C motif) ligand 17
TGFA	Transforming growth factor, α
TIMP-2	Tissue inhibitor of metalloproteinase 2
TIMP1	Tissue inhibitor of metalloproteinase 1 (erythroid potentiating activity, collagenase inhibitor)
TNFRSF1A	Tumor necrosis factor receptor superfamily, member 1A
TNF	Tumor necrosis factor superfamily, member 2
LTA	Lymphotoxin α (tumor necrosis factor superfamily, member 1)
TNFRSF10A	Tumor necrosis factor receptor superfamily, member 10a
TNFRSF10D	Tumor necrosis factor receptor superfamily, member 10d, decoy with truncated death domain
VEGF	Vascular endothelial growth factor
KDR	Kinase insert domain receptor (a type III receptor tyrosine kinase)
BTG2	BTG family, member 2
NM	Neutrophil migration

Development of methods for statistical analysis

With a statistical hypothesis test, the naive approach to multiple-hypothesis testing is to declare the discrepancy between the 2 adverse event groups to be statistically significant for a cytokine if the P value is <α, where typically α = 0.05. This threshold is calibrated to declare false statistical significance with a probability of ∼α. However, the probability of spuriously selecting cytokines increases with the number of hypothesis tests that are performed (N). The Bonferroni correction chooses a significance level of α′ = α/N. This method takes the number of hypotheses tested into account and insures that the probability of declaring at least 1 false rejection is no greater than α, but the probability of erroneously maintaining the null hypothesis approaches unity as N increases.

A simple statistical procedure that controls the number of type I errors while maintaining good statistical power when performing multiple hypothesis tests is the FDR method [15]. On the basis of the distribution of P values generated by a statistical test, the FDR procedure returns a statistical significance threshold that controls the average fraction of false discoveries made among the multiple hypothesis tests whose null hypotheses were rejected. We used the Wilcoxon rank-sum test as our significance test for the comparison of means between systemic adverse event and non–adverse event groups. Using the FDR q = 0.3, we found the threshold of significance to be 0.02. The unweighted voting procedure, which involved SVM and NSC methods described below, was used to further filter out spurious associations. Unless otherwise stated, methods were implemented in the MATrixLABoratory (MATLAB) programming language, version 7.1 (release 13).

NSC and SVM are supervised learning methods in which the classes are predefined. Here, we considered 2 classes—systemic adverse event and non–adverse event—although the individual adverse event subsets could be treated as multiple classes in multiclass implementations of these algorithms. For binary classification, the basic aim of SVM is to find a hyperplane that maximally separates training data from the 2 classes. For cases in which no linear separation is possible, SVMs use a kernel technique to automatically realize a nonlinear mapping of a feature space. The hyperplane found by the SVM in feature space corresponds to a nonlinear decision boundary in input space [16]. We used the Gene Expression Model Selector analysis software, version 2.0.2 [17], with a radial basis function kernel, Markov blanket feature selection [18], and 10-fold cross-validation. With an average prediction accuracy of 69% across the 10 cross-validation splits, SVM found 7 cytokines, 5 of which passed our significance criterion of consensus with the other statistical methods.

NSC [19] was the final statistical learning method used in our committee-based strategy, because of its ability to perform automatic feature selection. For each cytokine i, the k components of the class centroids ${\stackrel{‒}{x}}_{\mathit{ik}}$ (or the mean of cytokines i for individuals in class k) are shrunk toward the overall centroid ${\stackrel{‒}{x}}_i$ (or the mean of cytokine i across all individuals). In our case, k = 1 or 2, corresponding to the 2 adverse event classes. The centroids are shrunk by a t statistic–like quantity $d_{\mathit{ik}}^{\prime }$, which is a measure of the ability of cytokine i to distinguish the class-k centroid from the overall centroid. If $d_{\mathit{ik}}^{\prime }$ is zero, then the cytokine-i component of the class-k centroid is equal to the component of the overall centroid, and this cytokine does not contribute to classification for class k. We used the same discriminant score described elsewhere [19], but we tuned the regularization term s₀ as well as the shrinkage Δ through 10-fold cross-validation. We found an average prediction accuracy of 70%, with Δ = 2.1 and s₀ = 0.002. The prediction accuracy of NSC was slightly better and required less computational time than that of SVM, and 4 of the 5 cytokines selected by NSC overlapped with the SVM selections.

The final step of our analysis was to create an interpretable model from the cytokines found by consensus among the 3 feature-selection methods (FDR-Wilcoxon, SVM, and NSC). Whereas these 3 feature-selection methods were chosen for their high power to select non-noisy variables and for their complementarity, decision trees were chosen to build the final adverse event model because of their ready interpretability and explicit modeling of variable interactions. We used the implementation of the C4.5 decision-tree algorithm provided in the Weka machine learning software package [20] to obtain the model in figure 1. Individuals were classified into adverse event or non–adverse event groups by sorting down a dichotomous tree toward terminal leaves. Starting from the root, the tree splits at a cytokine according to how well the relative change of a given cytokine separates individuals into the appropriate classes. Using information gain to rank cytokines, we placed cytokines at tree nodes with the greatest gain among attributes not yet considered in the path from the root node. We used a 25% confidence value for pruning branches that do not improve training accuracy and found the cross-validation accuracy to be insensitive to changes in this value. We optimized the minimum number of instances that must be present from each adverse event class in the training data for a new leaf to be created to handle those instances. For these data, a minimum of 5 instances resulted in a more-parsimonious tree that more-readily generalizes to test sets.

Figure 1.

Final pruned decision-tree cytokine model for predicting adverse event (AE) status. Cytokines identified by the unweighted voting filter (stem cell factor, monokine induced by interferon-γ, tissue inhibitor of metalloproteinases [TIMP-2], granulocyte colony-stimulating factor [G-CSF], intercellular adhesion molecule–1 [ICAM-1], and eotaxin [EOT]) were selected to train the decision-tree classifier. Input (ovals) for the if-then rules is the percentage change of the subject's cytokine level during the acute phase relative to the baseline cytokine level. On the basis of the value of the input, the inequalities guide the decision of which branch to follow. Given an individual's cytokine profile, one follows the decision branches downward to 1 of the 6 terminal nodes (AE and non-AE boxes). An individual is predicted to be classified as having an AE or non-AE status, depending on which inequalities are satisfied: ICAM-1 ≤11% (non-AE), EOT ≤−10% (non-AE), G-CSF >97% and TIMP-2 ≤51% (AE), G-CSF >97% and TIMP-2 >51% (non-AE), EOT >−10% and TIMP-2 ≤37% (non-AE), or EOT >−10% and TIMP-2 >37% (AE). The misclassification rates are given in parentheses below each terminal node.

RESULTS

The relationship between systemic cytokine expression, lesion size, and adverse events

A high rate of adverse events has been reported in subjects receiving smallpox vaccines [14], but the underlying mechanisms are not well understood. We found significantly larger diameters of induration at the inoculation site for the 17 subjects who exhibited fever and/or lymphadenopathy, compared with individuals who did not experience an adverse event (P = .01). The larger local induration diameters in subjects experiencing systemic adverse events suggested a role for inflammatory cytokines in the development of adverse events, a hypothesis that is corroborated by our cytokine analysis.

To study systemic alterations in cytokine concentrations in subjects receiving the APSV, we measured the serum expression of 108 cytokines in vaccinia-naive healthy adults. Evaluation was performed at 2 separate time points: at prevaccination (baseline) and 1 week after vaccination (acute). We reasoned that systemic adverse events, including fever, generalized rash, and lymphadenopathy, were likely to be associated with inflammatory mediators. Thus, it was of interest to determine whether systemic alterations of circulating cytokines are associated with these adverse events. In our clinical study, systemic adverse events were reported in 22 subjects. Vaccination was not associated with any serious adverse events, which were defined by a need to visit the clinic or emergency department or the need for hospitalization related to vaccination. When viewed as 2 separate cohorts, subjects without a reported adverse event (52 subjects) exhibited serum cytokine profiles that were significantly different from those of subjects with a reported systemic adverse event (22 subjects).

By using the unweighted voting committee method described above, we found the following 6 adverse event–discriminating cytokines: stem cell factor, monokine induced by interferon-γ, tissue inhibitor of metalloproteinases–2 (TIMP-2), granulocyte colony–stimulating factor, intercellular adhesion molecule–1 (ICAM-1), and eotaxin. The feature-selection results are summarized in table 2, where the cytokines are ordered by their Wilcoxon P values. Interestingly, our feature selection results suggest that one cannot simply remove the FDR fraction of cytokines with the largest P values to avoid spuriously associating a cytokine with adverse event status. Because we spec-ified the FDR to be q = 0.3, the FDR procedure guarantees that, at most, 30% of the 8 cytokines selected in the FDR column of table 2 are false discoveries. However, the FDR method does not guarantee that the spurious associations are the 30% with the largest P values. The committee method helps isolate spurious associations, and this method eliminated 3 of the 8 cytokines selected by the FDR procedure (see table 2).

Table 2.

Cytokines listed in the first column were found to discriminate between adverse event (AE) and non-AE individuals by at least 1 of the 3 following statistical methods: false discovery rate (FDR) correction to the Wilcoxon rank-sum test, nearest shrunken centroids (NSC), and support vector machines (SVM).

Gene symbol	FDR method	NSC method	SVM method	Wilcoxon P	Change in the AE cohort, mean %	Change in the non-AE cohort, mean %
ICAM-1	X	X		.0013	37.2	17.8
G-CSF (CSF3)	X		X	.0029	994.5	48.3
TIMP-2	X	X	X	.0054	27.5	10.0
IL-10	X			.0124	378.0	−2.9
MIG (CXCL9)	X	X	X	.0128	53.2	19.4
ALCAM	X			.0151	21.2	10.1
SCF	X	X	X	.0166	19.4	−12.7
MPIF1 (CCL23)		X		.0274	75.8	36.2
Eotaxin		X	X	.0463	4.0	−6.6
IL-4			X	.0476	21.8	7.8
IL-8		X		.0577	12.4	−3.7
NTF3		X		.0990	17.8	−1.3

NOTE. For each cytokine row listed, an “X” in the FDR, NSC , and/or SVM method columns indicates that the cytokine was selected by the corresponding method. Cytokines identified by consensus of at least 2 of the 3 statistical methods are highlighted in bold. These highlighted cytokines were used to build a final decision-tree model (see figure 1). The cytokines above are ordered by Wilcoxon rank-sum P values. The 2 right-most columns show the relative percentage change from baseline, averaged across the AE and non-AE cohorts, respectively.

Using 10-fold cross-validation, we estimated that the decision-tree classifier yields a 77% classification accuracy, true-positive and false-positive rates of 83% and 36%, respectively, for the adverse event group and true-positive and false-positive rates of 64% and 17%, respectively, for the non–adverse event group. To obtain a descriptive, interpretive model of the functional relationship between the set of cytokines selected by our cross-validated committee method, the final decision-tree model in figure 1 was trained on the full data. The final model includes the cytokines ICAM-1, granulocyte colony–stimulating factor, eotaxin, and TIMP-2.

DISCUSSION

Few studies have examined serum cytokine expression in humans associated with vaccination, and comprehensive data regarding systemic cytokine concentrations after smallpox immunization are otherwise not available. Studies that examine alterations in cytokine profiles in the systemic compartment after smallpox immunization may provide important information about the mechanisms leading to the successful control of poxvirus infections and may elucidate the pathophysiology of vaccination-associated adverse events at all levels of severity. The aim of this study was to identify patterns in serum cytokine expression following vaccination and to determine whether cytokine dysregulation is associated with systemic adverse events following smallpox immunization.

We have shown previously that almost all subjects with vesicle formation exhibit strong vaccinia virus–specific cytotoxic T lymphocyte responses and increased numbers of interferon-γ– producing T cells following vaccination with APSV [3]. Our findings suggested that, regardless of the vaccine dose, brisk T cell and humoral responses are induced if a vesicle forms. Importantly, all subjects enrolled in the APSV clinical trial at our site developed clinically observable lesions at the site of inoculation. It is possible that production of biologically significant amounts of additional cytokines occurred locally in our subjects but did not present dramatic increases in the systemic compartment. The majority of circulating cytokines are produced within tissues or lymphoid structures, from where they can spill into the bloodstream. Our aim was to identify serum cytokine patterns predictive of systemic adverse events, which are most likely caused by a systemic excess of cytokines. In addition, serum cytokine expression is measured more easily and is reproducible for rapid clinical diagnostic purposes.

In the present study, we precisely measured systemic concentrations of 108 cytokines and chemokines in serum samples prior to vaccination and 1 week after vaccination, using a sensitive protein microarray technique that incorporated rolling-circle amplification technology [6-12, 21]. To extract a useful subset of cytokines that discriminates between subjects who experienced at least 1 systemic adverse event (fever, lymphadenopathy, or generalized rash) and subjects who did not experience an adverse event, we used the following 3 independent class comparison methods: FDR-Wilcoxon, SVM, and NSC. A cytokine was selected for building the final decision-tree model if it was identified by at least 2 of the 3 methods. Decision trees were used to derive a descriptive, interpretable model of the functional relationships between the 6 selected cytokines and adverse event status. It should be noted that these serum cytokine and chemokine expression levels were studied soon after vaccination, well before the maximal local lesion size was reached or most adverse events had occurred. Therefore, the model could be considered to be predictive of subsequent adverse events. Such profiling in future studies—at time points soon after immunization—may be useful in predicting the risk of experiencing an adverse event.

Of the 6 cytokines selected, 3 are in the proinflammatory IL-17 signaling pathway. In this pathway, fibroblasts, stimulated by IL-17, are induced to secrete inflammatory and hematopoietic cytokines, including granulocyte colony–stimulating factor, stem cell factor (both identified in our committee method), and IL-8 (CXCL8 identified by the NSC method). These cytokines provoke a range of activities, including neutrophil proliferation and differentiation. IL-17 has been shown to enhance cell surface expression of the endothelial cell adhesion molecule ICAM-1 on human fibroblasts [22]. In turn, increased expression of ICAM-1 was shown to aid in T cell recruitment during contact hypersensitivity (related to delayed-type hypersensitivity) [23]. In the present study, soluble ICAM-1 was a strong discriminator of smallpox vaccine–related adverse event status. In fact, ICAM-1 was the root node of the decision-tree model in figure 1, meaning that the effect of the other cytokines on adverse event status was conditionally dependent on ICAM-1. Although key cytokines in the IL-17 pathway play an important role in our analysis, IL-17 itself was not found to be differentially expressed at a statistically significant level. This may be because of our choice of cytokine measurement time point. We chose a time point that captures the peak concentration for most cytokines, but this time point may not be representative for all cytokines relevant to adverse events.

Eotaxin, a chemokine ligand for CCR3 that activates and recruits eosinophils to the site of inflammation and stimulates macrophage activation, was also 1 of the 6 discriminatory cytokines. Activated eosinophils can release reactive oxygen species, causing tissue damage during chronic inflammatory responses. The molecule monokine induced by interferon-γ (Mig/CXCL9), a member of the C-X-C subfamily of chemokines and an attractor of activated T cells expressing chemokine receptor CXCR3 [24], was found by our committee analysis. Interferon-γ–induced monokine is produced by macrophages and may play a critical role in enhancing the recruitment and activation of T cells [25].

TIMP-2 was found to be associated with adverse events by all 3 statistical learning methods, and its effect in the decision tree (figure 1) suggests that it plays a complex role in the development of adverse events. TIMP-2 appears in 2 branches of the decision tree, and it has a polar effect on the prediction of adverse event status, depending on the context of the other cytokines in the tree. Proteins in the TIMP family are natural inhibitors of the matrix metalloproteinases, a group of peptidases involved in degradation of extracellular matrix. TIMP-2 accelerates wound healing by enhancing the proliferation of epidermal keratinocytes and dermal fibroblasts. Following the branches toward TIMP-2 on the right shows that, when an individual's increase in TIMP-2 expression is <51%, then that individual experiences an adverse event. This is presumably because the balance between the MMP and its inhibitor tips toward the MMP causing excess extracellular matrix destruction and further inflammation. However, this situation only occurs when granulocyte colony–stimulating factor expression is substantially increased, compared with the baseline level. When the increase in granulocyte colony–stimulating factor is <97% and the expression of eotaxin does not decrease by >10%, then the role of TIMP-2 is qualitatively different. Hence, the expression of TIMP-2 must be taken in the context of the other cytokines to best predict adverse event status. This finding demonstrates one of the challenges of complex molecular investigations in clinical populations—that it is often necessary to take into account statistical interactions between variables when testing associations with a phenotype. It is possible that the right-hand branch of the decision tree may be the result of over-fitting; however, this type of complex TIMP-2 behavior has been reported elsewhere [26].

The results of our protein microarray analysis indicate a cytokine signature for the pathogenesis of adverse events involving stem cell factor, monokine induced by interferon-γ, TIMP-2, granulocyte colony–stimulating factor, ICAM-1, and eotaxin. This signature suggests that the development of adverse events involves excess stimulation of inflammatory pathways and the imbalance of tissue damage repair pathways. Figure 2 shows a smallpox vaccine adverse-event model of interactions involving the soluble cytokines found by this protein microarray. The initial local tissue injury in subjects experiencing adverse events following vaccination appears to trigger an acute inflammatory response not unlike a delayed-type hypersensitivity reaction. During the elicitation phase of delayed-type hypersensitivity, antigen presentation to Th1 cells in the dermis leads to the release of T-cell cytokines such as interferon-γ and IL-17 [27, 28]. A cascade of cytokines and chemokines is then released, enhancing the inflammatory response by inducing the migration of monocytes into the lesion and their maturation into macrophages and by further attracting T cells. The dominant cytokine responses in the systemic compartment were associated with robust macrophage activation. The role of inflammatory cytokines in adverse event development, coupled with our previous work demonstrating the role of T cell–derived factors and the similarities of systemic adverse events observed after smallpox vaccination with the clinical presentation of macrophage activation syndrome [29], suggest that systemic adverse events following smallpox vaccination may be consistent with low-grade macrophage activation syndrome caused by virus replication and vigorous tissue injury and repair.

Figure 2.

Model of early interactions involving soluble cytokines most associated with subsequent adverse events following vaccination. Tissue inhibitor of metalloproteinases (TIMP-2) modulates wound healing in response to tissue damage at the inoculation site and enhances dermal fibroblast proliferation. During the inflammatory response, IL-17 is secreted by T cells recruited by monokine induced by interferon-γ (MIG) and intercellular adhesion molecule-1 (ICAM-1). Stimulated by IL-17, fibroblasts secrete inflammatory and hematopoietic granulocyte colony-stimulating factor (G-CSF) and stem cell factor (SCF) and increase the surface expression of ICAM-1 and the production of eotaxin. Eotaxin and MIG stimulate macrophage activation. The excess local secretion of these soluble cytokines leads to their remote diffusion and later detection in the blood stream.