A transcriptomic reporter assay employing neutrophils to measure immunogenic activity of septic patients’ plasma

The study was approved by the ethical review committees of Khon Kaen University and Khon Kaen Regional Hospital (Khon Kaen, Thailand) and the Institutional Review Board of Benaroya Research Institute (Seattle, WA). Participants provided written informed consent to participate in this study. Written informed consent was obtained from parents or guardians on behalf of the minor/child participant in this study.

Septic patients were enrolled from Khon Kaen Regional Hospital, Khon Kaen, Thailand. Patients who met at least two of the criteria for severe inflammatory response syndrome (SIRS) were enrolled in the study[3, 15]. As part of the routine investigations, clinical specimens were collected for bacterial culture within 24 h following SIRS diagnosis. Only blood samples obtained from patients who were retrospectively diagnosed with culture-proven sepsis were retained for further analyses. Patients with negative blood cultures were excluded. Severe sepsis was defined according to the current guidelines from the Surviving Sepsis Campaign[15]. These criteria include several clinical and laboratory findings of sepsis-induced tissue hypoperfusion or organ dysfunction. We used the subset of these criteria for which the necessary data had been collected for our patient cohort: elevation in creatinine to >2.0 mg/dl, elevation in bilirubin to >2.0 mg/dl, platelet count <100,000/μL, and sepsis induced hypotension (septic shock). In hospital death was also counted as severe sepsis. Demographic and clinical data were recorded for all subjects (Additional files1,2, and3). Uninfected healthy controls were selected as individuals who had no signs of acute infectious diseases during the previous 3 months or at the time of the study. Uninfected controls also had to have normal blood counts, normal fasting blood glucose, and normal glycosylated hemoglobin. Three milliliters of whole blood was collected from each patient and healthy control into heparinized tubes (BD Biosciences). For sepsis patients, samples were grouped as drawn either in the first 48 h of admission, or at >48 h after admission. To separate plasma, blood samples were centrifuged at 2,000 rpm for 10 minutes and the plasma component was transferred into a cryogenic vial and stored at -80°C until used.

Blood samples from three additional healthy volunteers were used in subsequent cell isolation procedures. PMNs were isolated from heparinized venous blood by 3.0% dextran T-500 sedimentation and Ficoll-Paque PLUS centrifugation (Amersham Biosciences) as previously described[21]. The purity of isolated cells was generally more than 95% as determined by flow cytometry (FACSCalibur, Becton Dickinson)[22]. PBMCs were isolated from whole blood samples by centrifugation through a Ficoll-Paque Plus (Sigma Aldrich) density gradient.

A portion of the isolated PBMCs was subsequently used for MoDCs generation as previously described[22, 23]. MoDCs were harvested and resuspended in serum-free RPMI-1640 medium (Gibco), 5 × 10⁵ cells/well were plated into a 24-well tissue culture plate (Corning) for 24 h. The resulting cells were determined to be >95% CD11c⁺ by flow cytometry.

Cell cultures were performed as described by Pascual et al.[7]. Two million PMNs or one million PBMCs were resuspended in serum-free RPMI-1640 medium (Gibco) and added to either 5 ml or 2 ml culture tubes (Becton Dickinson), respectively. Five hundred thousand MoDCs were seeded into 24 well tissue culture plates (Corning) at 1 × 10⁶ cells/ml and rested for 24 h before the experiments. Cells were cultured with medium alone or a plasma sample in a final concentration of 20%. After 6 h incubation at 37°C in 5% CO₂, cells were harvested, washed twice with phosphate buffered saline, homogenized in RLT buffer (RNeasy mini kit; QIAGEN), and stored at -80°C until use.

Total RNA was isolated using the RNeasy Mini kit (QIAGEN) according to the manufacturer’s instructions. RNA integrity number (RIN) was determined by using an Agilent 2100 Bioanalyzer (Agilent). Qualified samples (RIN >6 or presence of 28s and 18s rRNA) were retained for further processing. Total RNA was amplified and labeled using the Illumina TotalPrep RNA Amplification Kit (Ambion). Labeled cRNA was hybridized overnight to Human HT-12 V4 BeadChip array (IIlumina), washed, blocked, stained and scanned on an Illumina HiScan instrument following the manufacturer’s protocols.

GenomeStudio was used to generate signal intensity values from the scans and perform background subtraction. Post-hybridization quality controls were done by the standard metrics provided by the manufacturer. Data from each cell type and each culture experiment were processed independently. All possible outliers were excluded from the expression data set by metrics for post-hybridization quality controls.

All data analyses were performed using R (version 2.14.0; http://​cran.​r-project.​org/​bin/​windows/​base/​old/​2.​14.​0/​). Data pre-processing of background subtracted data was performed by using the preprocessCore package from Bioconductor. Pre-processing included rescaling intensity by quantile normalization. After normalization, expressions were floored with intensities <10 set to 10. Transcripts with detection p-value of less than or equal to 0.01 in at least one sample (PALO) were selected for further analysis. Samples from the same cell type and batch were normalized to the average intensity of samples cultured in medium alone. A filter was set to include only transcripts that had at least two-fold changes and 100 intensity differences compared to medium control. Background subtracted and processed data from these experiments have been deposited at NCBI’s Gene Expression Omnibus database (http://​www.​ncbi.​nlm.​nih.​gov/​geo/​), with accession numbers GSE49758. To facilitate data sharing and interactive data analysis, we created a data portal (https://​gxb.​benaroyaresearch​.​org/​tra/​tra-paper/​tra-landing.​gsp) to store and analyze background subtracted data from all three experiments (see[24] for tutorial).

Transcripts that were differentially expressed between study groups were selected by Random k-Nearest Neighbor – Feature Selection (RKNN-FS) using the R package “rknn”[25]. A RKNN classifier consists of an ensemble of base k-nearest neighbor models (number of neighbors = 5), each constructed from a random subset of the input variables. Features were selected by ranking the importance of the PALO transcripts.

Gene ontology analyses were performed using GeneGo MetaCore pathway analysis tool (Thomson Reuters, NY). The default background gene list was used for all enrichment analyses including process networks, pathway map folders, and pathway maps. Pathway maps are the collection of pathways grouped into folders according to main cell processes, protein functions, and diseases. Map Folders are the collection of pathways grouped into folders according to main biological processes. Statistical significance was ascertained by using a threshold of false discover rate (FDR) <0.05. The network builder tool using the shortest possible path with no more than 2 steps was used to represent functional interactions. Upstream transcription factors were identified for lists of over-expressed genes using the Transcription Factors tool.

To determine the performance of our predictive signature, class prediction of binary variables was carried out by support vector machine (SVM; package “e1071”) and random forest (RF; package “randomForest”) algorithms. These machine learning methods are robust, well-accepted and commonly used methods for class prediction. Receiver operating characteristic (ROC) curves were constructed using the R package “ROCR”. Area under the curve (AUC) and confidence intervals were calculated using the R package “Hmisc”.

We first set out to determine which cell reporter system would be most amenable to detect meaningful changes in response to septic plasma in vitro. PMNs, PBMCs and monocytes were isolated from two healthy donors (Table 1). Monocytes were cultured with interleukin 4 (IL4) and granulocyte-macrophage colony-stimulating factor (GM-CSF) to generate MoDCs. Plasma samples were obtained from a first set of patients with culture-confirmed sepsis (n = 12) and from uninfected healthy controls (n = 6). Cells were cultured for 6 hours in medium alone (unstimulated) or in the presence of plasma (stimulated) using a final concentration of 20% (Figure 1). Microarray data were generated to assess transcriptional responses on a genome-wide scale. PALO filtering returned 21,236, 25,728, and 23,589 transcripts in PMNs, PBMCs, and MoDCs respectively. Transcripts were selected that changed by at least 2-fold and 100 intensity differences (2FC100DIF) in response to plasma stimulation compared to unstimulated cells. To reduce data dimensionality and facilitate the comparison among cell types, principal component analyses (PCA) were carried out. PCA plots of data obtained from the 3 different reporter cell systems (PMNs, PBMCs, and MoDCs) show changes in transcription that can be attributed to responses to plasma stimulation (Figure 2). The best separation between sepsis and uninfected plasma samples was observed in PMNs (Figure 2, top left panel). This result was confirmed by calculating the Euclidean distance from the center of uninfected controls (Additional file4). These results demonstrate the feasibility of a “reverse proteomics” approach to detect the presence of immunomodulatory factors in the blood of sepsis patients. While responses measured in PMNs and PBMCs were consistent for both cell donors (Figure 2, upper and middle right panels), significant donor-to-donor variation was observed in MoDCs (Figure 2, lower right panel).

To test how our approach could best be used for monitoring immunomodulatory factors in the blood of sepsis patients, we next compared the transcriptional programs elicited by septic plasma in PMNs, PBMCs, and MoDCs. Septic plasma samples (n = 6/24) that were found to induce the most robust responses consistently across the three cell reporter systems were selected (Figure 2, Additional file5). Transcripts changing by at least 2-fold compared to their own unstimulated controls in each reporter system (709, 366 and 452 transcripts for PMNs, PBMCs and MoDCs, respectively) were combined to generate a list of 1,366 transcripts (Additional files6 and7). The number of transcripts passing this 2-fold cut off was the highest in PMNs and the majority of these transcripts are over-expressed (Table 2). In contrast, there were fewer changes recorded in PBMCs and MoDCs and for the majority of those genes, transcript abundance decreased upon exposure to septic plasma.

In an effort to determine potential molecules present in septic plasma inducing these patterns of gene expression, we selected those genes over-expressed in each of the 3 reporter cell types and predicted the upstream transcription factors using GeneGo MetaCore. One hundred and sixty-one transcription factors were identified in total (Additional file8). There was significant overlap among the 3 lists of transcription factors with 27 predicted for all 3 cell types, including several well established in core immune pathways such as STAT3, STAT4, STAT5A, STAT5B, and CREB1. These findings are consistent with the idea that a relatively small number of molecules are driving the observed transcriptional responses and could suggest future studies to determine the relevant immunomodulatory molecules in septic plasma.

To better understand the observed gene expression patterns and the functions of those genes, a heatmap was generated with the 1,366 transcripts differentially expressed in at least one cell type (Figure 3). We performed hierarchical clustering and further characterized the function of transcripts in each of the resulting clusters by using the GeneGo MetaCore pathway analysis tool. Networks with statistically significant enrichment of functional annotations were found in 5 out of 12 clusters. Most of the enriched functional networks are relevant to immune processes. Cluster 1 (n = 267), the largest cluster, is formed by transcripts over-expressed in PMNs. This cluster contains many genes involved in neutrophil function (e.g. CD177, IL1R2, NLRP3, TNFSF8, FCGR1A, FCGR1B, and FCGR1C). Cluster 3 (n = 84) contains transcripts over-expressed in PBMCs, MoDCs and to a lesser extent PMNs. This cluster contains genes enriched for cell cycle and apoptotic mechanisms (e.g. CDKN1A, GADD45A, GAA45G, SMAD7, etc). Cluster 5 (n = 39) is over-expressed in PMNs but under-expressed in PBMCs. This is a smaller cluster but it contains several important immune-related transcripts (e.g. CCL2, CCL20, CXCL2, IL1RA, IL1A, IRAK3, and TLR2). Innate immune pathways and chemotaxis are the major functions of the genes in this cluster; specific induction of these genes in PMNs is consistent with their known role as the immune cells that respond immediately to bacterial infection. Cluster 6 (n = 135) is under-expressed in PBMCs and is enriched in immune mediators involved in chemotaxis (e.g. CCL3, CCL3L1, CXCL5, and CXCL16), IFN-gamma signaling (e.g. CASP5, FCAR and CFB), and proliferation and morphogenesis (e.g. MMP19 and NRP1). Lastly, cluster 7 (n = 178) is under-expressed in all reporter systems. The enriched functional networks for this cluster are relevant to both innate and adaptive immune responses, for example, antigen presentation (e.g. CD58 and STAT1), and neutrophil activation (IL6, NFKB2, and CXCL1). Taken together, these results support the use of PMNs as a useful reporter cell type for sensing immunomodulatory constituents of septic plasma in our reverse proteomic assay system.

Next we sought to further investigate factors that account for the heterogeneity of the responses observed in PMNs exposed to plasma from the initial cohort of septic patients. Plasma samples from a larger independent set of septic (n = 29) and uninfected (n = 18) subjects were tested on PMNs collected from one of the cell donors used in our first experiment. PALO filtering returned 21,374 transcripts. The results were similar to that of the first experiment in both quantitative and qualitative aspects. The first two principal components, PC1 and PC2, explained more than 50% of the variability, and a distinct response to septic plasma compared with uninfected plasma was observed (Figure 4A, left panel). However, there was significant heterogeneity observed among the samples stimulated with septic plasma.

Further analyses identified 2 parameters accounting for much of the observed heterogeneity among samples stimulated with septic plasma – sepsis severity and timing of sample collection. To stratify severity, patient samples were categorized into “severe” or “not severe” sepsis according to predefined criteria (see Methods). In contrast to patients with non-severe sepsis, patients with severe sepsis were markedly separated from uninfected controls. This result was confirmed by calculating the Euclidean distance from the center of uninfected controls (ellipse in Figure 4A, right panel). The second important explanatory parameter identified was the timing of sample collection from hospital admission. There was no difference between transcriptional responses stimulated by septic plasma samples collected more than 48 h after admission and uninfected controls, whereas there was significant difference between septic plasma samples collected less than 48 h after admission and controls suggesting potential resolution of observed immunologic changes after treatment initiation. Numerous other demographic and clinical variables were investigated which did not appear to explain the observed heterogeneity (Additional file9). The fact that most variability could be accounted for by sepsis severity suggests this is the primary factor determining the PMN response to plasma.

Since disease severity impacted the magnitude of transcriptional response to septic plasma, we evaluated the potential value of our PMN reporter system for clinical applications. We conducted a third experiment using PMNs from the same donor as in a second experiment with an independent set of septic plasma samples all collected within 48 h of admission (n = 35) and uninfected control samples (n = 19). PALO filtering returned 15,083 transcripts. The PCA plots generated from these transcripts showed similar results to that obtained in the second experiment confirming the marked separation of severe sepsis from not severe sepsis and controls and demonstrating sepsis severity the main variable affecting transcriptional responses (Figure 4B, Additional file10). To assess variation of the cell donor source, we tested the same set of plasma samples with PMNs from an additional donor not used in either of the previous two experiments. PCA plots and Euclidean distance comparisons demonstrate similar findings as with the first donor suggesting this sort of PMN reporter system can be reproducible independent of the cell donor (Additional file11).

We then used a RKNN-FS algorithm to identify markers for differentiating severe and not severe sepsis (Table 3). Data from the third experiment were used as the training set because it was the most balanced and largest dataset (severe sepsis n = 20; not severe sepsis n = 15). RKNN-FS identified a set of 30 transcripts as being able to provide the highest accuracy (82.02%) in predicting sample class in a leave one out cross-validation scheme. This biomarker signature reflects different amplitudes of PMN responses to septic plasma samples according to disease severity. Misclassification was observed only with some non-severe sepsis and uninfected controls. This result is due to the more similar transcriptional response between these two groups as seen in Figure 4. To illustrate the performance of the identified markers for predicting both severe and non-severe sepsis from uninfected controls by a ROC curve, we built prediction models by using SVM and RF algorithms. Data from experiment II (severe sepsis n = 17; not severe sepsis n = 12) were used as a test set. The results demonstrated the high accuracy of the 30-transcript panel in predicting the response to plasma from severe sepsis samples; AUCs are 1 for both SVM and RF. The prediction accuracies of response to plasma from non-severe sepsis samples are slightly lower; the AUCs are 0.82 (95% confidence interval, 0.78 – 1.00) and 0.90 (95% confidence interval, 0.66 – 0.99) for SVM and RF, respectively (Figure 5). Meaningful validation could not be done using data from experiment I because almost all of these plasma samples (11/12) were obtained from patients with severe sepsis, 3 of which were collected at >48 h after admission.

Enrichment analyses were performed to characterize the functional relevance of our severity signature panel in GeneGo MetaCore. Collectively, these genes are involved in immune system response as their major biologic function (Figure 6A). Analyses for significant cellular process (Pathway Maps) also suggested their roles in cell cycle regulation (Figure 6B). The genes participating in these functions can be visualized on an interaction network (Figure 6C).

In addition to these enrichment analyses, many of our 30 classifier biomarkers have been previously shown to have potential roles in several dysregulated inflammatory conditions including sepsis and secondary organ dysfunction. Most of the 11 over-expressed transcripts have been independently identified as being of potential value as severity biomarkers in such inflammatory conditions. FKBP5 has been shown to contribute to the regulation of myeloid-derived suppressor cells (MDSCs) and changes in MDSCs have been shown important in a sepsis model[26, 27]. TLR2 has been shown to be overexpressed in septic shock and also to play a role in acute kidney injury, which is an important consequence of severe sepsis and was part of our severity criteria[28, 29]. TPST1 over-expression can be induced by lipopolysaccharides (LPS) and inhibition of TPST1 has been shown to affect macrophage signaling suggesting its role in the innate immune response[30]. ECHDC3 over-expression was reported in onset of acute coronary syndrome[31]. IL18R1 and IL1R2 belong to the IL1R family. IL18R1 plays a role in neutrophil migration and activation and has been identified as a biomarker for several systemic inflammatory conditions, such as surgery-induced inflammation and bacterial meningitis[32‐34]. IL1R2 expression has been suggested to be a marker of sepsis and high circulating IL1R2 protein levels have been reported in critically-ill patients with sepsis[35, 36].

The functional relevance of the 19 under-expressed biomarkers was also investigated. MMP9 has been shown to have negative correlation with multiple organ dysfunction scores in sepsis[37]. FCGR2B deficient mice have been shown to have increased bacterial clearance and survival in sepsis[38]. TNFRSF9 is thought to be down-regulated by TNF-α, and to enhance anti-apoptosis and subsequently induce inflammatory responses[39, 40]. CCL22 is involved in regulating immune response by recruiting T-helper cells and regulatory T cells and a sepsis model demonstrated CCL22 playing a role in enhancing neutrophil activation and chemotaxis[41‐43]. P2RY2 has been shown to play a role in neutrophil chemotaxis in a mouse model and is also involved in the recruitment of PMNs to the lung resulting in acute lung injury in sepsis[44, 45].

Several other genes in our severity panel are involved in cell cycle and proliferation, for example, KLF9, CCND3, YIPF5, ALG10B, and ARID5A. Taken together this literature shows well-established pathogenic roles in sepsis or other inflammatory conditions for much of our candidate biomarker panel and constitutes an external validation of our reporter assay system.