Two New Neutrophil Subsets Define a Discriminating Sepsis Signature

Abstract

Rationale

Sepsis is the leading cause of death in adult ICUs. At present, sepsis diagnosis relies on nonspecific clinical features. It could transform clinical care to have immune-cell biomarkers that could predict sepsis diagnosis and guide treatment. For decades, neutrophil phenotypes have been studied in sepsis, but a diagnostic cell subset has yet to be identified.

Objectives

To identify an early, specific immune signature of sepsis severity that does not overlap with other inflammatory biomarkers and that distinguishes patients with sepsis from those with noninfectious inflammatory syndrome.

Methods

Mass cytometry combined with computational high-dimensional data analysis was used to measure 42 markers on whole-blood immune cells from patients with sepsis and control subjects and to automatically and comprehensively characterize circulating immune cells, which enables identification of novel, disease-specific cellular signatures.

Measurements and Main Results

Unsupervised analysis of high-dimensional mass cytometry data characterized previously unappreciated heterogeneity within the CD64⁺ immature neutrophils and revealed two new subsets distinguished by CD123 and PD-L1 (programmed death ligand 1) expression. These immature neutrophils exhibited diminished activation and phagocytosis functions. The proportion of CD123-expressing neutrophils correlated with clinical severity.

Conclusions

This study showed that these two new neutrophil subsets were specific to sepsis and detectable through routine flow cytometry by using seven markers. The demonstration here that a simple blood test distinguishes sepsis from other inflammatory conditions represents a key biological milestone that can be immediately translated into improvements in patient care.

At a Glance Commentary

Scientific Knowledge on the Subject

The diagnosis of sepsis remains a difficult exercise combining clinical scoring systems and the combination of multiple biomarkers. Still, many patients with suspected sepsis suffer from delayed diagnosis and may be unnecessarily treated with antibiotics. There is an unmet need for specific and rapid diagnostic tests for sepsis, which would differentiate patients with sepsis from patients with aseptic inflammation. Biomarkers expressed on the neutrophil cell surface provide new perspectives for disease diagnosis, disease severity, and prognosis.

What This Study Adds to the Field

This work represents the first comprehensive evaluation of whole-blood circulating immune cells in patients with sepsis to use cytometry by time-of-flight, high-dimensional technology coupled with computational analysis. It allowed the identification of two novel sepsis-specific neutrophil subsets: CD10⁻CD64⁺PD-L1⁺ neutrophils and CD10⁻CD64⁺CD16^low/−CD123⁺ immature neutrophils. Immunophenotyping immature neutrophil subsets by using routine flow cytometry could be used for early identification of sepsis in patients.

Sepsis is the leading cause of death in the ICU (1–3). The diagnosis of sepsis in patients relies on clinical data rather than on a robust biomarker that distinguishes sepsis from sterile inflammation and predicts its clinical outcome, and the prognosis can be evaluated by using several scores, including the Simplified Acute Physiology Score II (SAPS II) and Sequential Organ Failure Assessment (SOFA) scores. The SOFA and SAPS II scores are indicators of severity, show poor performance regarding sepsis diagnosis, and were consistently shown to be nonspecific to sepsis (1, 4–7). It is estimated that the survival rate decreases by roughly 10% every hour that appropriate antimicrobial medication is delayed, emphasizing the urgent need for early diagnosis techniques (8, 9). A comprehensive system immunology approach using mass cytometry is well suited to characterize the diversity of disease-specific cellular states (10). Neutrophils are a primary immune cellular barrier against pathogens, but they may be a double-edged sword in sepsis, as they have a role in both inflammation and immunosuppression (11–15). We hypothesized that phenotype of circulating neutrophils might provide crucial early insight into immune features that drive sepsis and distinguish this disease from noninfectious inflammatory syndrome.

For the system immunology approach here, it was critical to track features that had been identified as important in sepsis biology but that did not individually have the resolving power to specifically distinguish sepsis. Neutrophils expressing the high-affinity immunoglobulin–Fc receptor I (CD64) were described in numerous clinical studies over the last 2 decades (16). CD64 is normally expressed on monocytes, but its expression on circulating neutrophils could be due to its upregulation during inflammation (17) or to released immature granulocytes from the bone marrow (BM), especially when it is associated with decreased expression of neutral endopeptidase (CD10) and low-affinity immunoglobulin–Fc fragment III (CD16) (13, 14, 18) (see Table E1 in the online supplement) (19). Previous studies also identified IL-3 as an orchestrator of emergency myelopoiesis during sepsis and showed its association with hospital mortality (20, 21). In parallel, PD-L1 (programmed death ligand 1) expressed on monocytes was also described as a mortality predictor in patients with sepsis (22, 23).

A system-level view is likely needed to identify cellular features that specifically distinguish sepsis infection–induced immune phenotypes from those triggered by aseptic inflammatory signals. To identify such early sepsis-specific cellular biomarkers, we developed a multiparametric immune profiling strategy (Figure 1). A cytometry by time-of-flight instrument was used to measure 42 markers on whole-blood immune cells from patients with sepsis and control subjects (Figure 1A) (24). A computational analysis approach was used to comprehensively characterize circulating immune cells and identify disease-specific cellular signatures (25, 26). This approach consisted of a “discovery strategy” (Figure 1B) and a computational “validation strategy” (Figure 1C) based on two complementary sets of algorithms. We identified two unreported early and sepsis-specific neutrophil subsets. A conventional “expert-driven strategy” using a limited set of markers confirmed that these two sepsis-specific neutrophil subsets were associated with sepsis (Figure 1D). This result was confirmed by using an independent cohort of patients and conventional flow cytometry (Figure 1E). Some of the results of these studies have been previously reported in the form of a preprint (https://doi.org/10.1101/2020.05.29.123992).

Figure 1.

Study design. (A) Blood samples from patients with sepsis (S) (n = 17) or NIC, post–cardiothoracic surgery patients with inflammation (n = 12) were enrolled in the discovery cohort of the study; in addition, blood samples were obtained from HDs (n = 11), and BM biopsy specimens were obtained from orthopedic surgery patients (n = 5). Immunostaining targeting 42 parameters was performed, and results were analyzed by using mass cytometry. (B) A computational “discovery strategy” was used to identify S-specific subsets, (C) a “computational validation” analysis was used to check whether the identified S-specific subsets were strategy-dependent, and (D) an additional “expert-driven validation” was used to define a small set of markers to gate on the S-specific neutrophil subsets. (E) A second independent validation cohort, including patients with S (n = 24) and NIPs (n = 18), was used for the “biological validation” of these S-specific neutrophil subsets through conventional flow cytometry. BM = bone marrow; HD = healthy donor; NIC = noninfected control; NIP = noninfected patient; PD-L1 = programmed death ligand 1.

Methods

Study Design

This observational study was approved by the ethics committee (Comité de Protection des Personnes Ile-de-France VII A000142-53). Two cohorts were used in this study (Table E2). Seventeen patients with sepsis (S) and 12 patients undergoing cardiac surgery considered as noninfected control (NIC) subjects with inflammation were included in the discovery cohort of the study (Table E2). The validation cohort was composed of 24 patients with sepsis and 18 noninfected patients with confoundable symptoms of sepsis (NIP) (Table E2). Blood samples were drawn in heparin-coated tubes and were collected at the first and seventh day after admission from antibiotic-treated patients with sepsis, after surgery from NIC patients in the discovery cohort, and at the first day after admission from the patients in the validation cohort. In addition, blood samples of 11 age- and sex-matched healthy donors (HDs) were obtained from the French blood donation center. Five BM biopsy specimens from orthopedic surgery patients were also included in this study.

Mass Cytometry Analysis

Whole-blood samples were stained by using a 42-dimension mass cytometry panel (Table E3). A multistep staining protocol was set up and is detailed in the Methods section in the online supplement. Once the collection of samples was completed, stained cells were thawed and then measured on a Helios cytometry by time-of-flight instrument. Acquired data were normalized by using MATLAB-based software (MathWorks) (27) and were analyzed by using the Cytobank platform (28).

Computational Data Analysis

To identify immune subsets and visualize all cells in a two-dimensional map in which position represents local phenotypic similarity, we used two different dimensionality-reduction tools depending on the strategy: the Visualization of t-Distributed Stochastic Neighbor Embedding (viSNE) implementation of the t-Distributed Stochastic Neighbor Embedding (tSNE) algorithm (29) and the Uniform Manifold Approximation and Projection for Dimension Reduction (UMAP) algorithm (30). Cells were also grouped in phenotypically homogenous clusters by using either the Spanning-tree Progression Analysis of Density-normalized Events (SPADE) algorithm (31) or the flow or mass cytometry analysis algorithm using a Self-Organizing Map (FlowSOM) (32, 33). To phenotypically characterize these clusters, marker enrichment modeling (MEM) (34, 35) was used. The analysis process of each strategy is detailed in the Methods section in the online supplement.

Flow Cytometry Validation Panel

To validate the sepsis-specific neutrophil signature, a seven-marker panel (Table E4) was designed for conventional florescent flow cytometry. The samples from patients with sepsis, as well as those from noninfected patients, were analyzed by using blind cytometry testing. The staining protocol is detailed in the Methods section of the online supplement.

Activation and Phagocytosis Assay

To address neutrophil activation and phagocytic capacities, we used pHrodo-labeled BioParticles (Invitrogen) coated with Staphylococcus aureus or zymosan antigens. The staining protocol is detailed in the Methods section of the online supplement.

Statistical Information

Numerical data are given as the median and interquartile range (25th to 75th percentile), with the exception of activation and phagocytosis assay data, which are given as the mean ± SD. A nonparametric, two-tailed Mann-Whitney test with a significance threshold of α = 0.05 was used to compare the cellular abundance of cell subsets and the mean fluorescence intensity (MFI) ratios between two groups of patients. A nonparametric, two-tailed Wilcoxon signed rank test with a significance threshold of α = 0.05 was used to compare the cellular abundance of cell subsets from patients at Day 1 and Day 7. The relationship between two data sets was assessed by using the Spearman rank correlation coefficient (r) and test, with a significance threshold of α = 0.05, and a linear regression line was drawn on the corresponding plot. Statistical tests were performed by using GraphPad Prism version 7 software (GraphPad Software), as well as receiver operating characteristic (ROC) analyses.

Results

Mass Cytometry and Computational Analysis Revealed a Sepsis-Specific Neutrophil Signature

We designed a longitudinal observational study including 40 individuals to explore the evolution of the circulating immune-cell phenotypes of patients with sepsis (n = 17) and NIC patients (n = 12) at Days 1 and 7 (Table E2), HDs (n = 11), and BM biopsy specimens (n = 5) (Figure 1A). Whole-blood immunostaining was performed with a 42-parameter mass cytometry panel designed to give a comprehensive evaluation of circulating leukocytes (Figure 1A and Table E3). We identified circulating immune-cell populations. By using the viSNE tool, neutrophils were gated, and other circulating immune cells were independently analyzed.

The neutrophils were analyzed by using a discovery strategy that included use of the viSNE and SPADE tools (Figure 1B). viSNE is an unsupervised algorithm that reduces feature dimensions and allows cell visualization in a two-dimensional map. SPADE is an unsupervised algorithm that aims to group cells into nodes that can be displayed on the viSNE map. This strategy allowed us to define an imprint for each sample group (Figure 2A). On the resulting map, the Day 1 neutrophils of patients with sepsis and NIC patients and the neutrophils of HDs were arranged in three different areas (Figure 2A). The neutrophils of patients with sepsis were clustered in specific nodes that were absent from the neutrophils of NIC patients and HDs (Figures E1 and E2). Some of these sepsis-specific nodes were shared with the BM samples, suggesting the occurrence of myelocytosis for patients with sepsis (Figures E1 and E2). Most cells from Day 7 samples were phenotypically similar to those from HDs (Figures 2A, E1, and E2). CD16, CD10, and CD64 markers split the neutrophil signature into a positive subset and a negative subset for each marker (Figure E1). To characterize all the nodes, their abundance in each sample and their average expression of each marker were extracted and used to generate two heatmaps (Figures E3 and E4). Hierarchical clustering was used to the arrange rows (nodes) and columns (samples) of the frequency heatmap (Figure E3) and the columns (markers) of the phenotype heatmap (Figure E4). In this unsupervised, three-arm analysis (nodes, samples, and markers), the resulting dendrograms led to the identification of three main sample clusters (columns), as shown in Figure 2B (Figure E3 before tree cut). Most of the samples were clustered according to patient groups. Sepsis Day 1 and BM samples were clustered together. Sepsis Day 7 samples were split in two sample clusters, with half of them clustering with HD samples (Figures 2B and E3), suggesting the acquisition of a “healthy” neutrophil phenotype profile (Figure 2A). In addition, this unsupervised strategy allowed the precise delimitation of four groups of cell nodes (Figure 2D): 1) HD-abundant nodes representing neutrophils with a CD16^highCD10^medCD64⁻ phenotype, 2) nodes harboring a CD16⁺CD10^medCD64⁻ phenotype common to NIC and sepsis samples at Day 7, 3) nodes defined as being CD16^lowCD10⁻CD64^low that were common to NIC and sepsis samples at Day 1, and 4) Sepsis Day 1 and BM nodes with a CD10⁻CD64⁺ phenotype. Node group 4 represents cells that are highly abundant in sepsis samples at Day 1 when compared with samples from other patient groups (Figure E5A). The statistical analyses of these nodes are presented in Figure E5B. Among the nodes that statistically discriminate between sepsis and NIC samples at Day 1 (Figure E5B), a specific phenotypic characteristic was observed: three nodes expressed CD123, and four other nodes expressed PD-L1 (Figures 2C, 2D, and E5A). On the basis of phenotypic homogeneity, meta-clusters were generated to group nodes that share similar expression of these two markers and represent two neutrophil subsets that are specific to sepsis at Day 1 and are observed to be lacking in NIC neutrophils (Figure 2E). Subset 1 was composed of CD10⁻CD64⁺CD16⁺PD-L1⁺ neutrophils (sepsis median proportion, 18.08% [6.69–48.33%]; NIC median proportion, 0.81% [0.53–3.01%]; P = 0.0002), and subset 2 was identified as being composed of CD10⁻CD64⁺CD16^lowCD123⁺ immature neutrophils (sepsis median proportion, 10.06% [1.12–39.35%]; NIC median proportion, 0.04% [0.02–0.42%]; P < 0.0001) (Figure 2). We also recapitulated previously described results (13, 14, 18) regarding the sepsis-related increase of circulating immature CD10⁻CD64⁺ neutrophils when compared with NIC samples at Day 1 (sepsis median proportion, 11.03% [1.41–40.39%]; NIC median proportion, 0.62% [0.12–1.46%]; P = 0.001), and we confirmed their phenotypic similarities with a third of BM neutrophils (BM median proportion, 37.39% [17.90–46.48%]) (Figure 2E). In addition, we noticed that all HD-specific nodes were absent in Day 1 samples from patients with sepsis (Figures 2B and 2D).

Figure 2.

Identification of Sepsis (S) Day 1 (D1)–specific neutrophils by using a discovery analysis strategy. (A) t-Distributed Stochastic Neighbor Embedding (t-SNE) analysis was performed on neutrophils from all samples, with cells being organized along t-SNE-1–2 and t-SNE-2–2 axes according to per-cell expression of CD11b, CD66b, CD16, CD10, CD64, CD123, and PD-L1 (programmed death ligand 1). Cell density for the concatenated file of each group is shown on a black to yellow heat scale for each group time point. (B) The heatmap shows sample clustering (columns) according to nodes’ log₂-transformed cell proportion and centered around the mean proportion of all samples’ nodes (rows). Samples and the mean-centered, log₂-transformed node cell proportion were arranged according to complete-linkage hierarchical clustering. Heat intensity (from blue to yellow) reflects the mean-centered log₂-transformed cell proportion of each sample’s node. (C) The heatmap shows the characterization of cell nodes identified using the Spanning-tree Progression Analysis of Density-normalized Events (SPADE) algorithm (columns) according to the mean expression of seven markers (rows). Markers were arranged according to complete-linkage hierarchical clustering, and nodes were preordered according to the heatmap node order in B. Heat intensity (from blue to red) reflects the mean expression of each marker for each node. (D) Four groups of nodes were back-viewed on a t-SNE1–2/t-SNE2–2 map. (E) The cell abundance of each meta-cluster subset (CD10⁻CD64⁺CD16⁺PD-L1⁺ cell subset in red, CD10⁻CD64⁺CD16^lowCD123⁺ cell subset in blue, and CD10⁻CD64⁺ cell subset in green) was presented as the cell proportion among the total neutrophils of each group of samples. A nonparametric, two-tailed Mann-Whitney test was used to compare differences in the cellular abundance of cell subsets between noninfected control (NIC)-D1 and S-D1 samples (see Methods). Sample sizes were as follows: HD = 11, BM = 5, NIC = 12, and S = 17. BM = bone marrow; HD = healthy donor.

With this strategy, we identified two novel neutrophil subsets that included CD123⁺ cells and PD-L1⁺ cells at an early stage of sepsis and found an absence of HD neutrophil phenotypes.

A Computational Validation Strategy Confirmed Sepsis Day 1–Specific Neutrophil Subsets

To test whether the previously identified neutrophil subsets were sepsis specific and robust, an independent unsupervised data analysis strategy was applied on the same data files used in the discovery strategy (Figures 1B and 1C). This validation strategy was based on the UMAP and FlowSOM algorithms. UMAP is an unsupervised dimensionality-reduction algorithm (Figure E6A), and FlowSOM is an unsupervised clustering algorithm. This strategy allowed the identification of 50 neutrophil clusters and the complete-linkage hierarchical clustering of their relative cell abundance, and the samples were arranged according to patient groups (Figure E6B). Two main cell-cluster groups appeared to be more abundant in sepsis samples (Figures E6B and E6C), and almost all HD-associated clusters were absent in Day 1 samples from patients with sepsis.

To phenotypically characterize these two main cell clusters, the MEM phenotype annotation tool was used. The MEM label of each cluster is an objective description of what makes that subset distinct from all the other clusters. Among these clusters, three cell meta-clusters were identified, one with CD10⁻CD64⁺ immature cells and two meta-clusters that were phenotypically identical to the sepsis-specific neutrophil nodes identified in the discovery strategy (Figures 3A and E6D). Subset 1 clusters contained CD10⁻CD64⁺PD-L1⁺ neutrophils with median cell proportions of 5.50% (1.15–38.03%) for Day 1 sepsis samples and 0.09% (0.02–0.33%) for Day 1 NIC samples (P < 0.0001) (Figure 3B). Subset 2 clusters gathered CD10⁻CD64⁺CD16^low/−CD123⁺ immature neutrophils with median cell proportions of 2.43% (0.98–6.32%) and 0.04% (0.03–0.28%) for Day 1 sepsis and Day 1 NIC samples, respectively (P = 0.0006) (Figure 3B). We also visually noted that subset 1 clusters (PD-L1⁺ cells) and subset 2 clusters (CD123⁺ cells) from the validation strategy are colocalized with subset 1 nodes (PD-L1⁺ cells) and subset 2 nodes (CD123⁺ cells), respectively, from the discovery strategy when back-mapped onto the t-SNE1–2 and t-SNE2–2 axes (Figure 3C).

Figure 3.

Validation of Sepsis (S) Day 1 (D1)–specific neutrophil subsets by using a second computational strategy. As a first step, Uniform Manifold Approximation and Projection for Dimension Reduction (UMAP) analysis was performed on all samples’ neutrophils, and cells were organized along UMAP-1 and UMAP-2 axes according to per-cell expression of CD11b, CD66b, CD16, CD10, CD64, CD123, and PD-L1 (programmed death ligand 1). As a second step, FlowSOM clustering was done to separate neutrophil subsets into 50 clusters. MEM was then used to quantify the enriched features of the 50 clusters. Protein enrichment was reported on a +10 to −10 scale, in which +10 indicates that the protein’s expression was especially enriched and −10 indicates that the protein’s expression was excluded from those cells, relative to the other neutrophil clusters. (A) Among these clusters, two meta-clusters were identified as being phenotypically identical to the strategy 1 S-specific neutrophils: clusters 18 and 19 (in red), composed of CD10⁻CD64⁺PD-L1⁺ neutrophils, and clusters 6 and 7 (in blue), composed of CD10⁻CD64⁺CD16^lowCD123⁺ neutrophils. (B) The cell abundance of each meta-cluster subset (the CD10⁻CD64⁺CD16⁺PD-L1⁺ cell subset in red and the CD10⁻CD64⁺CD16^lowCD123⁺ cell subset in blue) was presented as the cell proportion among the total neutrophils of each group of samples. A nonparametric, two-tailed Mann-Whitney test was used to compare differences in the cellular abundance of cell subsets between NIC-D1 and S-D1 samples (see Methods). Sample sizes were as follows: HD = 11, BM = 5, NIC = 12 ,and S = 17. (C) Each meta-cluster of cells (red and blue) was back-viewed on both a UMAP-1/UMAP-2 map and a t-SNE1–2/t-SNE2–2 map. BM = bone marrow; HD = healthy donor; MEM = marker enrichment modeling; NIC = noninfected control; t-SNE = t-Distributed Stochastic Neighbor Embedding.

Expert Gating Strategy Based on a Limited Set of Markers Validated the Sepsis Day 1 Neutrophil Signature That Correlates with SAPSII and SOFA Scores

After cell subsets were identified by automatic and high-dimensional analysis strategies, we determined whether the identified neutrophil signature could be found by using conventional analysis applicable by experts. The use of such a gating strategy would make it easier to transpose it to clinical use.

A biparametric gating strategy on a limited set of markers allowed the identification of neutrophils expressing CD123 and PD-L1 (Figure 4A). When CD123⁺ and PD-L1⁺ sepsis-specific neutrophils were mapped back onto both SNE1–2 and t-SNE2–2 axes and and UMAP1 and UMAP2 axes, they were located in the same regions as the cells identified by using the two previous computational strategies, meaning that they share the same phenotype (Figure 4A). This expert gating strategy applied on the current data set allowed for the selection of PD-L1–expressing neutrophils that were significantly more abundant in the blood of patients with sepsis on Day 1 (9.25% [3.61–36.97%]) than in the blood of NIC patients on Day 1 (0.12% [0.07–0.60%]; P < 0.0001) or HDs (0.01% [0.00–0.03%]; P < 0.001) (Figure 4B). Similarly, expert gating allowed the selection of sepsis-specific neutrophils (2.47% [0.44–17.42%]) that were consistent with the CD123⁺ subset cell phenotype and that were almost absent from NIC samples (0.04% [0.07–0.87%]; P < 0.0001) and HD samples (0.04% [0.02–0.10%]; P < 0.0001) (Figure 4B).

Figure 4.

Sepsis (S) Day 1 (D1)–specific neutrophil subsets validated by using expert gating and correlated with severity scores. (A) An expert gating strategy with a seven-marker set allowed for the selection of a CD10⁻CD64⁺PD-L1⁺ cell subset (in red) and a CD10⁻CD64⁺CD16^lowCD123⁺ cell subset (in blue), which were back-viewed on both discovery (t-SNE1–2/t-SNE2–2) and validation (uniform Manifold Approximation and Projection for Dimension Reduction [UMAP]-1/UMAP-2) maps. (B) The two neutrophil subsets are significantly more abundant in the blood of patients with S collected at D1 after admission to the ICU than in the blood collected at D1 or D7 from noninfected control (NIC) post–cardiothoracic surgery patients or healthy donors (HDs). (C) Correlation between the log₁₀-transformed frequency of the CD10⁻CD64⁺PD-L1⁺ neutrophil subset (in red) or the CD10⁻CD64⁺CD16^lowCD123⁺ neutrophil subset (in blue) and Simplified Acute Physiology Score II (SAPS II) score (green squares) or Sequential Organ Failure Assessment (SOFA) score (purple squares). (D) The area under the receiver operating characteristic curve (AUROC) obtained by using only the CD123⁺ neutrophil subset. (E) The AUROC obtained by the using CD123⁺PD-L1⁺ neutrophil subsets. (F and G) The AUROCs obtained by using the SOFA (F) and SAPS II (G) clinical scores. A nonparametric, two-tailed Mann-Whitney test was used to compare the cellular abundance of cell subsets between S-D1 samples and NIC-D1, NIC-D7, or HD samples. A nonparametric, two-tailed Wilcoxon signed rank test was used to compare cellular abundance between the two matched groups: S-D1 and S-D7. Linear regression lines and Spearman rank correlation were used to assess the relationship between neutrophil subset frequency and severity scores (see the Methods). Spearman r coefficients and two-tailed P value are presented. Sample sizes were as follows: HD = 11, BM = 5, NIC = 12, and S = 17. BM = bone marrow; CI = confidence interval; PD-L1 = programmed death ligand 1; Std. = standard; t-SNE = t-Distributed Stochastic Neighbor Embedding.

Although the proportion of CD10⁻CD64⁺CD16⁻CD123⁺ neutrophils could be used to distinguish between sepsis samples and NIC samples at Day 1, we observed a large variability among patients. Interestingly, we noticed that patients with the highest CD123⁺ neutrophil subset proportion (>20%) tended to have a more severe condition (requirement for mechanical ventilation and catecholamine support). Later correlation with severity scores confirmed this observation. The proportion of CD123⁺ sepsis-specific neutrophils, as assessed by using the simple gating strategy on mass cytometry data, positively correlated with the SAPS II score (Spearman r, 0.62; P = 0.0192) and the SOFA score (Spearman r, 0.55; P = 0.0437) (Figure 4C). However, the proportion of CD123⁺ neutrophils was not influenced by the sepsis endotype. The proportions of PD-L1 neutrophil subsets did not correlate with severity scores or sepsis endotypes. ROC analysis of the abundance of these CD123⁺ neutrophils was performed to determine the optimal threshold separating patients with sepsis from noninfected patients. A cutoff point of 0.38% of the CD123⁺ neutrophil subset abundance enabled identification of patients with sepsis with a specificity of 91.67% and a sensitivity of 81.25% and displayed an area under the ROC curve (AUROC) of 0.91 (Figure 4D). When combining the abundance of the CD123⁺ and PD-L1⁺ neutrophil subsets, the cutoff point changed to 0.93%, lowering both the sensitivity (to 75%) and the specificity (to 83.33%). (Figure 4E). However, using a clinical SOFA score >2 enabled discrimination with good sensitivity (94.12%) but poor specificity (45.45%) and resulted in a lower AUROC of 0.79 (Figure 4F). In addition, the AUROC when using the SAPS II score was also lower (0.82), demonstrating a sensitivity of 88.24% and a poor specificity of 45.45% (Figure 4G).

Thus, a simple gating strategy assessing only seven key markers successfully identified CD123⁺ and PD-L1⁺ sepsis-specific neutrophils and indicated that CD123⁺ neutrophils may be a marker of sepsis severity with better discriminating efficiency than that of clinical scores.

Mass Cytometry and Unsupervised Analysis Identified Classical Sepsis Immune Hallmarks

By using two complementary computational strategies, we identified a sepsis-specific signature on neutrophil cells. We asked whether a signature on nonneutrophil cells could reinforce the CD123⁺ and PD-L1⁺ neutrophil subsets as sepsis biomarker candidates. The nonneutrophil circulating immune cells were computationally analyzed by using the t-SNE and SPADE algorithms. A heatmap was generated to characterize node phenotypes and delimitate the main circulating nonneutrophil immune-cell populations, according to complete-linkage hierarchical clustering (Figure E7A). These populations were then color-coded and back-gated on the t-SNE map (Figure E7B). Classical hallmarks of sepsis were identified, including lymphopenia, monocytopenia, and a persistently lower level of monocyte HLA-DR in patients with sepsis when compared with the HD group (P < 0.0001, P = 0.0426, and P < 0.0001, respectively; Figure 5A). In parallel, we observed an elevated number of circulating neutrophils (P = 0.0039), and consistent with that, a higher neutrophil-to-lymphocyte ratio (P < 0.0001) in patients with sepsis versus HDs (Figure 5B). These trends were not exclusive to sepsis but were also observed in the NIC group when compared with the HD group (P = 0.0003, P < 0.0001, P = 0.0034, and P < 0.0001 for lymphocyte and neutrophil counts, monocyte HLA-DR expression levels, and neutrophil/lymphocyte ratios, respectively). No significant difference was observed between the group with sepsis and the NIC group at Day 1 within these main immune-cell populations (Table E2).

Figure 5.

Nonneutrophil cell analysis identifies sepsis (S) immune hallmarks. (A) The lymphocyte and monocyte numbers and the intensity of mHLA-DR were obtained from nonneutrophil computational analysis and presented for each group. (B) Neutrophil numbers were obtained previously from the computational separation of neutrophils from nonneutrophils and were used to calculate the neutrophil/lymphocyte ratio. (C and D) Cell numbers for the main immune-cell subsets that were differentially abundant in the S group as compared with the healthy donor (HD) and NIC groups (C) and in the S group as compared with the HD group only (D). D1 = Day 1; mDC = mococyte-derived dendritic cells; mHLA-DR = HLA-DR expression on monocytes; NIC = noninfected control.

To identify an early sepsis-specific signature within these immune-cell populations, we compared the abundance of the identified cell nodes of these immune-cell populations among HD, NIC, and sepsis samples at Day 1. The abundance of 22 nodes was found to be selectively regulated in sepsis samples at Day 1 when compared with both NIC and HD samples, and 25 nodes differentiated sepsis-only samples from NIC samples at Day 1 (Figures E7C and E7D). Notably, these nodes included 15 nodes identifying classical monocytes with high expression of HLD-DR; 3 nodes identifying CD4⁺ T lymphocytes and CD8⁺ T lymphocytes expressing CCR2 and CCR6, all of which were highly reduced in patients with sepsis; 1 node identifying B lymphocytes with low expression of pan–B-cell markers (HLD-DR, CXCR5, CD19, and CCR6); and 1 node identifying monocyte-derived dendritic cells (Figure 5C). Among the nodes that were massively reduced in both the sepsis and the NIC samples, 15 nodes out of 55 represented basophil and eosinophil subsets (Figure 5D); the others were scattered among other cell populations.

Taken globally, analyzing circulating nonneutrophil cells by using a computational strategy allowed us to identify sepsis hallmarks as well as the abundance differences among several circulating immune-cell subsets.

CD123⁺ and PD-L1⁺ Sepsis-Specific Neutrophils Are Detectable through Conventional Cytometry and Discriminate between Patients with and without Infection

We identified two neutrophil subsets by including 40 individuals and using 42-marker mass cytometry and computational analysis. These subsets might be detectable by using the conventional cytometry approach that is used routinely in the clinic. To evaluate the efficiency and specificity of CD123⁺ and PD-L1⁺ neutrophil subsets to differentiate patients with sepsis from patients without infection, we set up a fluorescent, seven-marker flow cytometry panel (Table E4). We monitored an independent validation cohort composed of noninfected patients (n = 18) and patients with sepsis (n = 24).

By using the overlay of full-minus-two panel–stained control samples and full panel–stained samples from three representative patients to analyze several expression levels of CD10, CD123, and PD-L1, we identified an increase among CD123⁺ and PD-L1⁺ sepsis-specific neutrophil subsets and a decrease of neutrophil CD10 expression (CD14⁻CRTH2⁻CD15⁺ cells) (Figure 6A). ROC analysis was performed by using CD123⁺ and PD-L1⁺ neutrophil subset abundance, which was measured by using conventional flow cytometry on samples from an independent validation cohort of patients with sepsis and noninfected patients. A cutoff point of 0.35% of the CD123⁺ neutrophil subset abundance enabled ruling out sepsis with a specificity of 94.44%, a sensitivity of 87.5%, and an AUROC of 0.95 (Figure 6B). When combining the abundance of the CD123⁺ and PD-L1⁺ neutrophil subsets, the cutoff point changed to 0.60%, demonstrating no effect on the sensitivity or specificity (Figure 6C). However, using a clinical SOFA score >2 enabled discrimination with good sensitivity (91.30%) but poor specificity (18.18%) and resulted in a lower AUROC of 0.61 (Figure 6D). In addition, the AUROC when using the SAPS II score was also lower (AUROC, 0.69), demonstrating a sensitivity of 91.67% and a poor specificity of 25.00% (Figure 6E). These results indicated that conventional flow cytometry recapitulates the results obtained by using mass cytometry and confirmed that the identified neutrophil subsets could be used as a marker of sepsis severity that is more efficient than using clinical scores and is reliably quantified by routinely performed clinical flow cytometric profiling.

Figure 6.

Sepsis-specific neutrophils are detectable by conventional cytometry and discriminate infected from noninfected patients. (A) The gating strategy applied on fluorescent flow cytometry data of three patients with sepsis from the validation cohort. The overlay of FMT-stained control samples and the FP-stained samples of each representative patient showed an increase of sepsis-specific neutrophil subsets and a decrease of CD10 expression by neutrophils (CD14⁻CRTH2⁻CD15⁺ cells). (B–E) The AUROCs obtained by using only the CD123⁺ neutrophil subset (B), the two CD123⁺ and PD-L1⁺ neutrophil subsets (C), or the SOFA (D) and SAPS II (E) clinical scores. AUROC = area under the receiver operating characteristic curve; CI = confidence interval; FMT = full-minus-two panel; FP = full panel; FSC = forward scatter; PD-L1 = programmed death ligand 1; S = sepsis sample; SAPS II = Simplified Acute Physiology Score II; SOFA = Sequential Organ Failure Assessment; SSC = side scatter; Std. = standard.

In addition, we evaluated whether the CD123⁺ neutrophil subset was only abundant in patients with the highest severity scores. We used the data generated in both the discovery (Figure 4) and the validation cohorts (Figure 6) and divided the cohorts by quartile of severity according to the SOFA and SAPS II scores (Figures E8A and E8B). Although patients with sepsis and noninfected patients overlap greatly in terms of their severity scores, the CD123⁺ neutrophil subset proportion efficiently distinguished between the sepsis and control groups of both the discovery cohort (Figure E8A) and the validation cohort (Figure E8B).

Immature Sepsis Neutrophils Exhibit Impaired Microbe-Specific Activation and Phagocytosis

To address sepsis-associated neutrophil activation and phagocytic capacities, whole blood of each tested individual was incubated with S. aureus– or zymosan-coated BioParticles labeled with pHrodo, a pH-sensitive fluorochrome (36) to identify immature neutrophil BioParticle uptake capacity and activation.

All immature circulating neutrophils (CD64⁺CD10⁻) were able to phagocyte S. aureus beads regardless of their group (HD, Sepsis Day 1, BM). However, the Day 1 neutrophil phagocytosis of zymosan beads among sepsis samples (mean ± SD, 28.12% ± 8.39%) was not as effective as that among HD samples (mean ± SD, 50.43% ± 13.04%; P = 0.02) (Figure 7A). This sepsis-associated decrease of phagocytosis correlates with the proportion increase of both CD123⁺ and PD-L1⁺ immature neutrophil subsets in the blood of the tested patients with sepsis when compared with HDs (Figure 7B). t-SNE visualization of the positive control (PC) and negative control (NC) neutrophils in both the S. aureus (Figures 7C and 7D) and the zymosan (Figures 7E and 7F) bead stimulations highlighted the lower expression level of CD11b markers by sepsis-associated neutrophils at Day 1 when compared with HD neutrophils and also highlighted the default activation of these cells after microbial bead activation. In fact, sepsis-associated neutrophils exhibited a lower CD11b and CD66b PC/NC MFI ratio after activation than did HD neutrophils after S. aureus (Figure 7D) or zymosan (Figure 7F) stimulation. The impaired phagocytic capacity of the immature sepsis neutrophils compared with HD neutrophils was confirmed by the measurement of the phagocytosed bead PC/NC MFI ratios. This ratio was three times lower for the Sepsis Day 1 S. aureus response (Figure 7D) and 30% lower for the Sepsis Day 1 zymosan response (Figure 7F). These data allowed identification of the impaired capacity of immature sepsis neutrophils to form efficient phagolysosomes after BioParticle stimulation and showed a default of activation when compared with HD neutrophils.

Figure 7.

Staphylococcus aureus– and zymosan-specific activation and phagocytosis are impaired in sepsis immature neutrophils. To address the phagocytic capacities of sepsis immature (CD64⁺CD10⁻) neutrophils, 100 μl of blood was incubated with 20 μl or 40 μl of beads coated with S. aureus or zymosan, respectively, coupled with pH acidification–sensitive fluorochrome. After 1 hour of incubation at 37°C (positive control [PC]) or 4°C (negative control [NC]), cells were stained and analyzed by using flow cytometry. (A) Gating strategy of CD15⁺CD14⁻CD3⁻CD19⁻ neutrophils from healthy donors (HDs), Sepsis Day 1 (S-D1) samples, and bone marrow (BM) samples. Cells were separated into two gates on the basis of CD10 expression and phagocytosis marker intensity (S. aureus or zymosan), and PC cells (red dots) were overlaid onto NC cells (blue dots). The proportions of total phagocytic neutrophils were presented for the three groups. t-SNE analysis organized cells along the t-SNE axes according to per-cell expression of five proteins and phagocytosis fluorescence. (B and C) The cell expression of CD11b after S. aureus (B) or zymosan (C) stimulation for one representative individual from the HD group and one representative individual from the S-D1 group stimulated at +4°C (NC) and +37°C (PC) is shown on a heat scale. (D and E) For each individual, the PC/NC mean fluorescence intensity (MFI) ratios for CD66b and CD11b and after S. aureus (D) or zymosan (E) stimulation in each group were plotted in histograms. CD10⁻ cells have less phagocytic capacity, whether measured by using the MFI or a proportion. Stimulated CD10⁻ cells exhibit a lower level of expression of CD11b and CD66b. A nonparametric, two-tailed Mann-Whitney test was used to compare differences in the cellular abundance of cell subsets and in the MFI ratios (see the Methods). Sample sizes were as follows: HD = 4, S-D1 = 6 and BM = 3; PD-L1 = programmed death ligand 1; t-SNE = t-Distributed Stochastic Neighbor Embedding.

Discussion

Whole-blood mass cytometry and computational analysis identified classical hallmarks of sepsis and revealed two novel neutrophil subsets that distinguish early sepsis from aseptic inflammatory syndromes. We identified two novel neutrophil subsets, CD10⁻CD64⁺PD-L1⁺ neutrophils and CD10⁻CD64⁺CD16^low/−CD123⁺ immature neutrophils, that could be used for early identification of sepsis in patients. The CD123⁺ and PD-L1⁺ neutrophil subsets could help improve sepsis diagnosis and guide sepsis treatment monitoring.

The results of this study recapitulated previous original findings and meta-analyses regarding the sepsis-related increase of circulating immature CD10⁻CD64⁺ neutrophils (13, 14, 18, 37). Despite all of these large efforts, CD64 detection–based tools are not yet standardized for sepsis diagnosis because of the heterogeneity of sepsis syndrome and the interindividual variability of the CD64 basal expression among patients with sepsis.

The CD10⁻CD64⁺CD16^low/−CD123⁺ population is most consistent with immature neutrophils. The frequency of this population among total neutrophils positively correlates with both SAPS II and SOFA severity scores and needs to be confirmed in a larger collection. The neutrophil expression of CD123 during sepsis has not been described before. A previous study by Weber and colleagues (20) used a mouse model of abdominal sepsis and reported that the cytokine IL-3 potentiates inflammation in sepsis by inducing myelopoiesis of neutrophils and that IL-3 deficiency protects mice against sepsis. Moreover, the authors described an association between high plasma IL-3 concentration and high mortality. This result was also obtained in a recent prospective cohort study, in which higher concentrations of IL-3 were shown to be independently associated with hospital mortality in patients with sepsis (21). All of these results identify IL-3 and its receptor CD123 as orchestrators of emergency myelopoiesis and reveals a new target for the diagnosis and treatment of sepsis.

To our knowledge, the expression of PD-L1 by neutrophils during sepsis has not been reported before. Expression was defined among monocytes, macrophages, and endothelial cells (38) but not among granulocytes. Monocyte PD-L1 expression was described as an independent predictor of 28‐day mortality in patients with septic shock (22, 23). Peripheral blood transcriptomic analysis done by Uhel and colleagues (15) revealed the PD-L1 gene to be among the top 44 immune-related genes differentially expressed between patients with sepsis and HDs. In parallel, mice in which the PD-1/PD-L1 interaction was inhibited show improved sepsis survival (39). Our results bring up a new target for the immune checkpoint therapies.

Controversial results were previously described regarding the functional aspects of neutrophils during sepsis. On one hand, Demaret and colleagues (40) described conserved phagocytosis and activation capacities among sepsis neutrophils characterized as CD10^dimCD16^dim immature cells after whole-blood IL8-, N-formylmethionine-leucyl-phenylalanine–, or fluorescein isothiocyanate–labeled Escherichia coli stimulation. On the other hand, by comparing mature and immature neutrophil functions, Drifte and colleagues (41) found that the latter were less efficient in phagocytosis and killing. Accordingly, we observed an impairment in the capacity of cells to form efficient phagolysosomes after BioParticle stimulation and a default of activation when compared with HD cells.

The immunosuppressive function was also attributed to a granulocyte-like myeloid-derived suppressor cell (G-MDSC) neutrophil subset during sepsis (13–15, 18). But, to date, human G-MDSC definition lacks consensus regarding the phenotypic characterization. Published results on G-MDSC in cancer were obtained according to various phenotypes. Condamine and colleagues (42) described them as LOX1 (lectin-type oxidized LDL receptor 1)–expressing cells. By using flow cytometry, we measured the expression of LOX-1 in patients with sepsis (data not shown). No LOX-1 co-staining was observed among either the CD123⁺ subset or the PD-L1⁺ subset. More investigation is needed to characterize whether CD123⁺ neutrophils and PD-L1⁺ subsets belong to the G-MDSC subset.

Further research should be conducted to identify appropriate clinical actions for each identified neutrophil subset and their evolution over the time course and in different cohorts of patients (undifferentiated shock patients, immunosuppressed patients, different types of infections, durability of neutrophil population after antibiotics) to understand whether altered neutrophil production is responsible for increased sepsis risk and to determine how these subsets can be therapeutically targeted.

In this study, we show that the use of the identified neutrophil subsets provides information complementary to that provided by severity scores, such as those from the SOFA and SAPS II, and that these subsets are specific to sepsis. In the discovery cohort, in which stringent selection was applied for the inclusion of patients with sepsis and NIC patients, few differences were observed among the AUROCs achieved when using CD123⁺ neutrophils, SOFA scores, and SAPS II scores (Figures 4D, 4F, and 4G). In contrast, in the validation cohort, in which blind analysis was performed, the SOFA and SAPS II scores lost their discrimination power (Figures 6D and 6E) and the CD123⁺ neutrophil biomarker remained highly specific and sensitive for the identification of sepsis in patients.

In addition, although the diagnosis of sepsis was considered for a significant proportion of patients (6 of 18) in the ICU control group of the validation cohort and a third of these patients received antibiotics because of their clinical characteristics, the diagnosis of sepsis was finally ruled out. They were found to be noninfected and indistinguishable from patients with other inflammatory and noninflammatory conditions (Table E5). Of note, the CD123⁺ neutrophil proportion of these patients was <0.3%, which was below the cutoff value identified in our ROC analysis. The use of this biomarker candidate would have enabled avoidance of this unnecessary administration of antibiotics, especially given that flow cytometry is a technique with wide availability in the clinic, reasonable costs, and rapid results.

The use of a whole-blood flow cytometry test to diagnose sepsis could change the fate of patient care. The clinician would have a rapid and specific result, obtained before microbiological culture results, that could guide their therapeutic decision.

In parallel, future studies using flow cytometry in the clinic should now be undertaken to validate the use of these new neutrophil subsets as early biomarkers predictive of sepsis. Larger cohorts that better represent not only sepsis but also the diversity of aseptic inflammatory syndromes need to be evaluated.

A delay in sepsis diagnosis has been shown to decrease survival and increase hospital costs, and a better diagnostic strategy will definitely help to improve patient care, enable avoidance of unnecessary treatments, and reduce hospital lengths of stay.