Unsupervised analysis of transcriptomics in bacterial sepsis across multiple datasets reveals three robust clusters

Abstract

Objective

To find and validate generalizable sepsis subtypes using data-driven clustering.

Design

We used advanced informatics techniques to pool data from 14 bacterial sepsis transcriptomic datasets from 8 different countries (N=700).

Setting

Retrospective analysis.

Subjects

Persons admitted to the hospital with bacterial sepsis.

Interventions

None.

Measurements and Main Results

A unified clustering analysis across 14 discovery datasets revealed three subtypes, which, based on functional analysis, we termed Inflammopathic, Adaptive, and Coagulopathic. We then validated these subtypes in 9 independent datasets from 5 different countries (N=600). In both discovery and validation data, the Adaptive subtype is associated with a lower clinical severity and lower mortality rate, and the Coagulopathic subtype is associated with higher mortality and clinical coagulopathy. Further, these clusters are statistically associated with clusters derived by others in independent single sepsis cohorts.

Conclusions

The three sepsis subtypes may represent a unifying framework for understanding the molecular heterogeneity of the sepsis syndrome. Further study could potentially enable a precision-medicine approach of matching novel immunomodulatory therapies with septic patients most likely to benefit.

Introduction

Sepsis is defined as life-threatening organ dysfunction resulting from a dysregulated immune response to infection (1). Despite its association with nearly half of all in-hospital deaths, there are still no approved therapies specific for sepsis (2, 3). In part, this is because the clinical syndrome of sepsis includes substantial heterogeneity and may in fact encompass many different subtypes, analogous to what is well established among patients with cancer (4, 5). Current sepsis groupings are based on clinical criteria such as the presence of shock, infection source, or organ failure, but such groupings may not represent the driving biology of the host response. They have also failed to adequately match patients for novel interventions. If the heterogeneity of sepsis truly reflects heterogeneity in the host response, characterization of these underlying host response types will be fundamental to enabling precision sepsis therapeutics(6).

In unsupervised analysis, data is sorted into subgroups (‘clusters’) that are defined only internally and without reference to external ‘supervisory’ outcomes, such as mortality or severity. Instead, the structure inherent within the data is used to define the subgroups. Such data-driven analyses have been successful in defining validated, clinically relevant disease subtypes in multiple diseases(4, 5, 7, 8). Since whole-blood gene expression reflects the temporal state of the circulating leukocytes, at least two academic groups have applied unsupervised clustering to whole-blood transcriptomic profiles in patients with sepsis to study the ‘host response’ in a data-driven framework(9–13). Their results have identified higher-mortality subtypes with evidence of immune exhaustion and diminished glucocorticoid receptor signaling, as well as lower-mortality subtypes with conventional pro-inflammatory signaling(9–13).

Clustering analyses often yield non-reproducible results for one of two reasons: either multiple arbitrary choices in methodology are used such that minor changes in analysis yield new results, or the clustered dataset is too small and not representative of the broad heterogeneity of a disease. However, recent advances in meta-clustering and data pooling can help solve both problems(14–16). Coupled with an unprecedented amount of publicly available transcriptomic data in sepsis(17, 18), here we tested the hypothesis that there exist robust, reproducible sepsis host-response subtypes (clusters) across the broad, heterogeneous spectrum of clinical sepsis.

Methods

Systematic search and dataset criteria

We performed a systematic search of GEO and ArrayExpress for gene expression studies of clinical studies in sepsis, as previously described (16). Individual datasets were renormalized as previously described (18). Datasets were only included if they studied whole blood gene expression at hospital or ICU admission (i.e., primary admission for sepsis). Since the host response differs substantially between bacterial and viral infections(15, 19), an unsupervised analysis would likely lead to groupings primarily based on infection type. We thus removed all samples with microbiologically confirmed viral infection unless a microbiologically confirmed bacterial infection was also present (only 3 confirmed co-infections were included). Studies that did not supply sample-level microbiological data but were identified in their manuscript as being drawn from patients with primarily bacterial sepsis were treated as all bacterial. We further removed patients that were sampled more than 48 hours after sepsis diagnosis given the potential impact of treatment on the host response(20, 21). All data used herein were de-identified and publicly available and so exempt from IRB review.

Pooling data with COCONUT to enable clustering

The recent development of the COmbat CO-Normalization Using conTrols method (COCONUT)(15) allows for bias-free correction of batch effects between multiple microarray datasets, enabling pooled analysis, provided that healthy controls are present. The core assumption is that healthy controls across datasets come from the same statistical distribution. This assumption allows for the calculation of correction factors that remove technical differences across pooled datasets without bias to the number or type of diseased samples present.

We split the datasets into ‘discovery’ and ‘validation’ groups based on whether healthy controls were present in the dataset, specifically so that we could use the COCONUT method. Since the inclusion of healthy controls in any given dataset is essentially random, the discovery/validation split was not expected to introduce bias. We used the COCONUT method to co-normalize the discovery datasets into a single pool, and then removed all healthy controls from further analysis.

Clustering the discovery data using COMMUNAL

In order to determine how many clusters were present in the COCONUT-conormalized discovery data, we used the COmbined Mapping of Multiple clUsteriNg ALgorithms (COMMUNAL) method, which integrates data from multiple clustering algorithms and validity metrics across a range of included variables to identify the most robust number of clusters present in the data (see Supplementary Digital Content) (14). We ranked the top 5,000 genes across the discovery datasets using an algorithm that accounts for both within-dataset variance and between-dataset variance (16). We ran COMMUNAL using consensus-clustering versions of two algorithms, K-means clustering and Partitioning Around Medioids (PAM), due to their robustness in large, noisy datasets. Both methods were run across a range of variables from 100 genes up to 5,000 genes (in ranked order). COMMUNAL then integrated these data (at its default parameters) to produce an optimality map of clustering. In the resulting map, the most stable optima were taken as indicating the most robust clustering.

Having chosen an optimal clustering using COMMUNAL, we integrated the sample assignments between clustering algorithms (i.e., the clusters into which the PAM and K-means algorithms assigned samples). The COMMUNAL method assigned all samples for which the clustering algorithms agreed to discovery clusters, and removed all samples for which there was disagreement between the PAM and K-means methods as ‘unclustered’. The hypothesis is that not every sample may be perfectly assigned to a given cluster (e.g., some samples may exhibit biology suggestive of two clusters). Since classifiers trained on data with fewer errors are more robust, removing these uncertain samples improves the classifier accuracy. Note that the classifier built for validation does not produce ‘unclustered’ assignments (see Supplementary Digital Content).

To check whether the discovery clusters appeared to be separated in gene expression space, we visualized them using both heat maps and principal component analyses. We further used pooled sample-level demographic and phenotypic data to investigate clinical differences between discovery clusters.

Biological and clinical investigations

The details of our treatment of complex clinical variables including illness severity, immunosuppression, and coagulopathy are explained in the Supplemental Digital Content. Gene ontology analysis (22), the construction of a cluster classifier (23), and testing of the validation datasets are described in the Supplementary Digital Content.

Jargon-free summary

In recognition of the highly technical nature of the paper, we have prepared a ‘jargon-free summary’ of the methods and results. This is available in the Supplementary Digital Content.

Results

Included studies, COCONUT conormalization, and COMMUNAL cluster selection

We first hypothesized that robust molecular subgroups exist in patients with bacterial sepsis. We thus performed a unified clustering across 14 bacterial sepsis discovery datasets from 8 different countries (N=700, Table 1a) using COCONUT co-normalization (24–37). We identified 9 validation datasets from 5 different countries that matched inclusion criteria but did not include healthy controls (N=600, Table 1b and Figure 1)(12, 38–43). We first co-normalized the 14 discovery datasets into a single pooled cohort using the COCONUT method (15), providing batch-corrected, pooled sepsis data across a wide variety of clinical conditions (Supplemental Figure 1). There were 8,946 genes that were measured in all 14 pooled discovery datasets. The pooled data were then clustered using the COMMUNAL algorithm across 11 test points ranging from the top 100 to 5,000 genes using consensus K-means and consensus PAM clustering (individual clustering algorithm results shown in Supplemental Figure 2) (14). Visual inspection of the COMMUNAL optimality map showed clear, stable optima at K=3 clusters from 500 genes to 5,000 genes (Supplemental Figure 3). Further, we chose the clustering at 500 genes as the optimal clustering assignment under the assumption that using the fewest number of genes had the least amount of noise or redundant signal. Based on gene ontology analysis described below, and to facilitate their easier understanding, we have named the three clusters “Inflammopathic”, “Adaptive”, and “Coagulopathic”.

Table 1.

Datasets included in the study. (A) Datasets with healthy controls were chosen for discovery and (B) datasets without healthy controls were chosen for validation.

Use	Accession	First author	Description of patients used here	Timing of sepsis diagnosis	Sample Size (N)	Age	Sex (percent male)	Severity	Mortality (percent)	Country
A. Discovery	EMEXP3567	Irwin	Children with meningococcal sepsis +/− HIV co-infection	Admission to ED	12	2.0 (IQR 0.6–6.9)	55	unk.	50	Malawi
	EMTAB1548	Almansa	Adult surgical patients with sepsis (EXPRESS study)	Average post-operation day 4 (hospital acquired)	82	69.7 (std. dev. 13.1)	67	APACHE II 17.0 (std. dev. 5.4)	32	Spain
	GSE11755	Emonts	Children with meningococcal septic shock	Admission to ICU	6	1.94	100	PRISM 24	0	Netherlands
	GSE13015 gpl6106	Pankla	Adults with sepsis, many from burkholderia	Within 48 hours of diagnosis; both community-acquired and hospital-acquired.	48	54.7 (std. dev. 11.7)	54	unk.	27	Thailand
	GSE13015 gpl6947				15				47
	GSE20346	Parnell	Adults with severe bacterial pneumonia	Admission to ICU	6	63 (range 52–75)	50	APACHE II 22 (range 10–33)	33	Australia
	GSE28750	Sutherland	Community-acquired sepsis with bacteremia	Admission to ICU	10	60	55	unk	unk	Australia
	GSE33341	Ahn	Adults with 2+ SIRS criteria and bacteremia	Within 24 hours of admission to hospital	51	58 (range 24–91)	61	unk.	4	USA
	GSE40012	Parnell	Adults in ICU with sepsis	Admission to ICU	21	61 (std. dev. 16)	40	APACHE II 21 (std. dev. 6)	26	Australia
	GSE40586	Lill	Infants, children, and adults with bacterial meningitis	Within 48 hours of hospital admission	21	43.4 (range 17d – 70y)	unk.	unk.	10	Estonia
	GSE57065	Cazalis	Adults with septic shock	Within 48 hours of admission to ICU	56	62	68	SOFA 10.5 (IQR 9–13)	18	France
	GSE65682	Scicluna	Adults with community-acquired pneumonia in ICU	Within 24 hours of ICU admission	101	64	57	APACHE IV 81	24	Netherlands
	GSE66099	Wong	Children in ICU with sepsis/septic shock	Admission to ICU	188	3.7	58	PRISM 15.7	14	USA
	GSE69528	Khaenam	Adults with sepsis, many from burkholderia	unknown	83	adults	unk	unk	unk	Thailand
B. Validation	EMEXP3850	Kwan	Children w/meningococcal sepsis	Admission to hospital; sampled at multiple times 0–48 hrs	24	1.3 (range 0.8–2.0)	40	PELOD; 29.2 (range 11–61)	21	UK
	EMTAB4421.51	Davenport	Adults with sepsis (GAinS study)	Within 24 hours of admission to ICU	178	64.2 (std. dev. 15.2	55	APACHE II 18.6 (std. dev. 9.7)	32	UK
	GSE10474	Howrylak	Adults in MICU with sepsis +/− ALI	Admission to ICU	34	57 (std. dev. 4.3)	45	APACHE II 20.7 (std. dev. 1.6)	33	USA
	GSE28658	Khoo	Children with bacteremia	At diagnosis	6	3.6 (std. dev. 2.2)	33	unk	unk	Nigeria
	GSE32707	Dolinay	Adults in MICU with sepsis +/− ARDS	Admission to ICU	48	57.1 (std. dev. 14.9)	53	APACHE II 26.7 (std. dev. 8.5)	35	USA
	GSE63042	Langley	Adults with sepsis (CAPSOD study)	Admission to ED	104	59.1 (std. dev. 18.3)	59	APACHE II 16.5 (std. dev. 7.3)	37	USA
	GSE63990	Tsalik	Adults with bacterial infection plus 2+ SIRS criteria	Admission to ED	70	49 (range 14–88)	50	unk.	9	USA
	GSE66890	Kangelaris	Adults in ICU with sepsis +/− ARDS	Admission to ICU	62	63 (std. dev 19)	56	APACHE III 100 (std. dev. 35)	25	USA
	GSE74224	McHugh	Community-acquired sepsis with bacteremia	Within 24 hours of ICU admission	74	62.5	55	unk	unk	Australia & Netherlands

Figure 1.

Overall study schematic.

To visualize their general separability, we performed principal components analysis on the discovery clusters using all genes both with and without the ‘unclustered’ sample (Figure 2). Details on the assignment of clusters in the Discovery datasets are available in the Supplemental Results, Supplemental Tables 1–2, and Supplemental Figures 4–5.

Figure 2.

The first two principal components (PCs) of the discovery clustering results (both with (A) and without (B) the 16% of samples that went unclustered in the final analysis, in gold) using all 8,946 genes present in the COCONUT conormalized data. Here we show that the cluster assignments that we recovered in an unsupervised manner are clearly separated in high-dimensional space, as demonstrated by the first two principal components.

Gene ontology across the different clusters

To better understand the biology represented by the clusters, we used gene ontology over-representation analysis. We assigned each of the 500 genes to one of the three discovery clusters based on absolute effect size (i.e., each gene was assigned to the cluster in which it was most different from the remaining two clusters). We then tested each of the resulting three gene lists for significance in gene ontology (GO) terms (Supplemental Table 3). The Inflammopathic cluster was significant for canonical pro-inflammatory signaling pathways such as IL-1 receptor, pattern recognition receptor activity, and complement activation. The Adaptive cluster was significant for several pathways related to adaptive immunity and interferon signaling. The third cluster was named Coagulopathic as it was significant for terms related to clotting and coagulation, such as platelet degranulation, glycosaminoglycan binding, and coagulation cascade.

Clinical findings across the different clusters

We investigated the differences between the discovery clusters in the demographic and clinical variables for which we had subject-level data (Table 2). We found significant differences in age (both the overall distribution, and the percent of patients >70 years of age), severity (as measured by percent of patients with clinical severity scores above the dataset mean, and/or in septic shock), and 30-day mortality. We also found that the Inflammopathic cohort had greater bandemia and a lower lymphocyte percentage on white blood cell differential; however, differential was only available in a single cohort. This suggests that the Adaptive cluster is comprised of less sick patients with fewer elderly patients, while the Inflammopathic and Coagulopathic clusters separate the sicker patients into a younger and an older group. Addition of the ‘unclustered’ patients showed they have a balanced phenotype with respect to age and shock; their addition did not substantially change the demographic or clinical findings (Supplemental Table 4). Since the unsupervised clustering did not take into account any clinical data whatsoever, finding a significant difference in mortality suggests that the clusters may represent distinct pathophysiological states of clinical relevance.

Table 2.

Demographic and clinical variables across discovery clusters. Not all variables were available for all samples, so the totals are not always consistent; N for each measured variable is included as a separate column. Statistics were calculated by pooling data among cohorts.

Variable	Inflammopathic	Adaptive	Coagulopathic	P value (chisq/ANOVA)	Total N used
Total Samples Assigned	175	219	108
Male (percent)	58.4	59.4	61.5	0.864	481
Age (years, +/− sd)	34.8 (32.1)	38.5 (28.7)	49.7 (29.4)	0.0001	452
Age < 18 (percent)	16.8	17.6	15.9	0.930
Age > 70 (percent)	27.7	20.0	36.4	0.007
WBC count (+/− sd)	18.02 (16.18)	13.83 (10.64)	12.87 (13.3)	0.176	133
Neutrophils (+/− sd)	59.67 (18.31)	61.14 (16.42)	58.15 (23.1)	0.843	107
Bands (+/− sd)	17.04 (12.77)	11.58 (11.57)	6.75 (6.13)	0.002	107
Lymphocytes (+/− sd)	15.89 (13.8)	20.17 (12.71)	27.05 (23.16)	0.024	107
Monocytes (+/− sd)	6.07 (4.33)	6.19 (3.82)	6.6 (6.66)	0.91	107
Immunosuppressed (percent)	5.80	8.90	11.50	0.62	140
Gram negative (percent)	46.2	48.4	51.4	0.860	285
Shock (percent)	73.0	32.2	62.2	4.58E-10	297
High Clinical Severity (percent)	50.8	32.4	56.3	0.002	313
Non-survivor (percent)	29.8	8.1	25.4	8.01E-06	355

We ran regression models on cluster membership (in a ‘1-vs-all’ format) to assess the joint ability of age, shock, severity, and their interaction to predict cluster membership. In each case, the percent of variance explained by age, shock and severity was 9.7%, 6.4%, and 0.7% for the Inflammopathic, Adaptive, and Coagulopathic groups, respectively, in discovery (total N=251, Supplemental Table 5). A sensitivity analysis showed that these results could only be explained away by an unmeasured confounding variable with a substantially greater effect size than the included variables (Supplemental Table 5). Thus, while age, shock, and severity are significantly different across the groups, cluster assignment is much more complex than these three factors alone.

Validation of cluster classifier in independent datasets

Having characterized the sepsis clusters in the discovery datasets, we next hypothesized that these same clusters could be recovered in independent validation datasets using a discrete classifier. We next built a gene-expression-based classifier for cluster assignment so that the cluster hypothesis could be tested and applied in external validation datasets. Briefly, the classifier assigns each sample three scores (one for each cluster type) and then applies multiclass regression to output a final cluster assignment (Supplemental Table 6A–B). The classifier used a total of 33 genes, and yielded an overall 83% accuracy in leave-one-out re-assignment of the samples on which it was trained (Supplemental Table 6C). The greatest classifier inaccuracy is in distinguishing Inflammopathic patients from Coagulopathic patients (Supplemental Figure 6). We applied the classifier to the 9 bacterial sepsis validation datasets (Supplemental Table 7)(12, 38–44), and judged the classifier’s accuracy by its ability to recover clusters with similar molecular and clinical phenotypes to the discovery clusters. Since the 9 validation datasets are independent from one another, we examined the same demographic and clinical variables as in the discovery clusters in both a pooled fashion (Table 3) and treating each dataset independently (Supplemental Table 8). As the individual datasets may be underpowered to detect differences, we ran statistical tests in the pooled data; compared to the discovery clusters, we observed the same patterns of significance. The Coagulopathic cluster had significantly more patients older than 70 years (P<0.05), whereas the Adaptive cluster had fewer patients with shock (P<0.01), fewer patients with high clinical severity (P<0.05) and a lower mortality (P=0.01).

Table 3.

Demographic and clinical variables across validation clusters. Not all variables were available for all samples, so the totals are not always consistent. N for each measured variable is included as a separate row. Statistics are shown both by aggregating cohort-level statistics, and by pooling data among cohorts.

Variable	Inflammopathic	Adaptive	Coagulopathic	P value (chisq/ANOVA)	Total N used	Number of Datasets
Total Samples Assigned	208	264	128		600	9
Male (pooled percent)	51.7	62.5	60.0	0.08153	519	7
Age (pooled mean, sd)	57.9 (20.9)	57.3 (19.7)	60.9 (23.1)	0.3210	520	7
Age > 70 Y (pooled percent)	32.2	28.0	43.5	0.016	520	7
WBC count (+/− sd)	18.48 (11.12)	16.94 (21.61)	14.57 (7.79)	0.67	104	1
Neutrophils (+/− sd)	81.27 (17.33)	76.8 (17.51)	84.19 (11.72)	0.22	93	1
Bands (+/− sd)	12.82 (17.81)	2.5 (6.62)	5.83 (9.07)	0.035	51	1
Lymphocytes (+/− sd)	6.96 (4.76)	11.84 (8.46)	5.95 (4.94)	0.001	93	1
Monocytes (+/− sd)	4.24 (2.82)	6.85 (4.44)	5.03 (3.19)	0.01	93	1
Immunosuppressed (percent)	2.9	6.4	13	0.32	104	1
Gram negative (pooled percent)	66.7	78.3	61.1	0.468	68	3
Shock (pooled percent)	69.8	36.7	45.5	0.0036	136	2
High Clinical Severity (pooled percent)	45.5	31.8	39.6	0.030	450	6
Non-survivor (pooled percent)	29.3	18.5	31.1	0.01095	514	7

The Coagulopathic cluster also was associated with clinical coagulopathy, including disseminated intravascular coagulation (P<0.05, Table 4, Supplemental Tables 9–10 and Supplemental Results).

Molecular similarity between clusters identified in discovery and validation

Since the validation clusters were assigned with information from only 33 genes, we investigated whether similar biology was present in the full gene expression profiles across discovery and validation clusters. First, we calculated the mean gene expression profiles for all 500 clustering genes, and tested for correlation between the clusters. Significant correlation would indicate that the classifier was capturing most of the information from the original clustering; the 33 genes used in the classifier were thus excluded from this analysis to avoid bias. Pearson correlations in mean gene expression profiles within the assigned clusters were high (Inflammopathic cluster, 0.59±0.18; Adaptive cluster, 0.67±0.19; Coagulopathic cluster, 0.20±0.21, Figure 3A). These correlations were significant (P<0.01) between the discovery and validation clusters for all datasets for Inflammopathic, all datasets for Adaptive, and five out of nine datasets for Coagulopathic. As a comparison, 1000 random samples of 500 genes yielded mean correlations of 0.01 – 0.02.

Figure 3.

(A) Correlations of average 500-gene expression vectors between clusters assigned in the discovery and validation datasets; correlation coefficient is shown by color (legend at figure right). Notably, samples from Inflammopathic clusters are positively correlated with Inflammopathic samples from other datasets, and negatively correlated with Adaptive samples from other datasets (and vice-versa). The Coagulopathic clusters show less cohesion but are positively correlated with one another. (B) Heatmap of Gene Ontology (GO) codes found to be overrepresented in the different clusters, colored by significance levels. In both (A) and (B), the pooled ‘Core’ discovery datasets are represented by a single column for each cluster, while each cluster in each validation dataset is represented by a separate column. Both sub-figures show a block structure indicative of molecular similarity across datasets between clusters of the same type.

We next tested whether the same Gene Ontology (GO) codes were overrepresented between validation clusters, as compared to the discovery clusters (Figure 3B). On average, 68%, 87%, and 61% of the codes found significant at p<0.01 in the discovery clusters (Inflammopathic, Adaptive, and Coagulopathic, respectively) were identified as significant at p<0.05 in the same clusters in validation. In addition, a block structure is seen within clusters of the same type, indicating generally shared pathway enrichment within cluster types.

Comparison to previously established sepsis endotypes

Two groups have previously performed clustering using sepsis transcriptomic profiles. Wong et al. (9–11) and Davenport et al. (12, 13). We compared our cluster assignments to the previously published assignments and showed significant overlaps with the Inflammopathic and Adaptive clusters (Supplemental Results and Supplemental Table 10).

Discussion

We here performed an unsupervised clustering analysis on pooled transcriptomic profiles (N=700) from 14 datasets from a broad range of subjects with bacterial sepsis, demonstrating that there are three robust sepsis clusters (or ‘endotypes’). We have named these clusters Inflammopathic (higher mortality, innate immune activation), Adaptive (lower mortality, adaptive immune activation), and Coagulopathic (higher mortality, older, and with clinical and molecular evidence of coagulopathy), based on their molecular and clinical profiles. Next, we showed that a 33-gene classifier that assigns subjects to these three clusters is able to recover the clinical and molecular phenotypes in 9 independent validation datasets (N=600). Finally, we showed that these clusters can significantly explain the clusters derived by independent groups using different methods (9, 12). Taken together, these results demonstrate that the host response in the sepsis syndrome can be broadly defined by these three robust clusters.

Notably, each of the validation datasets had separate inclusion/exclusion criteria, providing a sort of sensitivity analysis that the identified clusters appear in both pooled settings (as in discovery) but also in more uniform, carefully phenotyped cohorts. For instance, we pooled samples from pediatric and adult datasets in discovery, but our methods did not simply cluster patients by age; then in validation, two datasets were pediatric and seven were adult, but all datasets contained a mix of all three sepsis clusters. The fact that we redemonstrate the same broad phenotypic and molecular differences in these independent applications of the cluster classifier is strong evidence that cluster membership is present across populations.

Despite the outcome differences across our three clusters, their clinical utility is not merely the ability to risk-stratify in terms of mortality. Mortality prediction is better achieved through purpose-built classifiers, which have been demonstrated with these same data(18). Instead, the hypothesis that underlies the search for sepsis clusters is that ‘sepsis’ represents multiple different disease states and manifests in many different ways(3, 6, 45). The aim of our study was thus to uncover these subclinical clusters using a very large pool of sepsis patients across a wide range of clinical conditions. Uncovering and defining this heterogeneity may allow for greater success in the discovery and validation of therapies that are beneficial only to one sepsis cluster, but may be neutral or even harmful to other clusters(11). For instance, both the molecular and clinical data suggest that the Coagulopathic cluster may be associated with functional coagulopathy. Given the association of sepsis with clinical coagulopathies, and despite (or perhaps because of) the failure of most therapeutic interventions for coagulopathy in sepsis (3, 46, 47), further study of the Coagulopathic cluster is warranted. Similarly, drugs being tested in sepsis that are known to modulate the innate or adapative immune systems (such as anti-IL-1 or anti-PD-L1 treatments (48, 49)) may potentially find efficacy in the Inflammopathic or Adaptive clusters, respectively.

We inferred pathobiology for the clusters by assigning each gene to the cluster in which it showed the greatest differential change from the other clusters. For instance, the association of innate immune pathways in the Inflammopathic cluster is indicative not of ‘normal’ innate immune activation, but rather of overactivation of the innate immune system, or of a relative lack of activation of adaptive immune genes, in Inflammopathic patients as compared to other septic patients. Similarly, the relatively higher adaptive immune gene activation in the Adaptive cluster may be linked to its lower mortality. Seen through this lens, the three sepsis clusters show biological insights that, to some degree, reflect clinical intuitions. The early overactivation of the innate immune system or coagulation cascade in sepsis is linked to higher mortality, while the relative lack of these changes and the expansion of the adaptive immune response may be linked to better outcome(50). Furthermore, since genes were selected based on absolute effect size, similarity in gene ontology pathway analysis between Inflammopathic and Adaptive clusters could be reflective of opposite modulation of similar pathways; this is further suggested by the strong inverse correlation between the Inflammopathic and Adaptive clusters in Figure 2. As above, these biological insights might allow for hypotheses about guided treatments for different subtypes. Still, we only included subjects at admission for sepsis; whether and how these profiles might change depending on time since initial infection onset, longitudinally during treatment, or whether patients might move between subtypes over time, is unknown.

Two independent research groups have identified sepsis subgroups similar to those described here: one focused on pediatric sepsis in a US-based cohort (9, 10); the other focused on adult sepsis in UK-based cohorts (12, 13). Notably, the two subgroupings do not broadly overlap. Comparison of our three clusters with the prior clusterings yielded several interesting findings. First, using subject-level comparisons, patients assigned to the Inflammopathic cluster were mostly assigned to Endotype B (11) or SRS 1 (12). However, Endotype B conferred a lower mortality in children compared to Endotype A, while SRS 1 conferred a higher mortality in adults compared to SRS2. Still, we are reassured that these independent studies identified the same grouping of patients using completely separate techniques. Similarly, patients assigned to the Adaptive cluster were primarily assigned to SRS 2; both studies identified this as a low mortality group associated with interferon signaling. We also identified a third (Coagulopathic) cluster. It is possible that the substantially larger sample size and greater heterogeneity of our discovery cohorts compared to prior work allowed us to detect this third Coagulopathic cluster.

Our study has some limitations. First, we provided validation only in historical independent datasets, not in a prospectively collected cohort. This limited us to only non-targeted gene expression profiling (microarrays and RNAseq). Second, we examined only datasets of patients with bacterial sepsis at admission, because the clustering algorithms may otherwise have been overwhelmed by the differing host responses to different types of infections (15, 19). The coming availability of rapid host-response diagnostics to distinguish between bacterial and viral infections (15, 40, 51) suggests that the cluster classifier could be applied to patients after diagnosis with bacterial infections. However, it is unknown whether these subtypes exist in patients with viral or fungal sepsis, or in non-infected critically ill patients. Third, one of the validation datasets (GSE74224) re-used 21 samples (20% of its total) from one of the discovery datasets (GSE28750), although they were re-profiled using a different technology (29, 42). Exactly which samples are duplicates are unknown, so they could not be removed; however, this makes up less than 4% of the total validation samples, suggesting that results are unlikely to be affected. Finally, we have presented analyses for all clinical variables that were available in more than one study at the sample level. This led to the inclusion of some analyses that were individually underpowered. In addition, variables may not be missing at random; it is thus possible that missingness biases the outcome (for instance, by not reporting mortality in less-severe cohorts). The various weaknesses make clear that a prospective clinical study of the clusters will be necessary to confirm and extend our results.

Overall, we used state-of-the-art methods in bioinformatics and data analysis to create the largest known unbiased pool of sepsis transcriptomic profiles, and to then show that three robust, distinguishable clusters exist across the sepsis spectrum. These sepsis clusters could feature prominently in the clinical trials domain, where they may serve as an enrichment tool or a companion diagnostic. The confirmation that multiple subtypes exist within the host response will hopefully enable more research into a precision medicine approach for sepsis.