Molecular Subphenotypes in Sepsis and Acute Respiratory Distress Syndrome

Sepsis and Acute Respiratory Distress Syndrome (ARDS) are heterogeneous critical illness syndromes. Molecular subphenotyping shows promise in furthering our understanding of these syndromes and may offer a more precise method for identifying appropriate therapies. 

Current Definitions of Sepsis and Acute Respiratory Distress Syndrome

Sepsis and the acute respiratory distress syndrome (ARDS) are heterogeneous clinical entities that afflict critically ill patients. Based on current consensus definitions, these syndromes are diagnosed based on clinical characteristics readily available to most clinicians. Sepsis is a dysregulated host response to infection that results in life-threatening organ dysfunction (Singer et al. 2016; Meyer and Prescott 2024). Based on updated consensus criteria, it is defined by an increase in a sequential organ failure assessment (SOFA) score of 2 or more points from baseline or a quick SOFA score of 2 or more in the context of suspected infection. The primary focus of this update was to highlight the role of organ dysfunction that differentiates sepsis from uncomplicated infection. The new definition places less emphasis on systemic inflammatory response syndrome, which was the main criterion for diagnosing sepsis based on the previous iteration. Concerns regarding the effect of this change on recognising sepsis among critically ill patients have led to studies evaluating the concordance between these two iterations. Among patients admitted to the intensive care unit (ICU), there is good agreement between the two definitions in recognising sepsis, but the previous Sepsis-2 definition more systematically captured septic shock compared to current sepsis definitions.

ARDS is characterised by bilateral lung opacities and acute hypoxaemic respiratory failure not explained by cardiogenic pulmonary oedema (Wick et al. 2024). Conceptually, ARDS is an acute diffuse inflammatory lung injury precipitated by a risk factor such as pneumonia, extrapulmonary sepsis, aspiration or trauma. A recent global consensus conference expanded the definition of ARDS to address the disparities in the detection of ARDS based on available resources and therapies applied such as invasive ventilation (Matthay et al. 2024). Given the increasing use of high-flow nasal oxygen, the expanded definition considers pathophysiological changes that may be shared among patients who are on high-flow nasal oxygen and intubated patients (Wick et al. 2022), which may allow for earlier recognition and treatment of ARDS. Furthermore, it considers the fact that invasive ventilation and arterial blood gas monitoring which may not be readily available in resource-limited settings and includes patients on oxygen and a peripheral pulse oximetric oxygen saturation (SpO₂) to the fraction of inspired oxygen ratio <315 if the SpO₂ £ 97% (Jayasimhan and Matthay 2024).

The definitions of these syndromes prioritise early detection of critically ill patients at a higher risk of death using easily available clinical characteristics. However, using these definitions likely captures a heterogeneous group of patients with differing underlying pathobiology, disease behaviour and characteristics. Furthermore, these simple clinical criteria do not characterise the complex pathological features of these diverse syndromes in detail. This may provide one reason why the development of clinically effective pharmacologic or cell-based therapies for sepsis and ARDS have been limited.

Subphenotypes in Critical Illness Syndromes

Considerable efforts have been made over the last decade to identify meaningful underlying subgroups within these heterogeneous syndromes. These may be on the basis of clinical, physiological or molecular characteristics (Matthay et al. 2020; Gordon et al. 2024). A better understanding of the different molecular subtypes may explain the lack of clinical efficacy of biological therapies that show pre-clinical promise or heterogeneity in treatment effects within clinical trials of patients with sepsis or ARDS. The aims of this article are to describe different molecular subtypes identified, evaluate the differences in patient-centred outcomes among patients based on the underlying molecular subphenotype, briefly summarise the evidence of heterogeneity in the treatment effect of different interventions among patients with sepsis or ARDS based on the underlying molecular subphenotype, and consider some areas for further research.

Molecular Subphenotype

The European Society of Intensive Care Medicine (ESICM) guidelines on ARDS recently defined a subphenotype as “a distinct subgroup (of ARDS patients) that can be reliably discriminated from other subgroups based on a set or pattern of observable or measurable properties” (Graselli et al. 2023). Discrimination is typically based on a data-driven assessment of a multidimensional description of traits. Subphenotypes should also be reproducible in different populations. Molecular subphenotypes for the purpose of this review are a subgroup of patients with sepsis or ARDS that share similar biological factors which either portend a different prognosis or show response to specific treatments. This may be on the basis of a single biomarker or multiple biomarkers. Molecular subphenotypes may hold extended promise over clinical and physiologic subphenotypes as they may provide better insight into the underlying host response and distinct pathological changes which are proximal to clinical or physiological changes.

Molecular Subphenotypes in ARDS

Single biomarkers

A number of important prognostic biomarkers in both ARDS have been discovered over the past two decades. These discoveries have been on the basis of plasma protein biomarkers, as well as protein biomarkers detected in bronchoalveolar lavage fluid. These biomarkers include measures of inflammation, lung epithelial injury, fibroproliferation, oxidative injury, endothelial injury, as well as extrapulmonary organ dysfunction (van der Zee et al 2020; Antonucci et al. 2024; Jayasimhan et al. 2021).

Among protein biomarkers measured in bronchoalveolar lavage fluid associated with the pathophysiology and outcomes of patients with ARDS, alveolar type-II pro-collagen peptide (PCP-III) has shown excellent accuracy in detecting fibroproliferation in patients with ARDS (Clark et al. 1995; Chestnut et al. 1997). These findings have subsequently been replicated, correlating the degree of fibroproliferation on cross-sectional imaging with measured PCP-III in bronchoalveolar lavage fluid (Forel et al. 2015). 

Several plasma protein biomarkers have shown an association with outcomes in patients with ARDS; however, it is unclear if these plasma proteins merely correlate with disease severity or offer insights into the underlying pathophysiology, which if targeted with specific therapies, may improve outcomes. An example includes the use of plasma receptor for advanced glycation end products (RAGE), a measure of type-1 alveolar epithelial injury (Uchida et al. 2006). A secondary analysis of the ARDS network low tidal volume ventilation trial showed patients with a high baseline plasma RAGE had a greater reduction in mortality with low tidal volume ventilation than those with low plasma RAGE (Calfee et al. 2008). This suggests some patients may benefit more from lower tidal volume ventilation than others and raises the possibility that ultra-lung protective ventilation with very low tidal volumes may benefit a subset of patients with ARDS, perhaps with more alveolar epithelial injury. In addition, RAGE proved to have both pathogenetic and prognostic value in studies in patients with COVID-19 pneumonia, especially when combined with elevated SARS-CoV-2 viral antigen levels (Wick et al. 2022; Matthay et al. 2023; Wick et al. 2024).

Multiple biomarkers

The use of multiple biomarkers to identify subphenotypes has shown promise in terms of prognostication as well as potential in predicting response to specific interventions. Using unsupervised clustering methods, two subphenotypes have been repeatedly identified in patients with ARDS from multiple cohorts.

An important study by Calfee and colleagues first identified these two subphenotypes using an unsupervised approach, latent class analysis, based on clinical and biologic criteria of patients with ARDS enrolled in two large, randomised trials (Calfee et al. 2014). One subphenotype exhibited a hyperinflammatory subphenotype characterised by high levels of inflammatory biomarkers such as serum interleukin-6 (IL-6), interleukin-8 (IL-8), tumour necrosis factor (TNF) receptor 1 (TNFR1) and low levels of serum bicarbonate and protein C. Using the same unsupervised clustering method, these two subphenotypes have been replicated in multiple cohorts of patients enrolled in randomised controlled trials and observational studies.

A simplified 3 and 4-variable model developed using a machine-learning algorithm both accurately predicted these phenotypes when derived from three ARDS clinical trial cohorts and validated in a fourth trial (Sinha et al. 2020). Both simplified models also showed good accuracy in predicting the hyperinflammatory and hypoinflammatory subphenotypes when externally tested in a further two cohorts of patients with ARDS enrolled in clinical trials. This parsimonious approach may be of value for identifying these subphenotypes in real time for clinical trials. A similar subphenotype was also identified among patients using a hierarchical clustering model in a European cohort of patients with ARDS (Bos et al. 2017).

The prevalence of the hyperinflammatory phenotype among these studies ranges from 27-34% of patients among multiple cohorts of patients with ARDS. The aforementioned studies have consistently shown that the hyperinflammatory subphenotypes are associated with an almost two-fold increase in the risk of death compared to the hypoinflammatory phenotype. These studies also suggest heterogeneity in treatment effect from specific therapies. A re-analysis of the Assessment of Low tidal Volume and elevated end-expiratory volume to Obviate Lung Injury (ALVEOLI) trial showed a significant interaction between subphenotype and treatment allocation (Calfee et al. 2014). Patients with the hyperinflammatory subphenotype had reduced mortality when randomised to a high PEEP strategy compared to a low PEEP strategy. Similar interactions have been observed in other clinical trials of patients with ARDS. A re-analysis of the Fluid and Catheter Treatment Trial (FACTT) showed patients with a hyperinflammatory subphenotype had a greater survival benefit when randomised to a restrictive fluid strategy (Famous et al. 2017). A study of patients enrolled in a trial of simvastatin treatment of patients with ARDS showed no significant difference in treatment effect in the overall cohort but exhibited heterogeneity in treatment effect based on subphenotype (Calfee et al. 2018). Patients with the hyperinflammatory subphenotype randomised to simvastatin had a reduced mortality compared to placebo, while this interaction was not observed among patients with the hypoinflammatory subphenotype.

Among patients with COVID-19 related ARDS, multiple studies have demonstrated the presence of these two subphenotypes. Patients with the hyperinflammatory subphenotype experienced a greater reduction in mortality when treated with glucocorticoids compared to those with the hypoinflammatory subphenotype (Sinha et al. 2021). It has been suggested that the higher mortality rate seen among COVID-19 associated ARDS was higher than historical cohorts of ARDS patients due to a higher prevalence of the hyperinflammatory phenotype among patients with COVID-19 ARDS. However, a study by Sinha and colleagues using a parsimonious regression model with limited variables found the prevalence of the hyperinflammatory phenotype to be no higher than in previous cohorts of ARDS among a small subset of patients admitted with COVID-19 ARDS in the United Kingdom (Sinha et al. 2020). Furthermore, another study identified novel subphenotypes based on latent class analysis in hospitalised patients with COVID-19 that were part of a clinical trial (Alipanah-Lechner 2024).

Molecular Subphenotypes in Sepsis

Single biomarkers

Single biomarkers including acute phase proteins, chemocan ligands, soluble receptor proteins, acute phase proteins, and soluble urokinase plasminogen activator receptor (sUPAR), have shown utility in predicting an increased risk of death among patients with sepsis (He et al. 2024). Thus far, few studies have evaluated the value of these biomarkers in predicting a response to specific treatments.

A dose-ranging early-phase randomised trial by Reinhart and colleagues showed promise in the use of an IL-6 cut-off  > 1000 pg/ml in predicting response to treatment with tumour necrosis factor antibody treatment (Reinhart et al. 1996). These results, however, were not replicated in a larger randomised controlled study by the same group (Reinhart et al. 2001).

Higher levels of plasma cell-free haemoglobin, a potential mediator of the pro-inflammatory response and oxidative injury, are associated with increased mortality in patients with sepsis. An observational study of critically ill patients showed the use of paracetamol, a potent haemoprotein reductant may attenuate the adverse effects of cell-free haemoglobin and is associated with improved survival in patients with sepsis and an elevated cell free haemoglobin (Janz et al. 2013). A subsequent phase two randomised controlled trial showed no improvement in days alive and free of organ support in patients with sepsis (Ware et al. 2024). However, patients treated with paracetamol were less likely to develop ARDS and had significantly greater improvements in SOFA scores on days 1 through 4. No significant interaction was observed between baseline cell-free haemoglobin levels and treatment with paracetamol, but there were numerical differences that favoured benefit from paracetamol; the study may have been underpowered to detect this subgroup difference. A larger phase three randomised and blinded trial could evaluate these interesting findings with measurement in real time of plasma levels of cell-free haemoglobin and focusing acetaminophen versus placebo treatment on those patients with a higher level of cell-free haemoglobin (> 10 mg/dl).

Multiple biomarkers

A number of studies have evaluated the use of multiple biomarkers to cluster patients into subphenotypes that may be associated with increased mortality and predict response to specific treatments. These unsupervised clustering methods have classified patients into subtypes based on transcriptomic data (Davenport et al. 2016; Burnham et al. 2017), clinical, and molecular biomarkers (Scicluna et al. 2017; Seymour et al. 2019).

Wong and colleagues evaluated the use of genome-wide expression profiling to evaluate the presence of subphenotypes among paediatric patients with sepsis (Wong et al. 2009). Using an unsupervised hierarchical clustering model, they identified three subphenotypes of sepsis in this paediatric cohort. Subclass A was associated with increased illness severity, organ dysfunction and higher mortality. These subclasses have been validated in subsequent studies and the first two subclasses were consistently identified, subclass A and B, which correspond with adaptive immunity and glucocorticoid receptor signalling (Wong et al. 2011). These endotypes were also reproduced using a rapid multiplex digital messenger ribonucleic acid (mRNA) quantification platform (Wong et al. 2015). These endotypes have subsequently shown heterogeneity in treatment effect of glucocorticoids among patients with septic shock enrolled in a multifactorial clinical trial of vasopressin and hydrocortisone. Patients with Endotype A, with downregulation of the adaptive immune response and glucocorticoid receptor signalling had an increased mortality when treated with hydrocortisone compared to placebo (Wong et al. 2021). A similar model which showed prognostic utility in paediatric septic shock was shown to have predictive utility in an animal model of sepsis. In a study using the Pediatric Sepsis Biomarker Risk Model (PERSEVERE), a 12-gene-based biomarker model was used to guide antimicrobial dosing found in mice with a higher PERSEVERE score, indicating a greater bacterial burden could be rescued by higher doses of antibiotics (Wong et al. 2019).

In a study of adults admitted to U.K. intensive care units with sepsis, Davenport and colleagues discovered two distinct subphenotypes of sepsis response signatures (SRS) based on peripheral blood leukocyte transcriptomic analysis (Davenport et al 2016). The class SRS-1 exhibited evidence of immunosuppression with downregulation of Human Leukocyte Antigen Class-II, T-cell exhaustion and endotoxin tolerance and was associated with a higher mortality rate than SRS-II. These findings have been replicated in patients with faecal peritonitis and severe community-acquired pneumonia. Interestingly, the host response was not predicted by the source of infection. Much like the subphenotypes identified by Wong and colleagues, there is evidence to suggest a treatment interaction between the underlying SRS class and treatment with hydrocortisone. A secondary analysis of mRNA data available for a subset of patients enrolled in the VANISH trial showed treatment with hydrocortisone was associated with higher mortality (Antcliffe et al. 2019). Seymour and colleagues derived and validated four clinical subphenotypes of sepsis from two observational cohorts (Seymour 2019). They reproduced these subphenotypes in a subsequent cohort of patients with severe community-acquired pneumonia and three cohorts of patients with sepsis enrolled in clinical trials. They found a strong correlation between the clinical subphenotypes and host response biomarkers. The γ and δ phenotypes had higher measures of inflammation and abnormal coagulation compared to the α or β phenotypes.

Cohen and colleagues have attempted to use candidate gene expression for glucocorticoid receptor signalling to identify patients with septic shock who would benefit from treatment with hydrocortisone (Cohen et al. 2021). Although none of the candidate gene expressions were associated with improved outcomes when assigned to hydrocortisone treatment, four genes unrelated to the candidate genes showed an interaction with hydrocortisone treatment. These findings, however, were not corrected for multiple comparisons and remain hypothesis-generating.

The hyper-inflammatory and hypoinflammatory subphenotypes exhibited among patients with ARDS have also been identified among patients with sepsis. An analysis of two prospective observational cohorts and two cohorts of patients enrolled in randomised controlled trials replicated the same two-class model with a clinical classifier model with strong concordance with the two-class model among ARDS patients (Sinha et al. 2023). These findings remained robust despite sensitivity analyses excluding patients with ARDS. Among patients enrolled in a clinical trial evaluating the effectiveness of Activated Protein C in sepsis, there was heterogeneity in treatment effect based on subphenotype. When compared to placebo, APC treatment was associated with lower mortality in the hyperinflammatory subphenotype and higher mortality in the hypoinflammatory subphenotype. 

Translational Testing of Subphenotypes in Clinical Trials

Despite strong evidence suggesting the presence of underlying clinical and transcriptomic subtypes in both sepsis and ARDS, challenges remain in translating this knowledge to improve clinical outcomes. The main barrier to implementing these approaches clinically has been the difficulty in identifying these subphenotypes at the bedside. Most unsupervised methods require a large number of variables and specialised testing (e.g. gene-expression panels). However, single specific biomarker variables such as cell-free haemoglobin or simplified limited 3 variable models that measure IL-6 and TNFR1 and the lowest serum bicarbonate can now be done in real-time and will be the basis for a phase two randomised trial to begin in the U.K. and will be extended to the US and other countries (https://panthertrial.org).

In conclusion, with the advent of molecular phenotyping, the management of critically ill patients could be more precise and targeted when compared to currently employed methods of clustering (Figure 1).

Conflict of Interest

Dilip Jayasimhan has no conflicts of interest to declare. Dr Matthay has received research grants from Roche-Genentech and Quantum Health and has been an ARDS consultant for Merck Pharmaceuticals and CSL Behring.