Do you know your "sepsis phenotypes"? (Review)

Don't call it "sepsis phrenology." DON'T.

From the 1990s through, let’s say, 2011 (when Xigris™ was removed from the market), sepsis research meant molecule-hunting. Drug companies backed hundreds of investigations led by the Indiana Joneses of academic medicine as they sought the Cure for Sepsis in the heart of darkness. With the level of brainpower and resources being deployed, a cure seemed to be just around the corner.

Dozens of distinct compounds were tested in more than 100 phase II/III clinical trials: afelimomab, eritoran, various anti-TNF monoclonal antibodies and receptor fusion protein, anti-endotoxin antibodies, and many more.

None of them ever showed a replicable mortality benefit in phase III trials. After Xigris’s™ withdrawal from the market, it seemed like the search for the sepsis cure had reached a dead end, or run out of road, or maybe its horse had become very tired. The metaphor doesn’t matter. What mattered was making sure science kept moving forward, and NIH’s annual sepsis research budget had somewhere to flow.

There was a ray of hope: in those negative trials, there were often intriguing suggestions of benefits in subgroups (e.g., those with elevated IL-6 or some other cytokine). Could the benefit in these subgroups have been diluted to non-significance in the larger pool of undifferentiated septic patients?

As the classic sign-off from a hundred thousand manuscripts goes, more research would be required.

Sepsis Phenotypes: Some Assembly Required

Sepsis produces tremendous variation and complexity in physiology and immune response, but also significant overlap between patients.

So the first pass at defining sepsis phenotypes consists of collecting massive amounts of data through high-throughput systems like transcriptomics, proteomics, metabolomics, as well as clinical data from RCTs, and commanding computers to find patterns in it.

Unsupervised learning algorithms mine datasets for statistical correlations between constellations of patient factors and outcomes. Next, the phenotype construct is applied to other available trial data or prospectively collected data from registries or new RCTs (“validated”). If the correlation persists in the new data set, a new phenotype is christened and imbued with meaning. Rarely, however, have phenotypes been prospectively evaluated in clinical trials, largely because the methods to identify them are so cumbersome.

Broadly, sepsis phenotypes have been constructed from clinical or immunologic variables (although this choice is arbitrary and convenience-based, and should disappear as research methods evolve).

Clinical/Physiology Sepsis Phenotypes

The prototype for “clinical phenotypes” of sepsis is an analysis that was dubbed SENECA (JAMA 2019). Among 20,189 patients meeting Sepsis-3 criteria at 12 UPenn hospitals, 29 clinical variables (vitals, labs, vasopressor use, etc) were identified and fed to machine learning tools, which found clusters of the variables that could be stratified by mortality. Greek letters were added as a garnish.

Voilá, the apparent discovery of four clinical phenotypes of sepsis: α, β, γ, and δ. Mortality rises from 5% to 40% from alpha to delta. Gammas have more inflammation and lung dysfunction, while deltas have more liver dysfunction, shock, and mortality.

You may wonder what this adds to clinical gestalt (i.e., shocky patients with DIC die more often), or whether the patterns reflect any real categories or are just assembled by a computer largely out of epiphenomena and chance associations, or question how the schema could be used either for research or clinical purposes.

Clinical phenotypes did not “scale”, but a vast new research enterprise seeks to identify sepsis phenotypes tied to more fundamentally causative factor(s) at the biochemical or immunologic levels.

Biochemical/Immunologic Sepsis Phenotypes (Endotypes)

While clinical sepsis phenotypes don’t claim to be more than hypothesis-generating or as potential filters for clinical trial enrollment, this category of sepsis phenotyping seeks to identify subsets of patients with distinctly different operative immunologic processes.

These phenotypes get their own name: endotypes, to emphasize that there is Something Inside that is producing the observed manifestations, a Something we might precisely identify and influence for the patient’s benefit. (Using endotype also ditches the baggage of the word phenotype as used in phrenology, eugenics, etc.)

And investigators really are measuring something inside, a whole lot of it. Endotypes are proposed/constructed out of the patterns computers discover inside staggeringly complex “multi-omics” combinatory datasets that can include the expression of 20,000 genes, thousands of proteins, and metabolites per patient, each of which evolves over hours during sepsis.

Human researchers then work with the computer-identified patterns, each of which may implicate one known sepsis pathway over another, or suggest potential biochemical targets.

For example, two sepsis patients might both be hypotensive with elevated lactate, but one has sepsis driven primarily by hyperinflammation (an endotype detectable by elevated IL-6, e.g.) while the other by immunosuppression (an endotype which could be identified by low HLA-DR expression on monocytes on flow cytometry, e.g.).

If these two patients could be quickly differentiated at presentation, the argument goes, steroids could be given to the hyperinflammatory patient and withheld from the immunosuppressed patient, tailoring treatment beneficially to each’s need. (Keep in mind, this is all theoretical, today.)

Here’s just a slice of the research in sepsis phenotypes-slash-endotypes:

SRS and MARS

Unsupervised machine learning on transcriptomics signatures (gene expression) of peripheral blood leukocytes in patients with sepsis led Davenport et al and Scicluna et al to propose distinct endotypes of sepsis, with variable immune response patterns.

Davenport et al’s group proposed the SRS1 and SRS2 endotypes:

• Immunosuppressed (SRS1) pattern: T-cell exhaustion, downregulation of HLA class II

• Hyper-inflammatory/immune-competent (SRS2), with more normal inflammatory signaling

SRS1 (immune-suppressed) had higher mortality than the hyperinflammatory SRS2 group, and this was replicated in a cohort of 106 patients.

Seven genes were implicated in the differential gene expression between clusters. This 7-gene screen was then applied post hoc to data from 176 patients enrolled in the VANISH trial (JAMA 2016), suggesting that hydrocortisone may have harmed patients with immune-competent sepsis: hydrocortisone use was associated with an odds ratio of ~8 for increased mortality in those with the immunocompetent/hyperinflammatory SRS2 phenotype.

However, the trail stopped there. A bedside PCR test for the 7-gene expression profile has never been tested prospectively, and no test is commercially available.

A separate group, the MARS consortium identified four patterns (MARS 1 through 4) using expression from 140 genes: an immune competent/adaptive, immunosuppressed, hyperinflammatory, and interferon-gamma driven response to sepsis.

Subsequent groups (e.g., Chenoweth et al) have more or less replicated similar broadly sketched patterns, often with nuances or sub-subtypes added. This research has thus far defied systematization because of the protean variability in patients and their illness states, the differences in biological samples, and methodologies between research labs, among other factors.

Macrophage-activation-like syndrome (MALS) and “immunoparalysis”

An uncommon but particularly lethal observed sepsis pattern is macrophage activation-like syndrome (MALS), which produces cytokine storm predominantly mediated by IL-1, similar to hemophagocytic lymphohistiocytosis (HLH). Only about 3% of sepsis patients display features of MALS, but they have >60% mortality.

The ImmunoSep trial represents the single instance to date in which investigators have selected patients based on prospectively defined sepsis endotypes, and applied targeted interventions in randomized, blinded (partially) fashion.

---

---

MALS patients were identified as those with sepsis criteria and very high ferritin levels (>4420 ng/mL) and were randomized to receive an intervention (IL-1 blockade with anakinra vs placebo).

Patients in this trial with a different endotype, so-called immunoparalysis (<5000 human leukocyte antigen DR receptors on CD45/CD14 monocytes, as computed by flow cytometric analysis) could receive gamma-interferon or placebo.

Organ dysfunction (SOFA score) was reduced in the intervention groups, but 90-day mortality was virtually identical at ~68%.

And Proteomics, Cytokinomics, Metabolomics, (Freakonomics?) … All Together Now

The same essential move—perform a tricorder scan on the patient, collect an enormous volume of data, and feed it into a pattern-finding computer algorithm—can be performed with any high-throughput stream of biologic data.

This has produced claims that there are four “proteomic” subtypes, three “cytokine signature clusters,” at least two flow cytometry subtypes, and three “metabolomic subtypes” in sepsis, according to just a few recent publications.

The next frontier: feeding all these various streams (-omics) into one computationally gargantuan analysis (multi-omics). This is in its relative infancy.

There’s a sense that after the identification ten years ago of a handful of immunologic patterns, or endotypes, progress has slowed. There hasn’t been a process of continual refinement, but rather a rhyme-and-repeat vibe, describing the same associations in slight variations or riffs.

This is highly reminiscent of the early complex-genomics era, when a firehose of associations between SNPs and disease states seemed to show that geneticists were just about to crack this whole human disease thing wide open. Eventually, it became clear they were just reporting on the emissions, not understanding the engine, never mind showing us how to fix it.

Ontology Before Epistemology

The high-throughput technologies and computational processes that create the sepsis phenotypes are indeed impressive. They provide near-miraculous X-ray vision into the real-time immunologic and physiologic processes unfolding in this notoriously heterogeneous syndrome, identifying real, possibly vital differences between patients that can’t be detected at the bedside.

It remains to be seen, though, whether any of these suddenly visible and astoundingly complex pathophysiologic phenomena are modifiable to patients’ benefit (beyond what we already routinely do).

It’s possible (and in fact likely) that endotypes mistake phenomenology for mechanisms: endotypes can be real statistical clusters of phenomena, and also have no practical use.

The real answers most certainly lie deeper, and are beyond the capacities or the role of this newsletter to ascertain. The most intriguing thing I read (or skimmed) was also the last: this 2025 Nature Medicine SUBSPACE consortium paper, which plumbs the existing datasets to propose a consensus immune-dysregulation framework in critical illness more broadly, drawing a throughline between sepsis, ARDS, and trauma.

Although high-throughput techniques won’t hand us the Unified Theory of Sepsis anytime soon, they currently seem like the only shot we have of wrestling its heterogeneity complexity into more orderly and compliant shapes. This would at the very least allow for more rationally designed clinical trials like ImmunoSep.

The new catchphrase is “precision medicine,” but that label seems a bit premature: first, we need to see the “medicine” part in action.

References

Moore AR, Zheng H, Ganesan A, et al. A consensus immune dysregulation framework for sepsis and critical illnesses. Nature Medicine. Published online September 30, 2025:1-13. doi:https://doi.org/10.1038/s41591-025-03956-5

‌Giamarellos-Bourboulis EJ, Kotsaki A, Kotsamidi I, et al. Precision Immunotherapy to Improve Sepsis Outcomes: The ImmunoSep Randomized Clinical Trial. JAMA. Published online August 2025:e2524175. doi:https://doi.org/10.1001/jama.2025.24175

Classification of Patients With Sepsis According to Blood Genomic Endotype: A Prospective Cohort Study. The Lancet. Respiratory Medicine. 2017. Scicluna BP, van Vught LA, Zwinderman AH, et al.

Derivation, Validation, and Potential Treatment Implications of Novel Clinical Phenotypes for Sepsis. The Journal of the American Medical Association. 2019. Seymour CW, Kennedy JN, Wang S, et al.

Predicting Sepsis Severity at First Clinical Presentation: The Role of Endotypes and Mechanistic Signatures. EBioMedicine. 2021. Baghela A, Pena OM, Lee AH, et al.

The Immune Landscape of Sepsis and Using Immune Clusters for Identifying Sepsis Endotypes. Frontiers in Immunology. 2023. Tang G, Luo Y, Song H, et al.

Classification of Patients With Sepsis According to Immune Cell Characteristics: A Bioinformatic Analysis of Two Cohort Studies. Frontiers in Medicine. 2020. Zhang S, Wu Z, Chang W, et al.

Defining Critical Illness Using Immunological Endotypes in Patients With and Without Sepsis: A Cohort Study. Critical Care. 2023. Balch JA, Chen UI, Liesenfeld O, et al.

A Transcriptomic Classifier Model Identifies High-Risk Endotypes in a Prospective Study of Sepsis in Uganda. Critical Care Medicine. 2024. Cummings MJ, Bakamutumaho B, Tomoiaga AS, et al.

Profiling the Dysregulated Immune Response in Sepsis: Overcoming Challenges to Achieve the Goal of Precision Medicine. The Lancet. Respiratory Medicine. 2024. Cajander S, Kox M, Scicluna BP, et al.

Sepsis Phenotypes, Subphenotypes, and Endotypes: Are They Ready for Bedside Care?. Current Opinion in Critical Care. 2024. Scherger SJ, Kalil AC.