The molecular ICU: a primer on omics, informatics and the future of precision critical care

Abstract

Critical care has produced hundreds of neutral randomized trials, in part because therapies have been tested in biologically incoherent populations that dilute meaningful treatment effects. Syndromic diagnoses such as sepsis, acute respiratory distress syndrome, and traumatic brain injury group distinct pathobiological states under a single label, limiting the ability to detect treatment-responsive subgroups. While high-dimensional omics technologies have revealed this biologic heterogeneity, the field lacks a practical framework to translate these insights into clinical trial design and bedside decision-making. This Review addresses the translational gap in precision critical care through a pathway-level framework. We synthesize advances in genomics, transcriptomics, proteomics, and metabolomics, highlighting their roles in capturing susceptibility, host response, effector function, and real-time physiology. We propose pathway-focused biomarkers as clinically translatable signatures that preserve biological mechanisms while enabling practical measurement. We outline how pathway enrichment, network analysis, and multi-omic integration can identify these programs, and how feature selection can derive parsimonious biomarker panels for clinical use. This approach also supports pathway-guided drug repurposing by linking dysregulated molecular programs to existing therapies. Together, these signatures provide a framework for predictive enrichment, aligning patient selection with therapeutic mechanisms and facilitating implementation within adaptive platform trials. This Review serves as a practical primer that outlines the concepts and methods needed to translate omics into clinically actionable tools. By shifting from syndromic classification to pathway-defined biology, it provides a framework for biomarker development, trial design, and precision critical care.

Introduction

Why syndromes fail: biological heterogeneity in critical illness

For decades, critical care has been defined by syndromes. Syndromes such as sepsis, acute respiratory distress syndrome (ARDS), and traumatic brain injury (TBI) are treated as coherent disease entities, enrolled into clinical trials as unified populations, and targeted with therapies intended to apply broadly. Yet despite extraordinary scientific effort, transformative therapies in these syndromes remain non-existent. Clinical trials can cost between approximately $11 million and well over $100 million depending on design complexity and therapeutic area [1]. In critical care, where trials have been disproportionately neutral, this has translated into the expenditure of hundreds of millions of dollars by funding agencies with minimal therapeutic return [1]. Similarly, these non-significant trials invite reflection on the ethical justification for continuing similar approaches, particularly if the underlying assumptions and design strategies remain unchanged. The repeated shortcomings of critical care trials are no longer surprising. It reflects a structural flaw in how critical illness is conceptualized. Syndromes are clinical conveniences, not biological truths.

Sepsis, ARDS, and traumatic brain injury are not single biological entities, but heterogeneous syndromes encompassing diverse inflammatory, metabolic, vascular, and immune programs despite similar clinical presentations [2–4]. Consequently, therapies tested across these biologically diverse populations may produce benefit in some patients but not others, obscuring meaningful therapeutic signals at the trial level. This heterogeneity of treatment effect (HTE) has been repeatedly demonstrated in re-analyses of critical care trials [5–9]. The problem is not that therapies have failed, but that they have been tested in biologically incoherent populations. Over the past decade, critical care has increasingly shifted from syndromic labels toward biologic subgroups defined by high-dimensional omics technologies, including transcriptomics, proteomics, metabolomics, and multi-omic integration [10, 11]. These approaches, coupled with advanced bioinformatics, have revealed coordinated host-response pathways and reproducible subphenotypes in sepsis, ARDS, and TBI that differ in prognosis and, in some cases, treatment response [4, 12–15]. Yet despite this progress, a critical gap remains. Molecular subgroups have been discovered and rapid bedside identification has recently become feasible, but translation into routine clinical decision-making remains the next frontier [16, 17]. High-dimensional omics can define molecular states and bioinformatics can identify biologic patterns, but a unifying framework is still needed to convert these discoveries into actionable therapeutic strategies.

In this Review we introduce the concept of ‘pathway-focused biomarkers’ as a translational framework for precision critical care medicine. Pathway-focused biomarkers are parsimonious, biologically informed signatures derived from coordinated molecular programs that preserve mechanistic information while remaining suitable for clinical implementation. By reducing thousands of molecular signals into clear, biologically meaningful pathway signatures, this approach bridges the gap between omics discovery and practical use in clinical trials and treatment decisions.

By synthesizing evidence across sepsis, ARDS, and TBI, we outline how pathway-driven multi-omic approaches can (1) more precisely characterize underlying patient pathobiology, (2) provide refined biological strategies for predictive enrichment in critical care trials, and (3) accelerate the translation and implementation of mechanism-aligned therapeutics. Collectively, the promise of the Molecular ICU does not lie in measuring more biomarkers; rather, it depends on organizing biology at the pathway level.

From bedside biomarkers to bedside omics

Biomarkers provide the interface between molecular biology and bedside decision-making in the ICU, offering objective signals of organ injury and treatment response (Box 1). ICU practice has long relied on a small set of single-analyte markers that anchor therapeutic decisions (Table 1). These markers reveal a recurring pattern: clinically useful biomarkers tend to be biologically coarse, while those offering deeper mechanistic insight remain clinically unavailable. No single analyte captures the molecular complexity of these syndromes [18, 19].

Box 1: Definitions and terminology in precision critical care
Omics: Technologies that comprehensively measure large sets of biological molecules (genes, transcripts, proteins, metabolites) to characterize biological systems and their functional states
High-dimensional: Datasets containing a large number of measured variables or features relative to the number of observations, often encompassing hundreds to thousands of molecular measurements per sample
Phenotype: A patient subgroup defined by observable clinical or physiologic features, regardless of underlying mechanism
Subphenotype: A biologically coherent subgroup identified through statistical or data-driven methods using clinical or molecular data. Subphenotypes may differ in outcomes and, in some cases, treatment response, but the specific underlying mechanism driving the observed differences is not fully elucidated
Endotype: A subgroup defined by a shared, clearly delineated underlying molecular mechanism, establishing not only that patients differ biologically, but how, at the level of specific cell types, molecular pathways, or causal programs. Endotypes represent the most mechanistically grounded form of patient classification
Predictive enrichment: A trial strategy that selectively enrolls patients who are more likely to respond to a specific therapy based on biological, clinical, or molecular characteristics identified before treatment
Prognostic enrichment: A trial strategy that selectively enrolls patients at higher risk of a clinical outcome or disease progression in order to increase event rates and improve trial efficiency/
Prognostic biomarker: A biomarker that predicts clinical outcome independent of treatment. Prognostic biomarkers identify high- or low-risk patients but do not indicate whether a patient will respond differently to a specific intervention
Predictive biomarker: A biomarker that identifies patients more likely to benefit (or be harmed) by a specific treatment. Predictive biomarkers enable targeted therapy by stratifying heterogeneous populations into those who will and will not respond to a given intervention
Treatable trait: A measurable biological or clinical characteristic that is causally linked to a disease mechanism and can be targeted by a specific intervention. Treatable traits operationalize precision medicine by connecting a measurable feature to a therapeutic strategy, moving beyond diagnosis-based to mechanism-based treatment
Pathway-focused biomarker: Parsimonious, biologically coherent signatures that capture coordinated molecular activity within dysregulated pathways to reflect the dominant biological processes driving an individual patient’s disease state
Heterogeneity of treatment effect (HTE): Variation in the direction or magnitude of a treatment’s effect across subgroups of a trial population, such that an intervention may benefit some patients, have no effect in others, or potentially cause harm in certain populations. In critical illness, HTE may arise from biological, physiologic, etiologic, or baseline risk differences across patients

Adapted from: Sinha et al. (2023) [18], Gordon et al. (2024) [188], Maslove et al.(2022) [10]

Table 1.

Canonical biomarkers and their relevance to precision critical care

Biomarker	Biological process captured	Clinical role	Key strengths	Key limitations	Biomarker availability
Sepsis
Lactate [189]	Tissue hypoperfusion and metabolic stress	Risk stratification at presentation; resuscitation guidance; serial clearance used as a global marker of response to therapy	Rapid, inexpensive, universally available; strong association with mortality	Non-specific; elevated in non-hypoxic states (β-agonists, liver dysfunction, seizures)	Routine clinical
Procalcitonin (PCT) [19, 56, 63, 64, 190–193]	Innate immune response to bacterial PAMPs	Support for bacterial infection likelihood; guide antibiotic discontinuation and duration in selected populations	Predictable kinetics; supported by randomized trials for stewardship	Limited specificity; less reliable in early infection, trauma, or surgery	Routine clinical
C-reactive protein (CRP) [19, 62]	IL-6 driven acute-phase response	Monitoring inflammatory trajectory over time; adjunct to clinical assessment	Widely available; low cost	Poor specificity; delayed kinetics; minimal mechanistic insight	Routine clinical
Interleukin-6 (IL-6) [59–61]	Early innate inflammatory signaling	Early severity assessment; biological characterization in research and trials	Strong correlation with severity and mortality	Short half-life; limited availability for real-time bedside use	Available; limited bedside use
suPAR [57, 58]	Chronic immune activation and leukocyte trafficking	Baseline risk stratification and triage; identifies patients at high risk of deterioration	Biologically stable; reflects global immune dysregulation	Limited utility for monitoring response; not treatment-guiding	Available; not universally routine
Acute respiratory distress syndrome
Angiopoietin-2 (Ang-2) [194–196]	Endothelial activation and vascular instability	Biological stratification; prognostic enrichment in trials	Reproducible association with severity and mortality	Not lung-specific; not routinely available clinically	Research-grade
von Willebrand factor (vWF) [194–196]	Endothelial injury and activation	Marker of vascular injury burden; prognostic enrichment	Reflects systemic endothelial damage	Non-specific to lung injury	Available; not routine in ICU
sRAGE [197, 198]	Alveolar epithelial injury	Identification of epithelial-dominant lung injury; mechanistic stratification	Strong biological relevance to alveolar damage	Limited assay standardization and availability	Research-grade
Surfactant protein D (SP-D) [199]	Alveolar epithelial and surfactant dysfunction	Prognostic and biological characterization of lung injury	Lung-relevant biology	Assay variability; primarily research use	Research-grade
Interleukin-8 (IL-8) [200, 201]	Neutrophil recruitment and activation	Inflammatory subphenotyping; prognostic enrichment	Consistently associated with outcomes across cohorts	Reflects systemic inflammation, not lung-specific	Research-grade
Club Cell Secretory Protein (CC16) [202]	Airway epithelial injury and increased alveolar-capillary permeability	Biological characterization of epithelial-predominant lung injury; prognostic enrichment in ARDS and lung injury syndromes	Lung-specific biology; reflects epithelial barrier disruption; associated with ARDS severity and outcomes	Influenced by renal clearance; limited assay standardization; not routinely available clinically	Research-grade
PAI-1 [203, 204]	Impaired fibrinolysis and coagulation imbalance	Identification of prothrombotic lung injury biology; prognostic stratification	Integrates coagulation and inflammation	Not specific to ARDS; limited clinical availability	Research-grade
Brain injury
Neuron-specific enolase (NSE) [54, 55, 205–208]	Neuronal cell injury	Neuroprognostication after hypoxic-ischemic injury (post-cardiac arrest)	Validated thresholds in guidelines	Susceptible to hemolysis; influenced by renal failure Lengthy processing time (multi-day) at most centres	Routine clinical
S100B [209–214]	Astroglial injury	Early detection of brain injury; adjunct prognostication	Rapid kinetics; used in some triage algorithms	Limited specificity for CNS injury	Routine clinical
GFAP [54, 55, 215, 216]	Astroglial injury and blood-brain barrier disruption	Diagnosis of intracranial injury; outcome prognostication	High brain specificity; FDA-authorized assays	Captures astroglial injury only	Routine clinical; FDA-authorized
UCH-L1 [73]	Neuronal cell body injury	Complementary diagnostic and prognostic marker to GFAP	Neuron-specific; FDA-authorized	Limited insight into secondary injury mechanisms	Routine clinical; FDA-authorized

Bold — Routine clinical: Lactate, PCT, CRP, NSE, S100B, GFAP, UCH-L1. Bold italic — Available; limited/not universally routine: IL-6, suPAR, vWF. Bold underline — Research-grade: Ang-2, sRAGE, SP-D, IL-8, CC16, PAI-1

Until recently, high-dimensional molecular profiling required centralized facilities and turnaround times incompatible with critical care. That bottleneck is now lifting. Yet, significant technical challenges remain, including quality control, establishment of gold standards, multiplexing capacity, interferent management, and the temporal dynamics of biomarker measurement [20–22]. However, multiplexed immunoassays and electrochemical biosensors can quantify protein panels within minutes [23, 24], targeted transcriptomic panels are achievable in under 30 min [25–27], and benchtop NMR and microfluidic platforms increasingly enable rapid metabolite profiling [28–30]. Prospective work in both ARDS and sepsis has now demonstrated that biological subphenotypes can be identified at the bedside using rapid clinically deployable platforms, with turnaround times compatible with ICU decision-making [16, 17]. The constraint is no longer whether biology can be measured at the bedside, but whether we use it to align therapy with mechanism.

Beyond single-analyte thinking: a pathway-level framework

Moving beyond single-analyte thinking requires a new translational unit. We propose ‘pathway-focused biomarkers’ - parsimonious, mechanistically coherent signatures of coordinated biological activity - as that unit. Rather than expanding panels of single analytes or deploying thousands of omic features at the bedside, pathway-focused biomarkers capture the dominant biological program of an individual patient by reflecting groups of related molecules acting together within a defined system and indicating whether that system is activated, suppressed, or failing at a given moment. In doing so, they distil thousands of molecular signals into interpretable readouts of biology, with opportunity to preserve the mechanistic depth of omics while remaining practical for clinical deployment.

Constructing pathway-focused biomarkers requires knowing which pathways matter, which markers within them carry the signal, and how those programs differ between patients with the same clinical syndrome. This is not achievable through targeted measurement of a few candidate analytes; it requires the unbiased, high-dimensional measurement of biology that only omics technologies provide. Omics platforms serve as discovery engines for identifying pathway-focused biomarkers.

Omics as the discovery engine of precision critical care

Omics refers to the high-dimensional, comprehensive measurement of entire classes of biological molecules: genes (genomics), RNA (transcriptomics), proteins (proteomics), and metabolites (metabolomics). To contextualize these approaches in critical illness, it is important to understand the types of omics technologies, and the type of biological information each provides. Genomics can be viewed as the biological instruction manual, defining the DNA sequence that confers inherited susceptibility to host responses, but remains static during acute illness. Transcriptomics represents the cellular readout of these instructions, capturing which genes are actively transcribed in response to injury, infection, or stress at any given moment. Proteomics reflects the molecular machinery in action, quantifying the proteins that execute inflammatory, endothelial and metabolic programs that mediate organ dysfunction. Metabolomics captures the downstream byproducts of these processes, profiling small molecules generated by enzymatic activity and substrate flux that integrate genetic background, transcriptional regulation, and environmental influences. Understanding what each modality measures, and what it does not measure, is essential for interpreting multi-omic signatures, and for designing precision strategies that are biologically coherent and clinically actionable in critical illness. The complementary biological information captured by genomics, transcriptomics, proteomics, and metabolomics, and their respective roles in characterizing critical illness biology, is summarized schematically in Fig. 1.

Fig. 1.

Complementary biological information captured by omics technologies in critical illness. Schematic overview of the major omics modalities used in critical care research and the distinct biological information each captures. Genomics defines inherited susceptibility and static risk architecture; transcriptomics reflects dynamic gene-expression responses to injury, infection, and therapy; proteomics quantifies functional effectors that mediate inflammation, endothelial injury, coagulation, and organ dysfunction; and metabolomics captures downstream biochemical activity integrating cellular metabolism, substrate flux, and environmental influences. Together, these complementary layers provide a hierarchical view of critical illness biology, from predisposition to real-time physiological state, illustrating why no single modality can fully characterize complex ICU syndromes. Of note, although omic layers are often presented within a temporal hierarchy, their turnover is highly variable, with some transcripts changing rapidly over minutes to hours, whereas certain proteins may persist considerably longer, resulting in substantial temporal overlap across molecular layers

Genomics: the blueprint

Genomics examines the complete DNA sequence of an organism, providing a blueprint that influences downstream biology. In critical illness, genomic approaches have identified variants that affect susceptibility to sepsis, inflammatory responses, and risk of organ dysfunction [31, 32]. Genome-wide association studies (GWAS) comparing DNA across many people have revealed loci linked to ARDS susceptibility, sepsis mortality, and response to vasopressors [33, 34].

The principal strength of genomics is its temporal stability. The genome does not change in response to acute illness (with the exception of epigenetic changes). This makes genomics well suited to identify pre-existing risk factors and to inform prospective enrichment strategies. Genomic biomarkers can be measured at any time, including before ICU admission, without repeated sampling. Polygenic risk scores aggregating many variants provide quantitative measures of inherited susceptibility that may enhance prognostic models when combined with clinical variables [33]. For example, this could mean identifying patients genetically predisposed to vasopressor hyporesponsiveness before shock develops. Yet, genomic variants explain only a modest fraction of outcome variance in critical illness, lack mechanistic interpretability, and cannot capture the dynamic host responses that drive acute trajectories [33].

Transcriptomics: the real-time host response

Transcriptomics quantifies the RNA molecules expressed by cells at a given moment, providing a dynamic readout of gene activity in response to injury or treatment [35]. In critical illness, transcriptomic profiling has identified subphenotypes that predict clinical trajectories and immune states in sepsis, as well as inflammatory subphenotypes in ARDS [13, 36–39].

The key strength of transcriptomics is its ability to capture the real-time host response. Unlike genomics, gene expression changes rapidly with infection, tissue injury, and therapy, enabling tracking of disease evolution, identification of time-sensitive therapeutic windows, and monitoring of treatment response. Transcriptomic subphenotypes can show differential treatment responses. In the VANISH trial, patients with the immunocompetent SRS2 subphenotype had higher mortality when given corticosteroids than with placebo, suggesting that transcriptomic profile at septic shock onset may be associated with response to therapy [13]. Transcriptomic signatures often map directly onto pathways (interferon signaling, complement activation, metabolic reprogramming), providing mechanistic insight alongside prognosis. The scalability of RNA-seq and emerging rapid platforms make transcriptomic subphenotyping increasingly feasible within hours of ICU admission [16].

Yet, RNA instability demands stringent pre-analytical handling, transcript abundance does not reliably predict protein-level activity, and high dimensionality increases susceptibility to batch effects and overfitting across platforms and cohorts [40, 41]. A further limitation is that most ICU transcriptomic subphenotyping uses bulk whole-blood RNA, which aggregates signal across leukocyte populations and obscures the cell type or tissue of origin, constraining mechanistic interpretation and rational therapeutic targeting [42]. Single-cell RNA sequencing (scRNA-seq) directly addresses this by profiling individual cells, identifying disease-associated cell states invisible in bulk data; for example, scRNA-seq identified an immunosuppressive CD14 + monocyte state (MS1) expanded preferentially in the SRS1 sepsis subphenotype, providing a cellular basis for a previously bulk-derived classification [42, 43]. Because scRNA-seq remains costly and poorly suited to large clinical cohorts, cellular deconvolution methods (CIBERSORTx, MuSiC, BayesPrism) computationally estimate cell-type proportions and cell-type-specific expression from bulk RNA using reference signatures from purified populations or scRNA-seq atlases, enabling retrospective cell-resolved analysis of existing ICU cohorts at scale (for example, recovery of MS1 abundance from archived whole-blood transcriptomes) [44–46]. Together, scRNA-seq for discovery and deconvolution for scaled application offer a tractable path toward cell-resolved subphenotyping in critical illness.

Proteomics: where mechanism meets clinical phenotype

Proteomics characterizes the levels, modification states, and interactions of proteins, the main functional drivers of cellular processes. Because proteins execute the enzymatic, signaling, and regulatory roles encoded by genes, proteomic profiling reflects disease biology closer to clinical phenotype than nucleic acid-based measurements. In critical illness, plasma proteomics has revealed patterns of endothelial activation, complement imbalance, coagulopathy, and organ-targeted injury that correlate with outcomes in sepsis, ARDS, and TBI [47–53]. These are the same processes clinicians infer indirectly from rising lactate, falling platelets or worsening oxygenation; proteomics quantifies them directly. At the present state of the field, sepsis and ARDS subphenotypes and associated molecular assays are more well characterized than TBI. Nevertheless, in TBI, blood protein biomarkers have begun to reach clinical use. Glial fibrillary acidic protein (GFAP) and ubiquitin C-terminal hydrolase‑L1 (UCH‑L1) are FDA-authorized for evaluation of mild traumatic brain injury, the first clinically approved blood assay for brain injury [54, 55].

Proteomics’ key strength is its close alignment with functional pathobiology. Proteins directly mediate inflammation, coagulation, vascular barrier disruption, and organ failure, which are central to critical illness. Proteomics can capture post-translational modifications (phosphorylation, glycosylation, proteolytic cleavage) that modulate function without corresponding mRNA changes, revealing regulatory complexity beyond transcriptomics. Many widely used clinical biomarkers (procalcitonin [56], suPAR [57, 58], IL‑6 [59–61], CRP [62], and presepsin [63, 64]) are proteins associated with inflammation, severity, and mortality in sepsis [19], underscoring the translational relevance of proteomic findings. Circulating proteins are generally more stable than RNA, simplifying processing and enabling retrospective studies with stored specimens. Yet, the extreme dynamic range of the plasma proteome, high instrumentation costs, limited cross-platform concordance, and the confounding effects of protein synthesis, clearance, and consumption in acute illness constrain both discovery and validation [65–68]. Measured protein levels reflect the net effects of synthesis, secretion, consumption, and clearance, which can change independently in acute illness and complicate interpretation.

Metabolomics: the cellular energy state, exhaust and flux

Metabolomics profiles small-molecule intermediates and end-products of metabolism (“cellular exhaust”), providing a functional readout of enzymatic activity, substrate availability, and energy state. Because metabolites arise downstream of genetic, transcriptomic, proteomic, and environmental factors, metabolomic profiles reflect real-time cellular status and offer immediate insights into dysregulated pathways. In critical illness, metabolomics has revealed major perturbations in energy metabolism (lactate, ketones, tricarboxylic acid cycle intermediates), amino acid handling (tryptophan-kynurenine pathway, branched-chain amino acids), lipid mediators (eicosanoids, sphingolipids), and redox homeostasis (glutathione, oxidized lipids) that correlate with clinical parameters in ARDS, sepsis, and TBI [69–73].

Metabolomics’ unique strength is its integrative nature. Metabolite concentrations reflect cumulative effects of genes, transcripts, proteins, and environment, providing a proximal readout of physiological state. Unlike fixed genotypes or cell-specific transcripts, circulating metabolites represent systemic biochemistry accessible via routine blood sampling. Many metabolites (lactate, glucose, creatinine, bilirubin) are already measured clinically, highlighting translational feasibility. Metabolic pathways are often druggable, so identifying dysregulated nodes can suggest targets such as immunometabolic modulators, amino acid supplementation, or lipid mediator inhibition.

Metabolomics also has substantial limitations. Metabolite concentrations are highly sensitive to pre-analytical variables (fasting, sampling time, hemolysis, storage temperature, freeze-thaw cycles), confounded by diet, medications, and the microbiome, and many detected spectral features remain unannotated, limiting both reproducibility and biological interpretation [74–76].

The microbiome: a counterpart to host responses

The host response in critical illness is shaped by microbial communities that colonize, infect, and translocate across mucosal barriers. Microbiome profiling characterizes the composition and diversity of microbial communities and links them to host responses and outcomes, typically using 16 S rRNA gene sequencing or shotgun metagenomics. Metagenomic next-generation sequencing (mNGS) applies similar unbiased sequencing technology to a clinical specimen for pathogen identification, detecting bacteria, fungi, viruses, and parasites without targeted assays [77, 78]. Microbiome studies have demonstrated that dysbiosis contributes to host-response heterogeneity in pneumonia, ARDS, and sepsis. In ARDS, the lung microbiome becomes enriched with gut-associated bacteria, correlating with alveolar and systemic inflammation [79], and lung microbiota composition independently predicts outcomes in mechanically ventilated patients [80]. Gut dysbiosis in sepsis is similarly associated with mortality and modulates susceptibility to secondary infection via the gut-lung axis [81]. Critically, recent evidence indicates that pathogen characteristics themselves independently shape host-response subphenotype assignment. Enterobacterales infections are strongly associated with the hyperinflammatory sepsis subphenotype independent of illness severity, suggesting that integration of pathogen identity with host-response biomarkers may resolve heterogeneity that neither dimension captures alone [82].

The principal strengths of these approaches are unbiased detection, integration of host and microbial signal from the same specimen, and direct clinical actionability. Key limitations include contamination susceptibility in low-biomass samples, difficulty distinguishing colonization from infection, and constraints of cost and turnaround time [83]. Despite these, host-microbe interactions add a critical dimension to molecular subphenotyping that pathway-focused biomarker frameworks should increasingly incorporate.

Subphenotypes: resolving hidden biological structure

Beyond characterizing individual molecular layers, omics approaches have increasingly been used to identify biologically distinct subphenotypes within critical illness syndromes [10, 22, 84]. These studies show that patients with similar clinical presentations can have different underlying molecular programs [11, 35, 85], leading to differences in prognosis, biology, and treatment response that may not be apparent with traditional syndrome-based classifications [5, 7, 8, 10]. Emerging subphenotypes in critical illness are summarized in Fig. 2; Table 2.

Fig. 2.

Emerging molecular subphenotypes in critical-illness syndromes. Conceptual representation of molecular subphenotypes across major critical illness syndromes, including sepsis, acute respiratory distress syndrome (ARDS), and traumatic brain injury (TBI). Despite shared clinical phenotypes, patients segregate into biologically coherent subgroups defined by distinct host-response programs, such as inflammatory activation, immune suppression, endothelial and coagulation dysfunction, metabolic dysregulation, and neuroinflammatory or neuronal injury pathways. These subphenotypes are identified through transcriptomic, proteomic, and integrative multi-omic analyses and are associated with differences in clinical trajectories, outcomes, and, in select contexts, heterogeneity of treatment effect. Subphenotyping reframes critical illness classification around underlying mechanism rather than syndromic label, providing a foundation for pathway-focused biomarker development and predictive enrichment in precision critical care. While ARDS and sepsis have begun to demonstrate reproducible omics-defined subphenotypes associated with heterogeneity of treatment effect, comparable predictive enrichment strategies in TBI remain in earlier stages of development

Table 2.

Main subphenotypes identified in critically-ill patients with sepsis, ARDS or TBI

Subphenotype	Dominant dysregulated biological programs		Primary omic modality	Clinical associations	Evidence for heterogeneity of treatment effect	Level of evidence/study design
Sepsis
Inflammopathic [18, 86, 87, 217]	Innate immune activation: • Enrichment for innate immune pathways • Pro-inflammatory cytokine expression • Neutrophil activation • Pattern recognition signaling	Whole blood transcriptomics (multi-cohort) 33-mRNA classifier		• High mortality (~ 18–30%) • High severity scores • Elevated CRP, IL-6 • Higher secondary infection rates	COVID-19 validation: • Subphenotype validated in viral sepsis • Associated with elevated CRP Theoretical: May benefit from IL-1 or IL-6 blockade	Post-hoc (randomization not stratified by subphenotype) Transcriptomic subphenotyping applied to existing RCT cohorts; COVID-19 external validation (observational) Prospective subphenotype assignment: ORANGES trial (subphenotype assigned, treatment not randomized to subphenotype)
Adaptive [18, 86, 87, 217]	Adaptive immune activation: • Enrichment for adaptive immune pathways • Balanced immune response • Intact lymphocyte function	Whole blood transcriptomics (multi-cohort) 33-mRNA classifier		• Lowest mortality (~ 5–8%) • Low severity scores • ~39–44% of patients • Lower secondary infection rates	ORANGES Trial: • Lowest mortality confirmed prospectively COVID-19: • No deaths in Adaptive group Theoretical: May not require immunomodulation	Post-hoc (randomization not stratified by subphenotype) Transcriptomic subphenotyping applied to existing RCT/observational cohorts; COVID-19 external validation (observational) Prospective subphenotype assignment: ORANGES trial (subphenotype assigned, treatment not randomized to subphenotype)
Coagulopathic [18, 86, 87, 217]	Coagulation dysregulation: • Disrupted coagulation pathways • Endothelial dysfunction signatures • Platelet activation • Fibrinolysis abnormalities	Whole blood transcriptomics (multi-cohort) 33-mRNA classifier		• Highest mortality (~ 25–42%) • Elevated D-dimers • ~27% of patients • Clinical coagulopathy	COVID-19 validation: • Highest D-dimers in Coagulopathic group • 42% mortality Theoretical: May benefit from early anticoagulation. Target for endothelial-directed therapy	Post-hoc (randomization not stratified by subphenotype) Multi-cohort transcriptomic discovery; COVID-19 external validation (observational) Prospective subphenotype assignment: ORANGES trial (subphenotype assigned, treatment not randomized to subphenotype)
SRS1 [12]	Immunosuppression: • T-cell exhaustion • HLA class II downregulation • Endotoxin tolerance • Impaired antigen presentation • Enriched cell death/apoptosis pathways	Whole blood leukocyte transcriptomics 7-gene classifier (DYRK2, CCNB1IP1, TDRD9, ZAP70, ARL14EP, MDC1, ADGRE3)		• Higher 14-day mortality (HR 2.4–2.8) • ~41% of patients • Higher organ dysfunction • Lower monocyte HLA-DR • Immature neutrophil expansion	VANISH Trial (post-hoc): • No differential response to vasopressin vs. norepinephrine • No significant change in mortality with hydrocortisone Dynamic: ~50% transition to SRS2 within 5 days	Post-hoc (randomization not stratified by subphenotype) Transcriptomic subphenotypes applied retrospectively to VANISH RCT HTE analysis: subgroup interaction, not pre-specified enrichment
SRS2 [12]	Immunocompetent: • Preserved adaptive immune function • Intact T-cell activation • Normal HLA class II expression • Appropriate antigen presentation	Whole blood leukocyte transcriptomics 7-gene classifier (DYRK2, CCNB1IP1, TDRD9, ZAP70, ARL14EP, MDC1, ADGRE3)		• Lower 14-day mortality • ~59% of patients • Less severe illness • Better baseline immune status	VANISH Trial (post-hoc): • Significant interaction with hydrocortisone (p = 0.02) • INCREASED mortality with hydrocortisone (OR 4.6, 95% CI 1.5–14.4) Suggests potential HARM from steroids in immunocompetent patients	Post-hoc (randomization not stratified by subphenotype) Retrospective application to VANISH RCT Interaction p = 0.02 for hydrocortisone; caution: risk-based HTE not confirmed as biologically-mediated predictive enrichment
MARS1 [38, 88]	Immune paralysis/exhaustion: • Downregulation of innate immune genes (TLR, NF-κB) • Downregulation of adaptive immune genes (T-cell receptor signaling) • Impaired antigen presentation • Upregulation of metabolic pathways (heme biosynthesis) Biomarkers: BPGM↑, TAP2↓	Whole blood transcriptomics 140-gene classifier; 8-gene bedside signature		• Highest 28-day mortality (39%) • ~29% of patients • Associated with septic shock • Higher APACHE scores • Increased 1-year mortality	Limited direct RCT evidence • Aligns with CTS2 (lymphoid dysregulation) • Aligns with Coagulopathic subphenotype • Theoretical target for immune stimulation (e.g., IFN-γ, GM-CSF)	No RCT evidence Discovery cohort only; no prospective or interventional validation
MARS2 [38, 88]	Hyperinflammatory: • Pattern recognition receptor upregulation • NF-κB activation • IL-6 signaling • iNOS pathway activation • fMLP signaling	Whole blood transcriptomics 140-gene classifier; 8-gene bedside signature		• Moderate 28-day mortality (~ 22%) • ~34% of patients • Elevated inflammatory markers	Limited direct RCT evidence • Aligns with CTS1 (myeloid dysregulation) • Aligns with Inflammopathic subphenotype • Theoretical target for anti-inflammatory therapies	No RCT evidence Discovery cohort only; no prospective or interventional validation
MARS3 [38, 88]	Adaptive immune activation: • Heightened adaptive immune gene expression • T-cell activation pathways preserved• Intact lymphocyte function	Whole blood transcriptomics 140-gene classifier; 8-gene bedside signature		• Lower 28-day mortality (~ 23%) • ~23% of patients • Better prognosis	Limited direct RCT evidence • Aligns with CTS3 (lymphoid protective) • Aligns with SRS2 and Adaptive subphenotypes • May not benefit from immunomodulation	No RCT evidence Discovery cohort only; no prospective or interventional validation
MARS4 [38, 88]	Interferon/myeloid protective: • Interferon signaling • B-cell development pathways • IL-4 signaling • Mature neutrophil and monocyte responses	Whole blood transcriptomics 140-gene classifier; 8-gene bedside signature		• Moderate 28-day mortality (~ 33%) • ~13% of patients • Variable prognosis	Limited direct RCT evidence • Part of myeloid protective cluster • May represent viral infection signatures	No RCT evidence Discovery cohort only; no prospective or interventional validation
CTS1 [38]	Myeloid dysregulation: • Inflammatory pathway activation • Endothelial activation • Immature neutrophil signatures • Emergency granulopoiesis Consensus alignment: SRS1 + MARS2 + Inflammopathic	Consensus transcriptomic framework 18-gene classifier		• High mortality • Fewer ventilator-free days • Greater organ dysfunction • Elevated inflammatory biomarkers	Integrates findings across subphenotyping schemas • Consolidates evidence from SRS, MARS • Framework for future trial enrichment	Consensus framework only No independent RCT validation; integrative analysis across cohorts
CTS2 [38]	Lymphoid dysregulation: • Coagulation dysfunction • Lymphocyte suppression • Immune exhaustion features Consensus alignment: MARS1 + Coagulopathic	Consensus transcriptomic framework 18-gene classifier		• High mortality • Coagulation abnormalities • Lymphopenia	Integrates findings across subphenotyping schemas • Represents immunosuppressed/coagulopathic overlap • Potential for immune stimulation + anticoagulation	Consensus framework only No independent RCT validation; integrative analysis across cohorts
CTS3 [38]	Lymphoid protective: • Preserved adaptive immune function • Intact T-cell activation • Normal antigen presentation Consensus alignment: SRS2 + MARS3 + Adaptive	Consensus transcriptomic framework 18-gene classifier		• Lower mortality • Better clinical outcomes • Preserved immune function	Integrates findings across subphenotyping schemas • Represents favourable immune profile • May not require immunomodulation • Caution with immunosuppressive therapies	Consensus framework only No independent RCT validation; integrative analysis across cohorts
Hyperinflammatory [218]	Amplified systemic inflammation: • Elevated IL-8, IL-6, sTNFR-1 • Elevated ICAM-1 (endothelial activation) • Low protein C; elevated PAI-1 • Dysregulated coagulation/fibrinolysis • Metabolic acidosis (low bicarbonate) • Organ dysfunction (↑ creatinine, ↑ bilirubin, ↓ platelets)	Plasma proteomics + clinical data: Latent class analysis (VALID, EARLI) Clinical classifier model (CCM; XGBoost) applied to PROWESS-SHOCK, VASST Parsimonious classifier: IL-8, protein C, sTNFR-1, bicarbonate, vasopressor use		• ~29–42% of sepsis patients • Higher mortality (OR 2.2–2.6) • In-hospital mortality 43–45% (VALID/EARLI) • 28-day mortality 36–51% (PROWESS-SHOCK/VASST) • Fewer ICU-free days • More vasopressor use/shock • More bacteremia	PROWESS-SHOCK (activated protein C): • Significant treatment interaction with APC (p = 0.0043) • Survival BENEFIT from APC (APC 32% vs. placebo 39%) • Continuous-probability interaction (p = 0.0048) VASST (vasopressin): • No interaction with vasopressin vs. norepinephrine (p = 0.72)	Post-hoc (randomization not stratified by subphenotype) LCA in two observational cohorts (VALID, EARLI); CCM applied to PROWESS-SHOCK and VASST RCTs (post-hoc) HTE with APC observed retrospectively; not confirmed as biologically-mediated predictive enrichment
Hypoinflammatory [218]	Lower systemic inflammation: • Lower plasma inflammatory biomarkers (IL-8, IL-6, sTNFR-1) • Preserved protein C; lower PAI-1 • Normal/near-normal bicarbonate • Less organ dysfunction; higher platelets • Less metabolic derangement • Majority class; biologically heterogeneous	Plasma proteomics + clinical data: Latent class analysis (VALID, EARLI) Clinical classifier model (CCM; XGBoost) applied to PROWESS-SHOCK, VASST Parsimonious classifier: IL-8, protein C, sTNFR-1, bicarbonate, vasopressor use		• ~58–71% of sepsis patients (majority) • Lower mortality • In-hospital mortality 17–20% (VALID/EARLI) • 28-day mortality 20–28% (PROWESS-SHOCK/VASST) • More ICU-free days • Less shock/vasopressor use • Less bacteraemia	PROWESS-SHOCK (activated protein C): • APC associated with HARM (APC 23% vs. placebo 17%) VASST (vasopressin): • No differential response to vasopressin vs. norepinephrine. Majority class remains biologically undifferentiated	Post-hoc (randomization not stratified by subphenotype) Same LCA/CCM analyses as Hyperinflammatory phenotype No prospective enriched trial confirming differential treatment benefit
Acute respiratory distress syndrome
Hyperinflammatory [3, 14, 18, 36, 89–95]	Amplified systemic inflammation: • Elevated IL-6, IL-8, sTNFR-1 • Low protein C • Metabolic acidosis (low bicarbonate) • Innate immune activation • Enhanced glycolysis • IFN-γ and T-cell activation (lung)		Plasma proteomics + clinical data: Latent class analysis Parsimonious classifier: IL-8, protein C, bicarbonate	• ~30% of ARDS patients • Higher mortality (HR ~ 2) • Fewer ventilator-free days • More vasopressor use/shock	HARP-2 Trial (simvastatin): • Significant survival BENEFIT (p = 0.008) • 13% absolute mortality reduction ARMA/ALVEOLI/FACTT: • BENEFIT from higher PEEP • BENEFIT from conservative fluid	Post-hoc (randomization not stratified by subphenotype) Latent class analysis applied to ARMA, ALVEOLI, FACTT RCTs (post-hoc) HARP-2: prospective enrichment-by-design (randomized, but treatment not selected by subphenotype) HTE not confirmed as biologically-mediated predictive enrichment
Hypoinflammatory [3, 14, 18, 36, 89–95]	Lower systemic inflammation: • Lower plasma inflammatory biomarkers • Preserved protein C levels • Normal/near-normal bicarbonate • Less metabolic derangement		Plasma proteomics + clinical data: Latent class analysis Parsimonious classifier: IL-8, protein C, bicarbonate	• ~70% of ARDS patients • Lower mortality • More ventilator-free days • Less shock	HARP-2 Trial (simvastatin): • No benefit from simvastatin ARMA/ALVEOLI/FACTT: • Different response to PEEP • Different response to fluid strategy	Post-hoc (randomization not stratified by subphenotype) Same post-hoc analyses as Hyperinflammatory phenotype No prospective enriched trial confirming differential treatment benefit
Traumatic brain injury
Early-inflammatory [97]	Early systemic immune activation: • Elevated IL-6 • Elevated IL-15 • Elevated MCP-1 • Systemic inflammatory response • Monocyte/macrophage activation		Plasma/serum proteomics: Multiplex inflammatory panel (30 mediators) Hierarchical clustering	• Unfavourable outcomes (GOS-E ≤ 4) • Older patients (median 49–62 yrs) • ~60% in validation cohort • NOT associated with injury severity	No RCT evidence yet Hypothesis-generating: • May be target for immunomodulation • Suggests stratified approach needed	No RCT evidence Discovery cohort; observational only; hypothesis-generating
Pauci-inflammatory [97]	Minimal early inflammation: • Lower inflammatory mediator concentrations • Attenuated systemic immune response		Plasma/serum proteomics: Multiplex inflammatory panel (30 mediators) Hierarchical clustering	• Relatively better outcomes • Younger patients (median 24–32 yrs) • NOT associated with injury severity	No RCT evidence yet Hypothesis-generating: • Immunomodulation may be unnecessary • Different therapeutic targets needed	No RCT evidence Discovery cohort; observational only; hypothesis-generating

All heterogeneity of treatment effect (HTE) analyses presented in this table are post-hoc subgroup analyses of trials not prospectively designed for subphenotype-based enrichment. Observed differential treatment effects may reflect risk-based heterogeneity rather than biologically-mediated predictive enrichment. No critical care syndrome has yet demonstrated definitive pathway-level evidence for biologically-mediated predictive enrichment, and these two concepts should not be conflated

HTE, heterogeneity of treatment effect; RCT, randomized controlled trial; VANISH, Vasopressin vs. Norepinephrine as Initial Therapy in Septic Shock; ARMA, ARDSnet Low Tidal Volume trial; ALVEOLI, Assessment of Low Tidal Volume and Elevated End-Expiratory Volume to Obviate Lung Injury; FACTT, Fluid and Catheter Treatment Trial; HARP-2, Hydroxymethylglutaryl-CoA Reductase Inhibition with Simvastatin in Acute Lung Injury to Reduce Pulmonary Dysfunction

Transcriptomic analyses in sepsis consistently identify molecular subphenotypes with distinct immune and endothelial programs associated with differences in organ dysfunction, mortality, and, in secondary analyses, differential responses to therapies such as corticosteroids [12, 18, 38, 86–88]. Integrated transcriptomic and proteomic studies in ARDS reproducibly define hyperinflammatory and hypoinflammatory subphenotypes that differ in clinical outcomes and exhibit opposing responses to fluid management, ventilation strategies, and statin therapy [3, 14, 18, 36, 89–95]. Unsupervised clustering in ICU TBI cohorts has identified biologically distinct subphenotypes based on GCS and composite metabolic stress profiles (lactate, oxygen saturation, glucose, base excess), which improve outcome prediction beyond clinical severity scores and show that metabolic derangement can outweigh neurological severity for prognosis [4, 96]. Proteomic and transcriptomic profiling in TBI reveals reproducible early inflammatory and pauci-inflammatory molecular states that are independent of injury severity and strongly associated with neurological outcomes, but have not yet been shown to modify treatment response [97].

Implications for precision critical care

Across sepsis, ARDS, and TBI, omic-informed subphenotypes consistently demonstrate three defining features: they capture biologically coherent host-response programs, they are associated with outcomes and trajectories, and in select contexts they identify heterogeneity of treatment effect that is obscured by traditional syndrome-based trial designs [5, 7, 8, 10, 18]. These properties position subphenotyping as foundational to predictive enrichment. By shifting classification from syndromes toward mechanisms, subphenotyping bridges high-dimensional molecular data and precision trial design. Yet single-omic measurements capture only one biological dimension at a time, motivating the integration of complementary molecular layers [22, 84, 98]. Because subphenotypes defined by single molecular layers capture only fragments of disease biology, pathway-focused biomarkers must integrate complementary omic dimensions to faithfully represent coordinated host-response programs.

Multi-omic integration: why one omic layer is never enough

Why integrate?

Multi-omics is the simultaneous measurement and integration of complementary molecular layers (DNA, RNA, proteins, metabolites) to characterize biological processes. Critical illness is uniquely suited to a multi-omic approach to subphenotype patients because it evolves rapidly, affects multiple organs, and reflects systemic dysregulation that unfolds across biological layers. These syndromes represent dynamic host-response states in which molecular programs shift over hours to days, and no single omic modality can fully capture the processes determining trajectory, therapeutic responsiveness, or recovery. Transcriptomic, proteomic, and metabolomic signals frequently diverge within the same patient, not as technical artifact, but because each platform interrogates distinct regulatory, effector, and physiological dimensions of disease biology [71, 99].

The rationale for integration is therefore conceptual as much as technical. First, coordinated molecular programs may emerge only when layers are analyzed jointly, remaining invisible within single-omic analyses [100]. Second, critical illness perturbs interconnected biological systems, with effects propagating across transcripts, proteins, and metabolites; integrating complementary layers yields deeper mechanistic insight than any individual modality alone [101]. Third, multi-omic integration can enhance predictive modeling and patient stratification [102]. Most importantly, omic layers sample distinct anatomic compartments and cellular sources (proteomics captures tissue-specific leakage proteins, metabolomics reflects whole-organism extracellular biochemistry, and transcriptomics often samples circulating immune cells), such that integration approaches pan-compartmental and pan-cellular resolution unattainable from any single layer [103].

Each omic layer operates on its own mechanistic and temporal scale [99, 104, 105]. Transcriptomics reflects regulatory intent, proteomics captures effector proteins and tissue injury, and metabolomics provides a near real-time snapshot of physiological state and metabolic flux, while epigenomic, single-cell, and spatial approaches add longer-lived biological memory and cellular context. These temporal hierarchies mirror critical illness itself: metabolic collapse may precede transcriptional reprogramming, proteomic injury signatures may persist despite relative transcriptomic quiescence, and therapeutic decisions must align with the dominant biological state rather than static diagnostic labels.

Importantly, discordance across omic layers is often biologically informative rather than contradictory. mRNA abundance frequently fails to predict protein levels or activity due to post-transcriptional regulation, altered translation, degradation, or cellular exhaustion [99, 104]. Metabolic perturbations without parallel transcriptomic shifts may reflect mitochondrial dysfunction, substrate depletion, microcirculatory impairment, or immunometabolic rewiring [106, 107]. In sepsis, profound metabolic derangement may coexist with an apparently muted transcriptome, consistent with energetic failure rather than overt transcriptional activation [71, 108, 109], while immune paralysis may manifest as reduced antigen-presentation transcripts despite proteomic evidence of ongoing innate effector activity [12, 110]. Consider a septic patient whose transcriptome shows robust interferon signaling, but whose metabolome reveals ATP depletion, mitochondrial dysfunction, and lactate accumulation. Despite evidence of immune activation, impaired cellular energetics may limit the capacity to mount an effective response, highlighting the need to consider both immune and metabolic states. These cross-layer mismatches are intrinsic to critical illness biology and cannot be inferred from any single modality, underscoring the necessity of multi-omic integration to avoid erroneous system-level conclusions.

How to integrate

Implementing multi-omics in critical care requires realism about ICU data. Cohorts are often modest in size, heterogeneously sampled, and vulnerable to missingness and batch effects. Two broad integration strategies are commonly used (Fig. 3A). Parallel integration analyzes each omic layer separately and then synthesizes findings at the interpretation stage [81, 87]. This approach is generally robust and interpretable, particularly when concordant pathway-level signals emerge across layers. However, it may fail to detect coordinated biological structure that becomes apparent only when datasets are analyzed jointly. Integrated approaches, in contrast, combine data layers within a shared analytic framework to identify coordinated multi-omic programs [81, 87]. These methods can reveal deeper biological relationships but are more sensitive to preprocessing decisions, scaling, and cohort size. When the number of molecular markers greatly exceeds the number of samples (as is common in ICU datasets) the risk of overfitting increases, making stability and interpretability more important than methodological novelty.

Fig. 3.

From Multi-omic Integration To Network Biology And Precision Drug Repurposing In Molecular Critical Care. Panel A depicts complementary multi-omic integration strategies, contrasting parallel layer-specific analyses with integrated approaches that model shared latent structure across transcriptomic, proteomic, and metabolomic datasets using joint factor or network-based methods; early integration combines raw features across layers, whereas late integration first compresses each modality into higher-order representations such as pathway scores or co-expression modules to improve robustness and interpretability in modest ICU cohorts. Panel B shows how molecular findings are contextualized within biological systems through protein-protein interaction networks and data-driven co-expression modules, with pathway enrichment methods collapsing long lists of differentially expressed molecules into coherent dysregulated biological programs that enable per-patient pathway activity profiling and mechanistically grounded subphenotyping. Panel C illustrates pathway-centric drug repurposing, in which subphenotype-specific dysregulated pathways are mapped onto integrated drug-target and drug-pathway knowledge networks, allowing candidate therapies to be prioritized based on their inferred ability to modulate coordinated disease-relevant programs while accounting for safety, pharmacokinetics, and feasibility in critical illness, thereby linking omics-derived biology to predictive enrichment and precision trial design in the Molecular ICU

Integration can occur at different stages. Early integration merges raw features across layers so that models learn cross-layer structure directly [90]. Late integration first reduces each omic layer into higher-order biological representations (such as pathway scores, co-expression modules, or network eigengenes) before combining them [91]. In critical care datasets, pathway- or module-level integration is often more robust, reduces dimensionality and platform-specific noise, and produces biologically interpretable units suitable for subphenotyping, risk stratification, and therapeutic targeting.

Operationalizing multi-omics in critical illness also requires aligning sampling with disease tempo. Standardized early sampling, anchored to key clinical events (e.g., ICU admission or shock onset), combined with serial sampling, allows capture of biological transitions during deterioration and recovery. High-dimensional profiling should function as a discovery engine that identifies dysregulated pathways and coordinated molecular programs. The clinically actionable output, however, should be a parsimonious pathway score or biomarker panel that represents the dominant biology of the patient. Multi-omics supports precision critical care not by bringing thousands of molecular measurements to the bedside, but by identifying the smallest biologically coherent representation of host-response state necessary to guide prognostication, predictive enrichment, and mechanism-aligned therapy.

Several computational frameworks have been developed to integrate transcriptomic, proteomic, and metabolomic datasets within unified models. Rather than analyzing each modality in isolation, these approaches identify shared biological structure that explains coordinated shifts across molecular layers. Methods such as MOFA+ [111], DIABLO [92], and iCluster [93] can uncover cross-layer patterns and define mechanistically grounded patient strata. While promising, their application in critical care remains early, and translation to practice requires further reduction of high-dimensional signatures into stable, clinically deployable biomarker panels.

Making subphenotypes actionable in critical care

Subphenotyping in critical illness requires analytical frameworks capable of identifying coordinated biological programs across high-dimensional molecular data rather than isolated biomarker-outcome associations. Regardless of integration strategy, the ultimate goal is not maximal molecular resolution but reduction to stable, interpretable pathway-focused biomarker panels that can operationalize biology at the bedside. High-dimensional omics platforms generate thousands of transcripts, proteins, and metabolites per patient, far exceeding the scope of conventional statistical analyses. While traditional statistical models remain essential and, in many cases, sufficient (particularly when the goal is interpretable classification using a known set of biologically meaningful predictors) [112], machine learning approaches are well suited to omics-driven subphenotyping when the dimensionality of the data exceeds what conventional methods can accommodate, enabling simultaneous interrogation of thousands of molecular features, capturing non-linear relationships, and revealing multivariate patterns that reflect underlying pathobiology [22, 113, 114].

Crucially, however, high-dimensional subphenotypes obtained with machine learning are not yet themselves clinically deployable. Subphenotypes defined by hundreds or thousands of molecular features, while biologically informative, are impractical for bedside implementation due to cost, assay complexity, and turnaround time. To make this useful in real-world critical care, the challenge is not collecting more data but simplifying it, and reducing complex molecular signals into a small, reliable group of biomarkers that remain accurate and practical for clinical use. This is enabled by combining machine-learning approaches with feature-selection strategies to derive parsimonious biomarker panels that preserve discriminatory power, of which a variety of computational methodologies exist (filter, wrapper and embedded reduction) [115]. Feature selection methods explicitly prioritize stability, redundancy reduction, and biological coherence, identifying small sets of biomarkers that work together to define a patient subgroup rather than maximizing classification accuracy alone [115]. Across these approaches, the goal is to reduce thousands of molecular features into a compact, mechanistically meaningful, and clinically feasible biosignature.

Feature selection addresses a core challenge in critical care omics: ICU syndromes produce thousands of measurable molecular changes, most of which are redundant or weakly informative. These methods identify the few biomarkers that consistently carry the dominant biological information needed to distinguish subphenotypes or predict outcomes, reducing dimensionality without discarding underlying mechanisms.

Molecular networks: from molecules to systems

Critical illness is not driven by isolated molecular abnormalities but by dysregulation of interacting biological systems. By measuring multiple omic layers together, researchers can now move beyond single biomarkers to identify coordinated cellular pathways that drive disease biology. Network analysis and pathway-based approaches shift biomarker discovery from single analytes toward integrated molecular programs, revealing how immune, endothelial, metabolic, and organ-specific processes act in concert to shape patient trajectories (Fig. 3B). In the ICU, where heterogeneous syndromes reflect overlapping and evolving pathobiology, these approaches provide a more faithful and clinically relevant representation of disease biology than individual markers alone.

Network-based analyses reinforce pathway-focused biomarker development by confirming that selected markers belong to coherent biological systems rather than representing isolated statistical associations. Protein-protein interaction analyses, most commonly using the STRING database [116], overlay differentially expressed proteins onto curated interaction networks to determine whether they cluster within known signaling pathways, complexes, or central hubs, thereby revealing how dysregulated markers relate to one another rather than acting in isolation. In contrast, Weighted Gene Co-Expression Network Analysis (WGCNA) [117] constructs networks directly from patient data, identifying modules of proteins that co-vary across individuals and reflect shared regulatory programs that may extend beyond predefined pathways. These data-driven modules can be summarized into eigengenes and linked to clinical outcomes, enabling molecular network activity to be related to patient trajectories. Together, these approaches move beyond identifying which biomarkers are altered to clarifying how they interact, which molecular networks and functional classes they belong to, and how coordinated pathway-level programs underpin biological heterogeneity. In doing so, it helps determine whether a set of biomarkers reflects a coherent biological mechanism, for example a complement module, an endothelial junction cluster, or a signaling axis anchored around NF-κB or JAK-STAT activation [118]. For example, network analysis of a sepsis cohort may identify a tightly co-regulated module containing complement C3, factor B, C5a receptor, and several endothelial adhesion molecules. This module doesn’t just correlate with severity; it defines a patient subgroup whose biology is dominated by complement-mediated endothelial injury. For that subgroup, complement-directed therapy has a mechanistic rationale that no single biomarker could illustrate.

Characterizing dysregulated pathways: the language of disease

Omics studies in critical illness routinely identify hundreds of differentially expressed molecules. But a list of 300 dysregulated proteins is not a molecular diagnosis: it is noise until it is organized into biology. Pathway enrichment analysis is the method that performs this translation (Fig. 3B). It is the fundamental methodology that generates ‘pathway-focused biomarkers’. It asks whether groups of biologically related molecules change together in a non-random, coordinated fashion, and in doing so converts raw molecular data into statements about which cellular programs are activated, suppressed, or failing in a given patient [100, 101]. This is the step that transforms omics from a data-generating exercise into a clinical tool.

The principle is that disease pathobiology operates through coordinated signaling networks, not through isolated molecular perturbations [119]. A modest, simultaneous shift across twenty markers of a shared signaling pathway carries far more biological meaning than a dramatic spike in any single marker. Pathway enrichment captures these coordinated shifts. It is what allows the field to move from “this patient has elevated IL-6” to “this patient has coordinated activation of NF-κB signaling with complement consumption and neutrophil degranulation”; a statement with mechanistic depth and therapeutic implications that a single biomarker measurement cannot provide [120].

All enrichment approaches depend on annotation databases that map genes and proteins to the pathways they participate in. Gene Ontology (GO) [121] organizes genes into categories describing biological processes, molecular functions, and cellular components, while Reactome [122] offers curated descriptions of human signaling pathways. These databases function as the biological dictionary that converts high-dimensional data into coherent pathway-level insights [120]. Several families of enrichment methods exist, each suited to different analytical questions. Over-representation analysis (ORA) tests whether predefined pathways contain more differentially expressed markers than expected by chance, offering simplicity but limited sensitivity to distributed, modest changes [123]. Gene Set Enrichment Analysis (GSEA) evaluates whether pathway members shift collectively across a full ranked dataset, detecting coordinated but subtle dysregulation that ORA would miss (exactly the kind of signal common in heterogeneous critical care syndromes) [124]. Modular enrichment analysis (MEA) identifies groups of co-varying molecules first and then determines what pathways they represent, making it particularly suited to multi-omic datasets [125]. Topology-based analyses incorporate known pathway structure and directionality to determine not only whether a pathway is dysregulated, but how it is functionally perturbed (such as whether it is activated or inhibited) and to identify key upstream regulators or driver nodes within the network [126]. The most clinically consequential advance, however, may be single-sample enrichment methods (ssGSEA, GSVA, PLAGE), which score pathway activity in individual patients rather than across cohorts [127, 128]. This is what makes precision enrollment and bedside decision-making possible: a pathway activity score for a single patient at a single time point, quantifying whether that patient’s complement system, interferon program, or glycolytic machinery is activated or suppressed right now. Of note, metabolomics requires its own enrichment methodologies (tools such as MetaboAnalyst [129], MSEA [130], and mummichog [131] map metabolite features onto pathway databases, though these approaches face distinct challenges including annotation gaps, isomer ambiguity, and uneven pathway database coverage. Moreover, annotation-based enrichment is inherently biased toward heavily studied biology, since databases such as GO, Reactome, and KEGG are derived from the published literature [132, 133]. Cancer-related pathways are therefore often overrepresented even in non-oncologic datasets, while processes specific to critical illness may be incompletely annotated. Enrichment findings should be interpreted within disease context and, where possible, validated across orthogonal databases.

Despite their transformative impact, pathway enrichment methods have been comparatively underused in critical illness. Most omics studies report individual biomarkers rather than the coordinated programs that more accurately reflect pathobiology. This is a missed opportunity. For example, consider two patients who present with septic shock and similarly elevated IL-6 levels. Pathway analysis reveals that in Patient A, the elevated IL-6 sits within a coordinated NF-kB inflammatory program with complement activation and neutrophil degranulation. In Patient B, the same IL-6 elevation occurs alongside suppressed interferon signaling and upregulated kynurenine metabolism, a pattern consistent with immune exhaustion rather than hyperinflammation. The biomarker is the same. The biology is opposite. The treatment should be different, and pathway enrichment helps the intensivist elucidate this. Once dysregulated pathways are identified and quantified, they become direct therapeutic targets.

Drug repurposing as a precision strategy

Once omics analyses have identified dysregulated molecular pathways underlying critical illness, the central challenge becomes therapeutic: can these pathways be meaningfully modulated within clinically relevant timeframes? In critical care, where disease trajectories evolve over hours to days and therapeutic windows are narrow, traditional de novo drug development is poorly aligned with biological urgency. Drug repurposing, the systematic identification of existing therapeutics capable of modulating disease states beyond their original indications, represents a pragmatic and coherent strategy for advancing precision critical care (Fig. 3C) [134].

Critical illness syndromes arise from distributed perturbations across cellular pathways. As a result, target-centric repurposing strategies based solely on known drug-target interactions are insufficient [135]. In contrast, pathway-centric approaches treat entire dysregulated molecular programs as the primary matching substrate, prioritizing drugs capable of modulating entire disease-relevant pathways or networks rather than individual analytes. This program-level framing aligns naturally with subphenotype-based classification and reflects the systems-level biology underlying heterogeneity of treatment effect in the ICU.

Early signature-based repurposing frameworks established the feasibility of this approach. The Connectivity Map and subsequent LINCS L1000 platform demonstrated that compounds could be computationally prioritized based on their ability to reverse disease-associated gene-expression programs, rather than on target overlap alone [136, 137]. These perturbational reference libraries enabled comparison of disease signatures against thousands of drug-induced molecular responses, establishing pathway reversal as a viable therapeutic principle. However, critical illness presents additional challenges; disease signatures are often subtle, heterogeneous, and temporally dynamic, and relevant therapeutic effects may arise through indirect or multi-step network mechanisms not captured by simple signature inversion.

Recent advances therefore extend repurposing toward network-aware, pathway-centric inference frameworks. In these approaches, drugs, molecular targets, and dysregulated pathways are embedded within unified interaction networks that capture known and inferred biological relationships [138, 139]. For example, multi-stage AI pipelines can integrate drug-target databases (DrugBank, ChEMBL, Therapeutic Target Database, Comparative Toxicogenomics Database), pathway-level mechanisms of action, and literature-derived drug-disease associations [135, 140–142] to identify candidate drugs that may target pathways of interest. Rather than assuming direct target engagement, candidate therapies are prioritized based on their inferred capacity to modulate disease-relevant cellular programs within the broader molecular network, including indirect or downstream effects [143, 144]. Integration of curated drug-target knowledge with pathway topology enables identification of compounds whose therapeutic potential emerges only at the systems level, an especially important consideration in polygenic, multi-organ syndromes.

Computational inference alone, however, is insufficient for translational prioritization. Robust repurposing frameworks increasingly integrate large-scale literature-derived evidence, allowing candidate ranking to reflect accumulated experimental and clinical knowledge alongside inferred pathway modulation [145, 146]. Systematic mining of the biomedical literature enables extraction of drug-pathway and drug-disease relationships, providing contextual support for mechanistic hypotheses and helping distinguish biologically plausible candidates from spurious matches [145, 146]. Natural language processing (NLP) algorithms using pre-trained biomedical language models (BioBERT, PubMedBERT) can mine the literature for drug-disease and drug-pathway associations, with named entity recognition identifying drug and pathway terms and relation extraction classifying semantic relationships (“inhibits,” “activates,” “modulates”) [145, 146]. In doing so, repurposing pipelines move beyond purely data-driven associations toward evidence-weighted inference. Of note, literature-derived drug-disease associations are inherently shaped by publication bias, since positive findings and well-studied compounds are disproportionately represented in the indexed literature. NLP-derived repurposing may therefore over-weight established pathways while overlooking less-studied candidates [147]. Notably, pathway-based drug repurposing pipelines have not yet been prospectively employed in the ICU, though they exist in other domains. Future prospective studies in the ICU will help define their clinical impact.

Crucially, critical care imposes constraints that demand feasibility-aware prioritization. Therapeutics nominated for ICU application must be compatible with acute illness physiology, narrow therapeutic windows, and existing care workflows. As such, contemporary repurposing strategies can explicitly integrate biological plausibility with safety profiles, pharmacokinetic suitability in critical illness, and pragmatic considerations such as route of administration and implementation feasibility [139, 148]. Candidate ranking is therefore multi-criteria and probabilistic rather than binary, enabling prioritization based on confidence and uncertainty rather than simple inclusion or exclusion. By linking subphenotype-specific pathway programs to existing pharmacologic levers, repurposing enables identification of patient subgroups most likely to benefit from a given intervention, reducing heterogeneity of treatment effect and rescuing therapies that fail in unselected populations. In the molecular ICU, drug repurposing thus functions as a translational bridge, from high-dimensional molecular heterogeneity to biologically informed, feasible, and testable precision trials.

How trials fail, and how to fix them: the rationale for predictive enrichment

Heterogeneity of treatment effect in critical illness arises from multiple sources, including baseline risk, physiologic phenotype, etiology, and underlying host-response biology, each motivating distinct enrichment strategies [8, 149–153]. Pathway-level signatures revealed by omics analyses address the biological dimension specifically, offering a mechanistic framework for predictive enrichment that complements physiologic and trait-based phenotyping approaches. By identifying patients more likely to respond to a specific therapy based on biological or clinical markers, predictive enrichment enables more targeted and biologically coherent clinical trials [98]. Pathway-level signatures may enhance predictive enrichment by identifying patients whose active biological programs correspond to the therapeutic targets of a given intervention. Rather than relying on syndromic definitions, disease severity, or single biomarkers, pathway-based enrichment yields the opportunity to stratify patients according to coordinated molecular dysregulation (such as immune activation, endothelial injury, metabolic failure, or neuroinflammation) that plausibly mediates therapeutic response [50, 103, 154]. These programs can be operationalized as parsimonious pathway scores or biomarker panels to guide eligibility, stratified randomization, or adaptive allocation, thereby aligning treatment testing with underlying biology, reducing heterogeneity of treatment effect, and improving the efficiency and interpretability of critical care trials.

Other fields (most prominently oncology) have transformed therapeutic development by implementing biomarker-guided enrichment strategies, where molecular signatures dictate who receives which therapy and why [155, 156]. The trastuzumab/HER2 model established the template, where patients with HER2-positive breast cancer derive substantial benefit from anti-HER2 therapy, while HER2-negative patients do not [156]. By restricting treatment to the molecularly defined responsive population, oncology transformed outcomes and established regulatory pathways for biomarker-guided therapy [157].

In contrast, critical care has struggled to operationalize biomarker-driven predictive enrichment [154]. Critical illness syndromes arise from multiple dysregulated molecular pathways, evolve rapidly, and rarely hinge on a single targetable mechanism [158]. Historically, biomarker research in this field has focused on individual proteins or genetic variants. While informative, single-analyte approaches fail to capture the complex, polygenic, and dynamic biology underpinning sepsis, ARDS, and TBI. A pathway-level approach is needed that identifies which molecular programs are dysregulated across patients and uses those signatures to meaningfully stratify treatment response. Cancer diagnostics tolerate turnaround times of days to weeks while critical illness demands results within hours [84]. Oncology biomarkers often reflect stable genomic alterations, whereas sepsis and ARDS phenotypes evolve rapidly as host responses transition from hyperinflammation to immunosuppression [18]. Nevertheless, the core principle translates: identifying patients more likely to respond to specific therapies based on biological markers offers a path to rescuing interventions that fail in heterogeneous populations [98].

From discovery to deployment: translating pathway-focused biology to the bedside

A defined pipeline is needed to translate pathway-focused biomarkers from discovery to prospective trial enrollment (Fig. 4). Multi-omic integration maps disease architecture, pathway analysis identifies dysregulated programs, and machine learning feature selection reduces high-dimensional signatures to parsimonious panels [38, 159–161]. Bedside deployment ideally uses rapid point-of-care platforms that return results within timelines suitable for patient management or enrollment in clinical trials. Pathway-level measurement also supports therapeutic discovery by linking dysregulated pathways to candidate therapies and advancing mechanism-informed interventions into adaptive platform trials.

Fig. 4.

Translating pathway-focused biology from discovery toward precision critical care implementation. High-dimensional multi-omic profiling (including transcriptomics, proteomics, metabolomics, and related molecular platforms) functions as a discovery engine to characterize disease architecture and identify molecular alterations associated with critical illness. Following preprocessing and integration, computational approaches including pathway enrichment, network analyses, clustering, and machine-learning methods organize thousands of molecular measurements into coordinated biological programs that reflect underlying host-response mechanisms. Feature-selection strategies may subsequently reduce these high-dimensional signatures into parsimonious pathway-focused biomarker panels intended to preserve biological coherence while improving feasibility for future clinical translation. Such pathway-focused biomarkers could ultimately support prognostication, predictive enrichment, therapeutic stratification, and adaptive clinical trial design by aligning therapeutic interventions with dominant biological programs. Dysregulated pathways may additionally inform therapeutic prioritization and drug repurposing strategies, providing a conceptual translational bridge between molecular discovery and mechanism-aligned interventions in critical illness

Recent prospective biomarker-guided trials demonstrate that the individual components of this pipeline (real-time bedside classification, biomarker-defined enrichment, and subphenotype-matched therapy) are each clinically operational, but none have yet been integrated under a pathway-level framework. PHIND demonstrated that rapid bedside subphenotyping in acute hypoxemic respiratory failure is feasible [17]. Tigris reported a mortality benefit with polymyxin B hemoadsorption in biomarker-defined endotoxic septic shock, a subphenotype unidentifiable by clinical criteria alone [162, 163]. ImmunoSep prospectively stratified sepsis into macrophage activation-like syndrome and immunoparalysis subphenotypes demonstrating that subphenotype-matched immunotherapy improves organ dysfunction [164]. Computational refinement of enrichment criteria is also yielding actionable signals: a 3-biomarker model identified a hydrocortisone survival benefit in CAPE COD limited to patients with severe immune dysregulation [165, 166], and sTREM-1 enrichment in ASTONISH identified a subgroup with nangibotide benefit that has informed phase 3 design [167].

These trials show that real-time bedside classification, biomarker-based enrichment, and subphenotype-matched therapy are each feasible, but not yet integrated into a pathway-level framework. Current enrichment remains simple, typically relying on one or a few analytes with threshold-based scoring, and no trial has enrolled patients using integrated multi-omic pathway signatures or tested pathway-defined treatment effects. This is not a weakness of the framework, but its central opportunity: the tools for pathway-focused predictive enrichment now exist, and the next generation of critical care trials must move from individual biomarkers to coordinated pathway-level biology.

Recent multi-omic studies illustrate the promise of this pathway-focused strategy. Integrated metabolomic and transcriptomic analyses show that even within the ‘hyperinflammatory’ ARDS subphenotype, distinct inflammatory programs exist, including (1) heightened innate immune activation with metabolic reprogramming, (2) hepatic dysfunction with impaired β-oxidation, and (3) suppression of interferon signaling with mitochondrial derangements [92, 168]. Each represents a plausible biologically distinct population, and possible subgroup for trial enrichment. Similarly, combined transcriptomic and proteomic profiling in ARDS has revealed coordinated dysregulation of innate immunity, tissue remodeling, zinc homeostasis, collagen synthesis, and neutrophil degranulation within patients broadly labeled as ‘hyperinflammatory’ [36, 92]. These insights highlight why syndromic definitions alone are insufficient, and how pathway-based subphenotyping may enable more meaningful, mechanism-aligned predictive enrichment. If trials are to meaningfully target biology rather than syndromes, enrichment strategies must begin to incorporate pathway-level signatures, moving beyond single biomarkers toward mechanistic profiles that better reflect the true drivers of critical illness.

An important tension underlies the use of pathway-focused biomarkers for predictive enrichment: critical illness biology is inherently dynamic, with host-response programs shifting over hours to days, yet predictive enrichment typically relies on pathway-focused biomarker measurement at a single early time point that may not reflect a patient’s dominant biology even 24–48 h later [10, 169]. Reconciling this will likely require strategies that embed biological dynamism into trial structure, including (1) longitudinal biospecimen collection to characterize pathway trajectories rather than static states, (2) dynamic predictive enrichment with re-assessment of pathway activity at predefined intervals, and (3) response-adaptive randomization within Bayesian platform trials that updates allocation probabilities as pathway-level and clinical response data accumulate [170, 171]. Whether single-timepoint, longitudinal, or response-adaptive pathway-focused enrichment proves optimal will likely vary by intervention, target pathway, and disease tempo, and remains an open methodological question for the next generation of biology-informed critical care trials.

Moreover, implementing longitudinal pathway-based enrichment introduces two practical challenges worth acknowledging. First, patients arrive at the ICU at very different points in their illness (some within hours of symptom onset, others days into a deteriorating course on the ward) so the same calendar time does not represent the same biological time. Anchoring sampling to a defined clinical event (such as shock onset, intubation, or initiation of vasopressors) rather than to ICU admission helps align patients to a common biological reference point, effectively redefining time zero from an administrative timestamp to a biologically meaningful inflection point, though the optimal anchor will vary by syndrome and intervention. Second, patients who die early are precisely those whose biological trajectories are most extreme, meaning loss to follow-up is informative rather than random, and trajectories built only from survivors will systematically misrepresent the biology of those who do not survive. Analytic approaches that jointly model biomarker trajectories with survival (estimating both the trajectory and the risk of death within a single statistical framework so each informs the other) [172, 173] or that incorporate death directly into the outcome (for example, that include death as a terminal state) [174] will be important tools as the field moves toward longitudinal pathway-focused enrichment.

Operationalizing omics for trial design in the ICU

Post hoc molecular stratification has revealed treatment effects are often masked in conventional intention-to-treat analyses, providing a compelling biological rationale for prospective predictive enrichment strategies. For example, secondary transcriptomic analysis of the VANISH trial demonstrated increased mortality with hydrocortisone among patients with the immunocompetent SRS2 subphenotype despite no overall treatment benefit in septic shock [13]. Similarly, retrospective analyses of the ARMA, ALVEOLI, and FACTT trials identified ARDS subphenotypes with differential responses to PEEP and fluid management strategies [14, 94], while latent class analysis of HARP-2 demonstrated improved 28-day survival with simvastatin in the hyperinflammatory subphenotype despite a neutral overall trial result [95, 175]. Collectively, these studies demonstrate that molecular stratification can uncover substantial heterogeneity of treatment effect within critical illness syndromes, revealing therapeutic benefit in biologically defined subgroups despite negative population-level trial results.

The field must now decisively move toward prospective validation of subphenotype-guided enrichment, increasingly leveraging adaptive platform trial designs that operationalize precision therapeutic targeting in real time. Adaptive platform trials are perpetual, flexible studies that test multiple treatments simultaneously and adapt allocation and eligibility over time based on accumulating evidence, typically using Bayesian frameworks with pre-specified stopping rules so arms can be added or dropped for efficacy or futility [176]. By incorporating molecular stratification, these trials can determine whether specific phenotypes preferentially benefit from particular treatments while also assessing overall efficacy. Table 3 summarizes major adaptive platform trials in critical care, including their target populations, design rationale, use of predictive or biomarker-based enrichment, and relevance to precision medicine in the ICU.

Table 3.

Adaptive platform trials in adult critical care medicine

Platform trial	Population	Why trial was designed	Predictive/biomarker enrichment used	Relevance for precision medicine	Geographic network	Molecular biology integration
REMAP-CAP [170, 176]	Severe community-acquired pneumonia requiring ICU care (including pandemic respiratory infections)	Efficiently identify which therapies improve survival and organ support-free days in severe pneumonia by continuously testing and updating treatment effects within a standing ICU population	Primarily pragmatic; stratification varies by domain and is not uniformly biomarker- or subphenotype-driven at enrollment	Demonstrates how a permanent adaptive infrastructure can rapidly generate high-quality evidence and subsequently incorporate biological stratification	Large international critical care network	Embedded biology infrastructure Pragmatic enrollment; protocolized biospecimen collection; pre-specified secondary biological analyses; not molecularly stratified at enrollment
PRACTICAL [176]	Acute hypoxemic respiratory failure (AHRF) in hospital/ICU	To determine which respiratory and supportive care strategies improve outcomes across the spectrum of acute hypoxemic respiratory failure using a seamless trial pipeline	Not inherently subphenotype-stratified; infrastructure supports embedded biology and downstream stratified analyses	Provides a bridge from pragmatic syndrome-based trials to future biologically enriched trials	Canadian-led, multi-centre network	Future biology-ready Platform supports addition of biological substudies; current enrollment not molecularly stratified
PANTHER [176, 219]	ARDS and hypoxemic acute respiratory failure	To test whether treatment effects differ by biological subphenotype, explicitly measuring heterogeneity of treatment effect across inflammatory ARDS subphenotypes	Yes; real-time assignment to hyperinflammatory vs. hypoinflammatory subphenotypes using rapid biomarkers and clinical data	Represents a direct implementation of predictive enrichment based on molecular subphenotyping in critical care	European / UK-led collaborative network	Molecular stratification from inception Pre-specified real-time assignment to hyperinflammatory vs. hypoinflammatory subphenotype; molecular enrichment criterion built into enrollment
TRAITS [176]	Adult acute critical illness (syndrome-agnostic ICU population)	To test whether targeting specific, measurable biological traits (rather than syndromic diagnoses) improves patient-centered outcomes	Yes; enrollment guided by trait definitions using bedside biomarkers or physiological surrogates	Advances a biology-first, cross-syndrome precision ICU paradigm focused on treatable mechanisms	UK / Scotland ICU research network	Molecular stratification from inception Enrollment guided by pre-specified biological trait definitions using bedside biomarkers; biology-first design from inception
INCEPT [176, 220]	Broad spectrum adult ICU admissions	To efficiently evaluate common ICU interventions (fluids, thromboprophylaxis, albumin) and determine their impact on survival and organ support outcomes	Primarily intervention-driven; not subphenotype-led by default	Establishes a scalable experimental backbone onto which biomarker- or subphenotype-guided stratification can later be layered	Denmark-led with international expansion	Future biology-ready Scalable experimental backbone; no current molecular stratification; infrastructure designed to support future subphenotype-guided substudies
TIGERS (tigerstrial.org)	Adults with sepsis	To accelerate identification of effective sepsis therapeutics through matching candidate immunomodulatory interventions to host-response subphenotypes (SRS classification)	Yes; enrollment and treatment allocation guided by prospective transcriptomic SRS subphenotype assignment using a rapid bedside classifier	First international platform trial designed around prospective transcriptomic subphenotype-guided predictive enrichment in sepsis, with multiple subphenotype-matched interventions	UK-Australia international collaborative platform	Molecular stratification from inception SRS subphenotype assignment at enrollment; subphenotype-matched interventions tested within an adaptive Bayesian design; embedded biospecimen collection

Collectively, these platforms address heterogeneity of treatment effect through complementary precision medicine strategies and are inherently compatible with multi-omic integration. Embedded biobanking and longitudinal biospecimen collection enable simultaneous evaluation of therapeutic efficacy and underlying pathobiology, while adaptive interim analyses iteratively refine biomarker-guided stratification in real time. Shared controls and multi-arm designs improve efficiency and power to detect biologically defined treatment effects, establishing a framework in which molecular discovery and therapeutic testing evolve in parallel to accelerate precision therapeutics in critical illness.

Challenges, limitations, implementation barriers and ethical considerations

Despite the promise of omics-driven subphenotyping and biomarker development in critical illness, several technical, biological, operational, and ethical challenges continue to limit translation from discovery to routine clinical and trial use. Importantly, these barriers are increasingly well defined and, with deliberate methodological, infrastructural, and governance solutions, are largely addressable rather than insurmountable. Nevertheless, they must be acknowledged.

Biological and technical limitations

A central challenge is the dynamic, time-dependent nature of critical illness biology. Host responses evolve rapidly over hours to days, transitioning among hyperinflammation, immunosuppression, metabolic collapse, and recovery, so a single blood draw may not capture the biology driving outcome or treatment response [10, 39, 177]. Subphenotypes defined early in illness may not remain stable, raising questions about temporal validity and optimal timing of molecular profiling.

High-dimensional omics data are also affected by batch effects, platform-specific bias, and limited cross-cohort reproducibility. Differences in biospecimen handling, assay chemistry, normalization, and analytical choices can produce discordant subphenotypes across studies, particularly in sepsis, where transcriptomic frameworks show only partial overlap [178, 179]. Batch effects can reduce statistical power and generate spurious associations when batches correlate with biological or clinical outcomes [180, 181]. Errors in experimental design, small sample sizes, inappropriate controls, lack of reference standards, and non-transparent reporting further threaten independent validation. Indeed, bedside implementation of pathway-focused biomarkers will require addressing a substantial set of practical barriers, including pre-analytic standardization (sample type, timing, handling, storage), inter-laboratory and platform concordance, device calibration and quality control, regulatory authorization and reimbursement, workflow and electronic health record integration, management of classification uncertainty, and cost-effectiveness [10, 20–22].

Analytical and methodological challenges

Machine learning can uncover complex molecular patterns but risks overfitting, limited interpretability, and poor generalizability, especially in modest ICU cohorts [182]. Many reported subphenotypes rely on unsupervised clustering without ground truth and are sensitive to modeling assumptions, distance metrics, and feature selection. Without rigorous external validation and transparent reporting, statistically derived clusters may be mistaken for biologically meaningful entities. Moreover, machine learning is not always optimal for clinical deployment. Simpler approaches such as regularized regression and biologically informed composite scores often perform comparably in ICU cohorts while offering greater interpretability and generalizability [112]. Machine learning may therefore serve best as a discovery and dimensionality reduction tool, with deployable pathway-focused biomarker panels favouring simpler, more interpretable scoring approaches. Pathway enrichment and network analyses also depend on curated knowledgebases that incompletely represent context-specific signaling in acute human illness and bias interpretation toward well-annotated pathways [15].

Implementation barriers in the ICU

Operational barriers include turnaround time, cost, and workflow integration. Critical care decisions often must be made within hours, whereas most omics platforms require centralized processing and batch analysis. Point-of-care assays and parsimonious biomarker panels are emerging but remain uncommon. Advances in rapid metabolic profiling, mass spectrometry, microfluidics, and portable biosensors may enable personalized medicine in critical care [30, 183]. ICU populations are also heterogeneous, unstable, and underrepresented in discovery cohorts, raising concerns about equity, generalizability, and access.

Ethical and regulatory considerations

Subphenotype-guided therapy raises ethical concerns. Assigning or withholding treatments based on probabilistic molecular classifications introduces risks of misclassification, algorithmic bias, and challenges for informed consent. Many omics datasets overrepresent high-income settings, specific socioeconomic groups, and individuals of European ancestry, limiting inclusion of communities with distinct biology and exposures [184]. This imbalance constrains biomarker applicability and may worsen inequities if tools validated in one population are applied to another without recalibration [185, 186]. Stratification criteria can shape access to treatment and, if correlated with socioeconomic status, residence, or ethnicity, may lead to unjust distribution of care [187]. Machine-learning-derived biomarkers further challenge transparency and explainability in high-stakes decisions. Finally, multi-omic data collection and integration raise issues of ownership, privacy, secondary use, and governance, especially as biobanking becomes embedded in adaptive platform trials.

Moving forward: what the molecular ICU means for intensivists

For intensivists, the central message of the Molecular ICU is that syndromic diagnoses are biologically insufficient. Patients who appear clinically similar often occupy fundamentally different molecular states, and outcomes and treatment responses are shaped more by underlying pathways than by bedside labels or isolated biomarkers. Single-analyte markers should therefore be interpreted as partial reflections of broader biological programs, and dynamic changes in immune, endothelial, metabolic, and neuroinflammatory pathways should be expected to evolve over time. For trialists, these insights demand a shift away from syndromic enrollment toward biologically informed predictive enrichment, with extraordinarily robust biospecimen collection, pathway-level biomarkers, and molecular stratification embedded prospectively into trial design. High-dimensional omics should be used to define biology, but operationalized through parsimonious biomarker panels aligned with the intervention’s mechanism, ideally within adaptive platform trials capable of testing heterogeneity of treatment effect. The central argument of this Review is that pathway-focused biomarkers represent the correct translational unit for precision critical care. They preserve mechanistic coherence, reduce dimensionality, enable bedside feasibility, and align directly with predictive enrichment and drug repurposing. Without pathway-level organization, omics remains descriptive; with it, molecular biology becomes operational.

Conclusion

Towards the molecular era of critical care medicine

Precision critical care requires pathway-level thinking and multi-omic integration. The complex, dynamic biology of critical illness resists reduction to single biomarkers, and the repeated failure of syndrome-based trials confirms that current approaches are insufficient. Three priorities are immediate. First, prospective biospecimen collection must become standard in critical care trials, enabling molecular stratification alongside therapeutic efficacy testing. Second, pathway-focused biomarker identification must be validated across diverse populations and clinical settings, with attention to turnaround time, assay standardization, and equity of access. Third, adaptive trial platforms must embed molecular subphenotyping into enrollment and allocation algorithms, closing the loop between biological discovery and therapeutic testing. These steps are not speculative. Post hoc molecular analyses of completed trials have already demonstrated that heterogeneity of treatment effect is detectable, biologically coherent, and clinically meaningful. The task now is to move this evidence from retrospective proof-of-concept to prospective implementation. Continuing to design trials and deliver therapies as though syndromic labels represent biological uniformity ensures that effective treatments remain undetected and patients receive interventions unlikely to benefit them. The era of the molecular ICU has arrived; what the field requires is the commitment to build it.

Abbreviations

ARDS

Acute respiratory distress syndrome

AUROC

Area under the receiver operating characteristic curve

CALC-1

Calcitonin-1 gene

CC16

Club cell secretory protein

CRP

C-reactive protein

CTS

Consensus transcriptomic subtypes

DIA

Data-independent acquisition

DNA

Deoxyribonucleic acid

ELISA

Enzyme-linked immunosorbent assay

FDA

U.S. Food and Drug Administration

GFAP

Glial fibrillary acidic protein

GO

Gene Ontology

GSVA

Gene set variation analysis

GWAS

Genome-wide association study

HLA

Human leukocyte antigen

HTE

Heterogeneity of treatment effect

ICU

Intensive care unit

IL

Interleukin

LC-MS/MS

Liquid chromatography–tandem mass spectrometry

LASSO

Least absolute shrinkage and selection operator

MCP1

Monocyte chemoattractant protein-1

ML

Machine learning

MOFA+

Multi-omics factor analysis plus

MS

Mass spectrometry

NGS

Next-generation sequencing

NMR

Nuclear magnetic resonance

NSE

Neuron-specific enolase

NULISA

Nucleic acid–linked immuno-sandwich assay

ORA

Over-representation analysis

PAI-1

Plasminogen activator inhibitor-1

PCA

Principal component analysis

PCT

Procalcitonin

PEA

Proximity extension assay

PEEP

Positive end-expiratory pressure

PPI

Protein–protein interaction

RNA

Ribonucleic acid

RNA-seq

RNA sequencing

RCT

Randomized controlled trial

scRNA-seq

Single-cell RNA sequencing

SOFA

Sequential Organ Failure Assessment

SP-D

Surfactant protein D

ssGSEA

Single-sample gene set enrichment analysis

suPAR

Soluble urokinase plasminogen activator receptor

TBI

Traumatic brain injury

TNF

Tumor necrosis factor

UCH-L1

Ubiquitin C-terminal hydrolase-L1

UMAP

Uniform manifold approximation and projection

vWF

Von Willebrand factor

WGCNA

Weighted gene co-expression network analysis

Author contributions

LRVN and DDF conceptualized the manuscript, led the writing, and developed the overall framework. HR contributed to drafting, figure illustration and critical revision of the manuscript. JB and MS provided intellectual input and revised the manuscript for important scientific content. DDF supervised the work, contributed to conceptual development, and critically revised the manuscript. All authors approved the final version of the manuscript.

Funding

Not applicable.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.