Integrative Biomedical Research

1. Introduction

Sepsis is, by the Third International Consensus Definitions (Sepsis-3), a life-threatening organ dysfunction caused by a dysregulated host response to infection (Singer et al., 2016). The definition is short. What it describes is not. Recent epidemiological estimates place the global burden somewhere between 48.9 and 166 million cases each year, with 11.0 to 21.4 million deaths attributed to sepsis (Fleischmann-Struzek et al., 2020; Lee et al., 2026), and the width of those ranges is itself telling: we still do not count sepsis very precisely. Once the syndrome progresses to septic shock and multiple organ dysfunction syndrome (MODS), intensive care unit (ICU) mortality commonly exceeds 40% (Gonzalez-Barbuzano et al., 2026).

For decades, severity has been judged with bedside scores: the Sequential Organ Failure Assessment (SOFA) and its quick version (qSOFA), the Acute Physiology and Chronic Health Evaluation (APACHE II), and, in children, the Pediatric Risk of Mortality (PRISM III) and Pediatric Logistic Organ Dysfunction-2 (PELOD-2) scores (Gonzalez-Barbuzano et al., 2026; Gürsoy Koca et al., 2026; Lee et al., 2026). These instruments are useful, and nobody seriously proposes abandoning them. But they register physiological derangement after it has happened; they say little about the biological heterogeneity, the timing, or the molecular endotypes that sit underneath (Gonzalez-Barbuzano et al., 2026). Two patients can share the same SOFA score and carry strikingly different immune profiles, and, perhaps unsurprisingly, strikingly different outcomes. That observation, more than any single discovery, is what has pushed the field toward biology-driven subtyping, prognosis and drug discovery.

Our picture of sepsis biology has shifted accordingly. The older model, centred on an uncontrolled "cytokine storm", has given way to something messier: a system-wide disorder in which hyperinflammation and immune paralysis can coexist, alongside endothelial injury, microvascular thrombosis and metabolic collapse (Ghazal et al., 2022; White et al., 2023). Using mathematical reasoning, Ghazal et al. (2022) proposed three cardinal rules for host-response trajectories. First, studying the immune-inflammatory response in isolation is not enough, because interlocking physiological, metabolic and neuroendocrine systems shape outcome. Second, cellular bioenergetic demand acts as a central force on host reactivity. Third, shifts in homeostatic set-points appear to steer both disease course and recovery potential. At the cellular and subcellular level, mitochondrial energy failure, loss of oxidative phosphorylation, excess reactive oxygen species (ROS) and dysregulated cell death, including apoptosis, pyroptosis, necroptosis and ferroptosis, seem to be shared drivers of tissue breakdown across organs (Z. Han et al., 2024; L. Zhang et al., 2026).

How, then, does one study a disorder that operates on so many levels at once? High-throughput omics technologies, spanning genomics, epigenomics, transcriptomics, proteomics, phosphoproteomics, metabolomics, lipidomics and metagenomics, have become the default toolkit (Gonzalez-Barbuzano et al., 2026; Z. Han et al., 2024; Hudu et al., 2026). Single-layer studies have been productive, identifying candidate biomarkers and disrupted pathways. Yet looking at one molecular scale at a time tends to hide multi-target interactions, feedback loops and temporal cascades (P. Li et al., 2026; Na et al., 2023). A good illustration comes from time-dependent multi-omics integration (TDMI), which combined phosphoproteomics at 4 h, transcriptomics at 6 h and proteomics and metabolomics at 18 h, and showed that both the TLR4/MyD88/NF-κB and the TRIF/IRF3 arms of Toll-like receptor 4 signalling drive early sepsis-associated liver dysfunction (SALD); neither snapshot alone had revealed this (Na et al., 2023). In a similar spirit, integrative profiling of sepsis-induced liver injury (SILI) suggested that salvianolic acid A protects the liver by coordinating amino acid remodelling, glutathione-dependent redox balance and anti-apoptotic signalling (P. Li et al., 2026).

Clinical systems immunology tells a comparable story. In adults with sepsis, regression models that combined immature CD10⁻ neutrophils, vascular permeability mediators (pentraxin-3, angiopoietin-2, endocan) and altered lysophosphatidylcholines and fatty acids predicted severity and organ failure reasonably well (White et al., 2023). Multi-marker strategies layered onto physiological scores, using procalcitonin (PCT), presepsin (PSEP), interferon-λ3 (IFN-λ3) and bioactive adrenomedullin (bio-ADM), improved diagnostic discrimination and roughly tripled the hazard captured for in-hospital death compared with SOFA alone, particularly in older patients (Lee et al., 2026).

Organ-specific vulnerability adds another dimension, and it is here that a systems view becomes almost unavoidable. In the heart, sepsis-induced cardiomyopathy (SIC) involves upregulated pyruvate dehydrogenase kinase 4 (PDK4), suppressed tricarboxylic acid (TCA) cycle flux and ferroptosis driven by VDAC2 malonylation (Z. Han et al., 2024). The liver is both a target and an amplifier of systemic inflammation; in critically ill children, secondary hepatic dysfunction is associated with longer ICU stays and greater morbidity (Gürsoy Koca et al., 2026). In the brain, widely targeted metabolomics combined with gradient-boosted classifiers (XGBoost) has pointed to disturbed glycerophospholipid and ether-lipid metabolism, and to succinate, as hallmarks of sepsis-associated encephalopathy (SAE) (Baizhanova et al., 2025; Hu et al., 2026). And artificial intelligence (AI) and machine learning (ML) applied to microbiome–metabolome data have implicated gut dysbiosis, short-chain fatty acid (SCFA) depletion and translocation of microbial products in systemic immune activation (Hudu et al., 2026).

None of this has yet changed routine care. Translation is held back by batch effects, high dimensionality, heterogeneous cohorts, a lack of standardised workflows and thin external validation (Gonzalez-Barbuzano et al., 2026; Hudu et al., 2026; Na et al., 2023). We think the most useful thing a review can do at this point is to line the evidence up across scales, from subcellular bioenergetics and immune-cell subsets to organ-specific injury and whole-patient multi-marker models, and to ask honestly where it holds together and where it does not.

With that in mind, the review had four aims. First, to synthesise the genomic, epigenomic, transcriptomic, proteomic, phosphoproteomic, metabolomic, lipidomic and metagenomic perturbations that characterise the dysregulated host response in sepsis. Second, to examine the subcellular, metabolic and cellular mechanisms of failure in individual organs, namely SILI, SIC, sepsis-associated acute kidney injury (SA-AKI), SAE and acute lung injury, and the crosstalk between them. Third, to appraise integrative computational approaches, including TDMI, network-based systems immunology and machine-learning algorithms such as random forest, XGBoost and deep neural networks, for endotyping, risk stratification and biomarker discovery. Fourth, to weigh the clinical utility and translational barriers of multi-marker and multi-omics approaches on the road toward endotype-guided therapy in critical care.

2. Multi-Omics Integration for Understanding Sepsis Pathophysiology and Organ-Specific Injury

2.1 Rethinking sepsis pathophysiology: beyond a unimodal inflammatory model

2.1.1 Burden of disease and the limits of physiological scoring

Sepsis continues to place an extraordinary load on intensive care services worldwide. Epidemiological work suggests tens of millions of hospital admissions and, by some estimates, close to 20 million deaths each year (Fleischmann-Struzek et al., 2020; Lee et al., 2026), and once shock and MODS set in, mortality above 40% is common (Gonzalez-Barbuzano et al., 2026). The Sepsis-3 definition, with its emphasis on organ dysfunction arising from a dysregulated host response (Singer et al., 2016), quietly acknowledges that it is the host, rather than the pathogen alone, that ultimately determines the course.

Clinical severity has long been measured with SOFA, qSOFA, APACHE II, PRISM III and PELOD-2 (Gonzalez-Barbuzano et al., 2026; Gürsoy Koca et al., 2026; Lee et al., 2026). These scores give a valuable physiological snapshot. Their limitation is structural, though, not incidental: they record downstream systemic dysfunction instead of the cellular and molecular heterogeneity that produces it (Gonzalez-Barbuzano et al., 2026; Gürsoy Koca et al., 2026). The practical consequence is familiar to anyone who has worked in an ICU. Two patients with the same SOFA or PRISM score may follow entirely different paths, one recovering within days and the other developing refractory multi-organ failure (Gonzalez-Barbuzano et al., 2026).

2.1.2 From cytokine storm to multi-system dysregulation

Over the past decade the conceptual centre of sepsis biology has moved. The "cytokine storm" is no longer thought sufficient to explain what happens. Instead, sepsis is now seen as a dynamic state in which hyperinflammation, immune paralysis, microvascular thrombosis, endothelial barrier failure and profound bioenergetic collapse can unfold at the same time, sometimes in the same patient on the same day (Ghazal et al., 2022; White et al., 2023). Ghazal et al. (2022), reasoning mathematically, distilled this into three cardinal rules. The host response is inherently systemic, so immunological parameters studied without their neuroendocrine and physiological feedbacks give only a partial picture. Bioenergetic capacity and cellular energy demand constrain both host reactivity and cell survival. And disease progression and recovery are ultimately governed by shifts in homeostatic set-points. These rules are abstract, admittedly, but they have a concrete implication: no single measurement layer is likely to be enough.

2.1.3 Subcellular drivers: mitochondria and regulated cell death

At the subcellular level, systemic collapse seems to be driven by mitochondrial dysfunction, failing oxidative phosphorylation, accumulating ROS and the dysregulated execution of regulated cell death (RCD) programmes, including apoptosis, necroptosis, pyroptosis and ferroptosis (Z. Han et al., 2024; L. Zhang et al., 2026). Mitochondrial injury and oxidative stress also feed immunometabolic reprogramming of immune cells, which may help explain why energy failure and immune paralysis so often travel together (H. Kim et al., 2026). Untangling these layered perturbations calls for analytical frameworks that can connect molecular events across several biological scales (Gonzalez-Barbuzano et al., 2026). Figure 1 sketches how the four scales discussed in this review relate to one another and to the omics layers used to read them (Figure 1).

2.2 Integrative multi-omics frameworks versus single-omics isolation

2.2.1 The reach and blind spots of single-layer omics

High-throughput platforms spanning genomics, epigenomics, transcriptomics, proteomics, phosphoproteomics, metabolomics, lipidomics and metagenomics have transformed systems medicine (Gonzalez-Barbuzano et al., 2026; Z. Han et al., 2024; Hudu et al., 2026). Single-layer studies have produced real insight, for instance transcriptomic signatures of sepsis endotypes and altered circulating amino acid profiles (Gonzalez-Barbuzano et al., 2026; Na et al., 2023). A consensus blood transcriptomic framework has, for example, tried to reconcile competing sepsis response signatures into a common classification (Scicluna et al., 2025). Still, any one layer offers a narrow view. It will miss post-transcriptional modification, the lag between gene expression and functional protein, enzymatic feedback and metabolite flux (Na et al., 2023; White et al., 2023). Metabolomics, arguably the layer closest to phenotype, has its own blind spots, since a metabolite concentration rarely reveals which pathway produced it (Pandey, 2025).

2.2.2 Time-dependent integration as a proof of principle

Integrative multi-omics strategies were developed partly to deal with these blind spots, and they are increasingly regarded as the preferred way to map sepsis pathophysiology. Time-dependent multi-omics integration (TDMI) is a particularly instructive case because it aligns sampling with the kinetics of each biological layer (Na et al., 2023). In a model of SALD, phosphoproteomics at 4 h, transcriptomics at 6 h and proteomics and metabolomics at 18 h were integrated, revealing early TLR4 activation and resolving its dual MyD88/NF-κB and TRIF/IRF3 arms, which had gone unnoticed when each layer was analysed alone (Na et al., 2023). The temporal logic of this design is summarised in Figure 2 (Figure 2). One caveat deserves mention: this was an experimental model, and whether the same sampling windows translate to patients, who rarely arrive at a known "time zero", remains uncertain.

2.2.3 Systems immunology and proteogenomic integration

Systems immunology offers a parallel route. By combining flow cytometry, multiplex cytokine profiling and untargeted lipidomics, White et al. (2023) linked sepsis severity to particular circulating immune-cell subpopulations. Immature CD10⁻ neutrophils, elevated vascular permeability mediators (pentraxin-3, angiopoietin-2, endocan) and shifted lysophosphatidylcholine-to-fatty acid ratios formed an interconnected signature that stratified risk and progression (White et al., 2023), in line with the discovery of discriminating neutrophil subsets in independent cohorts (Meghraoui-Kheddar et al., 2022). Integrated transcriptomic and proteomic profiling has also identified an interferon-dependent inflammatory endotype (Fratea et al., 2026). Meanwhile, proteogenomic integration, which brings genome-wide association studies (GWAS) together with plasma proteomics and expression quantitative trait loci (eQTL) mapping, has started to separate upstream causal drivers from downstream reactive biomarkers (Burnham et al., 2024; Gonzalez-Barbuzano et al., 2026; Sun et al., 2018). The main omics layers, their matrices and platforms, and their prognostic uses are compared in Table 1 (Table 1).

2.3 Organ-specific pathophysiology and subcellular vulnerabilities

Multi-organ failure in sepsis is not a uniform shutdown. It unfolds through distinct mechanisms in different organs, and systems biology has begun to map these vulnerabilities in some detail (Figure 3; Table 2).

2.3.1 Hepatic dysfunction and sepsis-induced liver injury

The liver acts both as an immunological filter and as a metabolic hub during severe infection. SILI and secondary hepatic dysfunction have been associated with mortality above 50% in some series and appear to act as independent drivers of multi-organ failure (Gürsoy Koca et al., 2026; P. Li et al., 2026; Saini et al., 2022). Multi-omics profiling points to broad metabolic reprogramming, with disrupted glutathione metabolism, impaired taurine and hypotaurine pathways and altered amino acid use (P. Li et al., 2026; Na et al., 2023). Experiments with salvianolic acid A suggest that restoring hepatic function requires several things at once: dampening STING/TBK1/IRF3 inflammatory signalling, suppressing hepatocyte apoptosis and rebuilding redox homeostasis (P. Li et al., 2026). In paediatric intensive care, secondary hepatic dysfunction occurs in nearly 9% of critically ill children, overwhelmingly

Figure 1. Sepsis as a multi-scale systems failure: biological scales of host dysregulation and the omics layers used to interrogate them. The left column arranges the septic host response into four nested scales, from subcellular events (mitochondrial failure, loss of oxidative phosphorylation, reactive oxygen species and regulated cell death) through cellular and organ-level injury to the systemic dysfunction captured by bedside scores; double-headed arrows indicate bidirectional influence between scales. The right column shows the omics or clinical measurement layer that most directly reads each scale (dashed arrows). The lower panel restates the three cardinal rules of host-response trajectories proposed by Ghazal et al. (2022), which motivate integration across scales. Clinical scores sit at the top because they record downstream dysfunction and are largely blind to the molecular endotypes below. Synthesised from Gonzalez-Barbuzano et al. (2026), Z. Han et al. (2024), Hudu et al. (2026), Lee et al. (2026), White et al. (2023) and L. Zhang et al. (2026).

Figure 2. Time-dependent multi-omics integration (TDMI) resolves dual TLR4 signalling arms in sepsis-associated liver dysfunction. Each omics layer was sampled at the time point matching its molecular kinetics: phosphoproteomics at 4 h (2,462 altered phosphosites), transcriptomics at 6 h (induction of DAMP genes such as S100A11, QPCT and IFITM2) and proteomics with metabolomics at 18 h (glutathione, taurine and amino acid changes), against a background of genomic and epigenomic priming. Two-step integration with xMWAS (correlation threshold |r| > 0.4) grouped features into six molecular communities, exposing parallel TLR4→MyD88→NF-κB and TLR4→TRIF→IRF3 pathways. Neither pathway was evident when any single layer was analysed alone. The data derive from an experimental model, and equivalent sampling windows in patients remain undefined. Adapted from the findings of Na et al. (2023).

with a hepatocellular pattern (96%), and tracks with longer stays, vasopressor need and mechanical ventilation, adding information beyond PRISM III and PELOD-2 (Gürsoy Koca et al., 2026). Earlier paediatric cohorts reported broadly similar risk factors and outcomes (Direito et al., 2021; Mogahed et al., 2020), and simple indices such as the bilirubin-to-albumin ratio and total bile acids have shown prognostic promise (Kang et al., 2023; Wang et al., 2025).

2.3.2 Sepsis-induced cardiomyopathy

Cardiac dysfunction in sepsis looks, at heart, like an energy crisis. Omics studies show that SIC involves upregulated PDK4, which phosphorylates and inactivates pyruvate dehydrogenase (PDH) (Z. Han et al., 2024; Shimada et al., 2022). With PDH blocked, pyruvate cannot enter the TCA cycle and the myocardium shifts from oxidative phosphorylation toward inefficient anaerobic glycolysis (Z. Han et al., 2024; Xu et al., 2020). At the same time, malonyl-CoA accumulates and drives hyper-malonylation of voltage-dependent anion channel 2 (VDAC2), leading to mitochondrial ROS accumulation and ferroptotic cardiomyocyte death (Z. Han et al., 2024; She et al., 2023). Targeted metabolomics in endotoxaemic myocardium further links energy-metabolic reprogramming to cardiac protection (M. Kim et al., 2022), and epitranscriptomic regulators such as METTL3 and YTHDC1 appear to prime myocardial inflammation upstream of these events (Shen et al., 2022; Xie et al., 2023).

2.3.3 Sepsis-associated encephalopathy and the gut–organ axis

SAE is a serious neurological complication, marked by acute cognitive dysfunction and, in survivors, long-term impairment. Widely targeted plasma metabolomics with ML algorithms such as XGBoost has identified disturbances in glycerophospholipid and ether-lipid metabolism, and accumulation of succinic acid, as core features separating SAE from uncomplicated sepsis (Baizhanova et al., 2025; Hu et al., 2026). Blood–brain barrier disruption through endothelial TREM-1 signalling adds a vascular component (Su et al., 2025), although the diagnostic value of individual brain biomarkers is still debated (Scarlatescu et al., 2025). Integrated microbiome–metabolome studies, for their part, emphasise the gut–brain and wider gut–organ axes. Sepsis-induced dysbiosis leads to loss of SCFA-producing taxa, breakdown of the intestinal epithelial barrier and systemic translocation of microbial damage-associated molecular patterns (DAMPs), which in turn seem to amplify immunometabolic instability and organ failure in distant tissues (Hudu et al., 2026).

2.4 Multi-marker risk stratification, AI integration and translational horizons

2.4.1 Combining scores and biomarker panels

To bring these insights to the bedside, much recent work has combined clinical organ-failure scores with targeted multi-marker panels. Combining established markers (PCT, PSEP) with emerging vascular and innate immune mediators (IFN-λ3, bio-ADM) gave better diagnostic and prognostic performance than SOFA alone (Lee et al., 2026). Patients with all four markers elevated had in-hospital mortality above 58%, reaching 66.7% in those aged 75 years or older, with a hazard roughly three times that derived from SOFA in isolation (Lee et al., 2026). In viral and COVID-19-associated sepsis, adding ferritin, C-reactive protein (CRP) and lactate dehydrogenase (LDH) to SOFA-2 sharpened cross-sectional assessment of organ-failure severity (Aytürk et al., 2026), consistent with earlier work on systemic inflammation and mortality risk (Carcillo et al., 2017). These clinical tools and their incremental value are compared in Table 3 (Table 3).

2.4.2 Persistent translational bottlenecks

The promise is real, but so are the obstacles. Omics assays often have turnaround times that do not fit the decision window of acute septic shock (Gonzalez-Barbuzano et al., 2026). Technical variation across sequencing platforms and mass spectrometers complicates comparisons between cohorts (Gonzalez-Barbuzano et al., 2026; Na et al., 2023). And many discovery cohorts are small, single-centre and drawn largely from adult European populations, which limits their relevance to children, neonates and more diverse groups (Gonzalez-Barbuzano et al., 2026; Gürsoy Koca et al., 2026).

2.4.3 Artificial intelligence and targeted delivery

AI and ML frameworks, including random forest, XGBoost and deep neural networks, are being used to extract small, highly predictive biomarker subsets from complex omics data (Baizhanova et al., 2025; Hudu et al., 2026; Min et al., 2017). On the therapeutic side, nanocarrier and biomimetic encapsulation platforms may allow

Figure 3. Organ-specific failure programmes in sepsis: shared systemic drivers expressed as distinct subcellular lesions and cell death pathways. A common dysregulated host response (centre), combining hyperinflammation, immune paralysis and bioenergetic collapse, reaches each organ (solid arrows) but is translated into organ-specific mechanisms. Each box names the dominant metabolic or signalling lesion and the regulated cell death (RCD) programmes implicated: glutathione/taurine loss and STING signalling in the liver, PDK4–PDH blockade and VDAC2-linked ferroptosis in the heart, succinate accumulation and TREM-1-mediated barrier failure in the brain, acylcarnitine accumulation and Hippo/ACSL4 pericyte loss in the kidney, histone- and NET-driven injury in the lung, and SCFA depletion with microbial translocation in the gut. The dashed arrow marks the gut–brain axis as an example of gut-derived amplification. Synthesised from Hu et al. (2026), Hudu et al. (2026), Z. Han et al. (2024), P. Li et al. (2026), Liu et al. (2024), Mohsin et al. (2025), She et al. (2023) and Su et al. (2025).

Figure 4. From bedside scores to integrated, AI-assisted risk stratification in sepsis, and the bottlenecks that currently limit translation. A patient with suspected sepsis can be assessed along three parallel tracks: physiological scores (fast and inexpensive but blind to endotype), targeted multi-marker immunoassays (available within hours), and multi-omics profiling processed through quality control, batch correction, feature selection and machine-learning classification with explainability tools. Combining tracks improves stratification; for example, positivity for all four markers (PCT, PSEP, bio-ADM, IFN-λ3) identified 58.3% in-hospital mortality (66.7% in patients aged ≥ 75 years) versus 31.3% for an elevated SOFA score alone, and an optimised SAE metabolite panel reached validation AUCs of 0.92–0.98. The intended end point is endotype-guided, organ-targeted therapy. Dashed boxes list the five main translational barriers. Synthesised from Gonzalez-Barbuzano et al. (2026), Hu et al. (2026), Hudu et al. (2026), Lee et al. (2026) and Z. Li et al. (2025).

anti-inflammatory, antioxidant or metabolic agents to be delivered to injured tissue while limiting systemic toxicity (Z. Li et al., 2025). How these strands might fit together, and where they currently stall, is outlined in Figure 4 and Table 4 (Figure 4; Table 4).

3. Methods

3.1 Review design and rationale

We designed this work as a structured narrative review rather than a formal systematic review with meta-analysis. The choice was deliberate, if not entirely comfortable. The literature on sepsis multi-omics mixes animal models, single-centre human cohorts, secondary database analyses and methodological papers, and outcome definitions, sampling times and analytical platforms differ so widely that statistical pooling would, we felt, give a false sense of precision (Gonzalez-Barbuzano et al., 2026; Hudu et al., 2026). We nonetheless borrowed the transparency elements of systematic reviewing: a pre-specified question, an explicit search strategy, stated eligibility criteria, duplicate screening and structured data extraction. The reporting follows the items of the PRISMA 2020 statement where these apply to a narrative synthesis. No protocol was registered, and no patient data were collected, so ethical approval was not required.

The guiding question was framed loosely in a population–exposure–outcome format: in patients or experimental models with sepsis (population), what do single-layer and integrated omics analyses, clinical multi-marker panels and AI/ML models (exposure) reveal about the mechanisms, prediction and stratification of organ dysfunction and death (outcomes)? Sepsis was defined according to Sepsis-3 (Singer et al., 2016), or the definition in use at the time for older studies.

3.2 Information sources and search strategy

We searched PubMed/MEDLINE, Scopus and Web of Science Core Collection for records published between 1 January 2016, the year Sepsis-3 was introduced, and [final search date, 2026]. Google Scholar was used only to trace citations and to locate articles in press, and the reference lists of key reviews (Gonzalez-Barbuzano et al., 2026; Z. Han et al., 2024; Hudu et al., 2026; L. Zhang et al., 2026) were hand-searched. A small number of earlier foundational papers, such as the plasma proteome genomic atlas (Sun et al., 2018) and work on deep learning in bioinformatics (Min et al., 2017), were included because later studies depended on them methodologically.

Search strings combined three concept blocks with the Boolean operator AND, and terms within each block with OR. The first block captured the condition ("sepsis" OR "septic shock" OR "multiple organ dysfunction" OR "MODS"). The second captured the analytical approach ("multi-omics" OR "multiomics" OR genomic OR epigenom OR transcriptom OR "single-cell" OR proteom OR phosphoproteom OR metabolom OR lipidom OR metagenom OR microbiome OR "eQTL" OR "Mendelian randomization" OR "machine learning" OR "artificial intelligence" OR "biomarker panel"). The third captured organ outcomes ("liver injury" OR "hepatic dysfunction" OR cardiomyopathy OR encephalopathy OR "acute kidney injury" OR "acute lung injury" OR ARDS OR "organ failure" OR mortality). In PubMed, the corresponding MeSH headings (Sepsis; Shock, Septic; Multiple Organ Failure; Genomics; Proteomics; Metabolomics; Machine Learning) were added to free-text terms. The full strategy for each database is available from the authors on request, and we would encourage readers to adapt rather than copy it, since indexing of newer omics terms is still uneven.

3.3 Eligibility criteria

Studies were eligible if they (a) investigated sepsis, septic shock or an accepted experimental model such as caecal ligation and puncture or endotoxaemia; (b) applied at least one high-throughput omics technology, a multi-marker biomarker panel, or an AI/ML model to host-response data; and (c) reported an outcome related to organ dysfunction, disease severity, endotype or mortality. Both human and animal studies were accepted, because much mechanistic evidence, particularly for the heart, liver and brain, currently exists only in animal models (Hu et al., 2026; Na et al., 2023; She et al., 2023). Adult, paediatric and oncology populations were all eligible, given the specific interest in whether findings generalise across age groups (Gürsoy Koca et al., 2026; "Predictors of PICU Mortality," 2026).

We excluded conference abstracts without a full report, editorials and letters without original data (short research letters with original data were kept, e.g., Rodríguez-Pérez et al., 2025), studies limited to pathogen identification without any host-response analysis, and articles not available in English. Narrative and systematic reviews were not treated as primary evidence but were read for context and for their reference lists.

3.4 Study selection

Records from all databases were exported into a reference manager and de-duplicated. Two reviewers independently screened titles and abstracts against the eligibility criteria, then assessed full texts of potentially relevant records. Disagreements, which were not rare given the fuzzy boundary between "omics" and conventional biomarker studies, were resolved by discussion and, where needed, by a third reviewer. 

3.5 Data extraction

Using a piloted extraction form, we recorded for each study: first author and year; country and setting; design (experimental, prospective or retrospective cohort, case–control, database analysis); population or model and sample size; sepsis definition; biological matrix (whole blood, plasma, serum, urine, cerebrospinal fluid, stool, tracheal aspirate or tissue); omics layer(s) and analytical platform (for example RNA-seq, single-cell RNA-seq, LC-MS/MS in data-dependent or data-independent acquisition, UPLC-QTOF-MS, ¹H-NMR, 16S rRNA or shotgun metagenomic sequencing); sampling time points; integration or modelling method (for example xMWAS, weighted LASSO, random forest, XGBoost, deep neural networks, SHAP or LIME); key molecular findings; organ system; and quantitative performance measures (area under the ROC curve [AUC], accuracy, hazard ratios, net reclassification and integrated discrimination improvement). Where a study reported several models, we extracted the best-performing validated model and noted whether validation was internal or external. Numerical values are reported as given in the source publications; we did not recalculate them.

3.6 Appraisal of evidence quality

Because the included designs were so varied, we did not compute a single quality score. Instead, each study was appraised on a small set of domains that, in our view, matter most for omics-based sepsis research: the adequacy of sample size relative to the number of features; whether batch effects and platform variation were addressed; whether sampling time was defined relative to sepsis onset; whether results were validated in an independent or external cohort; and whether clinical confounders, including severity scores, were adjusted for (Gonzalez-Barbuzano et al., 2026; Hudu et al., 2026; Na et al., 2023). These judgements were used to weight the narrative, not to exclude studies, and we flag weaker evidence explicitly in the text.

3.7 Synthesis strategy

Evidence was synthesised thematically across four biological scales: subcellular (mitochondrial bioenergetics and RCD), cellular (immune-cell subsets and endothelium), organ (liver, heart, brain, kidney, lung and gut) and systemic or clinical (scores, multi-marker panels and AI/ML models). For each scale we asked what was measured, which omics layer measured it, and whether integration added information beyond single layers. The resulting framework is summarised in Figure 1, the temporal integration logic in Figure 2, the organ-specific map in Figure 3 and the translational pathway in Figure 4. Study-level details are compiled in Tables 1 to 4. Figures were drawn by the authors from the synthesised evidence using Python (matplotlib); they are schematic and do not display pooled estimates.

4. A Multi-Scale Synthesis of Host Dysregulation, Organ Failure and Risk Prediction in Sepsis

Taken together, the studies reviewed here give a clearer, though still incomplete, picture of host dysregulation in sepsis. We present the findings in four connected domains: multi-scale omics signatures, organ-specific mechanisms, clinical risk stratification and AI-based integration (Figure 1).

4.1 Multi-scale omics mapping of host dysregulation

4.1.1 Genomic and epigenomic priming

Omics platforms now cover almost every layer of the host response, from DNA variation to the gut microbiome (Gonzalez-Barbuzano et al., 2026) (Table 1). At the genomic level, eQTL mapping identified 16,049 eQTLs in sepsis, and a subset of 1,578 SNP–gene pairs interacted with sepsis response signatures (SRS), linking inherited variation to particular immune-cell states (Burnham et al., 2024; Gonzalez-Barbuzano et al., 2026). Epigenetic machinery appears to act upstream. N⁶-methyladenosine (m⁶A) RNA methylation, through the METTL3/IGF2BP1/HDAC4 and YTHDC1/SERPINA3N axes, seems to prime myocardial inflammation and injury, at least in experimental models (Z. Han et al., 2024; Shen et al., 2022; Xie et al., 2023).

4.1.2 Transcriptomic and phosphoproteomic signalling

Bulk, single-cell and spatial transcriptomics capture broad induction of gene networks in acute sepsis (Gonzalez-Barbuzano et al., 2026) and, when harmonised, can be reduced to a smaller set of reproducible response signatures (Scicluna et al., 2025). In septic mouse myocardium, 158 tRNA-derived small RNAs changed expression, apparently modulating autophagic flux (Yuan et al., 2022). DAMP-related genes such as S100A11, QPCT and IFITM2 were rapidly induced through TLR4 signalling (Na et al., 2023). Phosphoproteomics picked up signalling even earlier, with 2,462 phosphosites altered (1,441 quantified) as soon as 4 h after the septic insult (Na et al., 2023).

4.1.3 Proteomic, metabolomic and metagenomic signatures

Plasma proteomics has produced a 9-protein signature (including GPX3, APOB, ORM1, SERPINF1, LYZ, C8A, CD14, APOC3 and C1QC) associated with acute organ dysfunction and a separate 22-protein panel associated with ICU mortality (Palmowski et al., 2025; Ruiz-Sanmartín et al., 2022). In the metabolome and lipidome, bioenergetic collapse shows up as rising lactate, branched-chain amino acid breakdown and loss of protective lysophosphatidylcholines (lysoPC C16:0, C17:0, C18:0), together with increased free fatty acids and succinate (Baizhanova et al., 2025; Hu et al., 2026; Pandey, 2025; White et al., 2023). Metagenomic sequencing describes marked gut remodelling, with loss of SCFA-producing Lachnospiraceae and Ruminococcaceae and circulating microbial cell-free DNA (Hudu et al., 2026). In the airway, metagenomics combined with host-response profiling improved diagnosis of lower respiratory infection (Langelier et al., 2018), and antibiotic resistance genes in tracheal aspirates were associated with mortality in non-pulmonary sepsis (Rodríguez-Pérez et al., 2025); clinical metagenomic next-generation sequencing (mNGS) is increasingly used when blood cultures are negative (Fang et al., 2026).

4.1.4 The added value of multi-layer integration

Where direct comparisons exist, integration outperformed single layers. TDMI uncovered parallel TLR4/MyD88/NF-κB and TRIF/IRF3 pathways in SALD that were invisible in each layer on its own (Na et al., 2023) (Figure 2). Likewise, combining immature CD10⁻ neutrophils with vascular mediators (pentraxin-3, angiopoietin-2) and lipid features in a weighted LASSO model reached a classification accuracy of 0.967 ± 0.039 and a ROC-AUC of 0.969 ± 0.040 for severe sepsis (Meghraoui-Kheddar et al., 2022; White et al., 2023). Such near-perfect figures in modest cohorts should probably be read as upper bounds rather than expected real-world performance.

4.2 Organ-specific pathophysiology, subcellular bioenergetics and regulated cell death

The systemic response resolves into distinct patterns of failure in each organ (Figure 3; Table 2).

4.2.1 Liver

In paediatric ICUs, secondary hepatic dysfunction affected 8.8% of critically ill children, predominantly with a hepatocellular pattern (96%) and raised ALT, AST and bilirubin (Gürsoy Koca et al., 2026). Similar rates and predictors were reported in earlier paediatric series (Direito et al., 2021; Mogahed et al., 2020; Saini et al., 2022). At the molecular level, multi-omics profiling showed disturbed amino acid metabolism, impaired taurine and hypotaurine pathways and glutathione depletion (P. Li et al., 2026; Na et al., 2023). Salvianolic acid A restored amino acid remodelling, suppressed STING/TBK1/IRF3-driven inflammation and blocked hepatocyte apoptosis, suggesting that hepatic protection needs coordinated action on several pathways (P. Li et al., 2026). Elevated total bile acids independently predicted death in paediatric sepsis (Wang et al., 2025).

4.2.2 Heart

Myocardial depression in sepsis appears to be driven by PDK4 upregulation and the resulting phosphorylation and inactivation of PDH (Z. Han et al., 2024; Shimada et al., 2022). Pyruvate entry into the TCA cycle falls, and metabolism shifts toward anaerobic glycolysis (Z. Han et al., 2024; Xu et al., 2020), with Sirt3 inactivation further impairing TCA enzymes (Xu et al., 2020). Elevated malonyl-CoA promotes VDAC2 hyper-malonylation, mitochondrial ROS accumulation and ACSL4-dependent ferroptosis (Z. Han et al., 2024; She et al., 2023; L. Zhang et al., 2026). Circulating extracellular histones above 75 μg/mL add to contractile failure by causing cardiomyocyte calcium overload (G. Zhang et al., 2022). Pharmacological activation of beclin-1 protected the endotoxaemic heart in association with metabolic reprogramming (M. Kim et al., 2022).

4.2.3 Brain

Table 1. Multi-omics profiling of the septic host response across biological scales: sample matrices, analytical platforms, key perturbations and prognostic applications. Each row summarises one omics layer, ordered roughly from inherited and regulatory priming (genomics, epigenomics) through signalling and effector molecules (transcriptomics, proteomics) to functional and ecological readouts (metabolomics, metagenomics), ending with integrated multi-layer approaches. Columns give the biological matrices and platforms typically used, the principal molecular perturbations reported in sepsis, and the way each layer has been applied to prognosis or endotyping. Numerical values (e.g., phosphosite counts, AUC) are reproduced as reported in the cited studies and were derived from both human cohorts and experimental models, so they should not be compared directly across rows. Abbreviations: GWAS, genome-wide association study; PRS, polygenic risk score; eQTL, expression quantitative trait locus; MeRIP-seq, methylated RNA immunoprecipitation sequencing; m⁶A, N⁶-methyladenosine; SRS, sepsis response signature; tsRNA, tRNA-derived small RNA; DAMP, damage-associated molecular pattern; LC-MS/MS, liquid chromatography–tandem mass spectrometry; DDA/DIA, data-dependent/data-independent acquisition; 2-DGE, two-dimensional gel electrophoresis; PDK4, pyruvate dehydrogenase kinase 4; PDH, pyruvate dehydrogenase; BCAA, branched-chain amino acid; lysoPC, lysophosphatidylcholine; SAE, sepsis-associated encephalopathy; SCFA, short-chain fatty acid; cfDNA, cell-free DNA; mNGS, metagenomic next-generation sequencing; TDMI, time-dependent multi-omics integration; PTX3, pentraxin-3; Ang-2, angiopoietin-2; AUC, area under the receiver operating characteristic curve.

Table 2. Organ-specific pathophysiology in sepsis: subcellular and bioenergetic mechanisms, regulated cell death pathways and candidate biomarkers. The table contrasts six organ systems commonly affected in sepsis-induced multi-organ failure, showing how a shared set of upstream disturbances (mitochondrial failure, oxidative stress and inflammation) is expressed differently in each tissue. For every organ, the dominant subcellular and bioenergetic lesion, the regulated cell death (RCD) programmes implicated, and the circulating or imaging biomarkers proposed so far are listed. Much of the mechanistic evidence, particularly for the heart, brain and kidney, comes from rodent models and awaits confirmation in patients. Abbreviations: SILI, sepsis-induced liver injury; SIC, sepsis-induced cardiomyopathy; SAE, sepsis-associated encephalopathy; SA-AKI, sepsis-associated acute kidney injury; ALI/ARDS, acute lung injury/acute respiratory distress syndrome; TCA, tricarboxylic acid; VDAC2, voltage-dependent anion channel 2; BBB, blood–brain barrier; NSE, neuron-specific enolase; ONSD, optic nerve sheath diameter; suPAR, soluble urokinase plasminogen activator receptor; bio-ADM, bioactive adrenomedullin; NET, neutrophil extracellular trap; mtDNA, mitochondrial DNA; LPS, lipopolysaccharide.

For the central nervous system, widely targeted LC-MS/MS metabolomics identified succinate as a key metabolite, rising stepwise from healthy controls to sepsis to SAE (Hu et al., 2026). Giving exogenous succinate to caecal ligation and puncture mice worsened cognitive deficits, hippocampal CA1/CA3 neuronal loss and Iba1⁺ microglial activation (Hu et al., 2026), which lends the association some causal weight. Endothelial TREM-1 activation disrupted the blood–brain barrier via PI3K/Akt signalling (Su et al., 2025). At the bedside, an optic nerve sheath diameter above 5.2 mm on ultrasound detected raised intracranial pressure with an AUC of 0.932 (Hakeem et al., 2025), and bioinformatic screens have proposed further SAE biomarkers (Lei et al., 2024), although no single marker is yet established (Scarlatescu et al., 2025).

4.2.4 Kidney, lung and gut

In SA-AKI, toxic acylcarnitines accumulate as β-oxidation fails, and pericytes are lost through Hippo/ACSL4-dependent ferroptosis (Liu et al., 2024). Transcriptomic endotyping and protein biomarkers seem to be complementary rather than interchangeable for risk stratification in SA-AKI (Tavris et al., 2025). In acute lung injury and ARDS, extracellular histones and neutrophil extracellular traps promote endothelial pyroptosis, while mitochondrial dysfunction and macrophage polarisation shape injury, and urolithin-A attenuated lung damage experimentally (Mohsin et al., 2025; White et al., 2023). In the gut, sepsis depletes faecal SCFAs such as acetate and butyrate, increasing epithelial permeability and allowing luminal DAMPs to reach the circulation (Hudu et al., 2026).

4.3 Clinical risk stratification and multi-marker panels

4.3.1 Conventional scores and paediatric hepatic criteria

Traditional tools, including SOFA, qSOFA, APACHE II, PRISM III and PELOD-2, offer standardised bedside assessment but cannot capture molecular endotypes (Gonzalez-Barbuzano et al., 2026; Gürsoy Koca et al., 2026; Lee et al., 2026) (Table 3). In paediatric cohorts, PRISM III and PELOD-2 separated survivors from non-survivors (p ≤ 0.01), whereas conventional liver criteria (ALT ≥ 2 × upper limit of normal, total bilirubin > 2 mg/dL) did not independently predict death after adjustment (p = 0.501) (Gürsoy Koca et al., 2026). Secondary hepatic dysfunction was, however, clearly associated with morbidity: a PICU stay longer than 7 days (p = 0.04) and prolonged vasopressor need (Gürsoy Koca et al., 2026). The bilirubin-to-albumin ratio may add prognostic information in this setting (Kang et al., 2023).

4.3.2 The four-marker immuno-vascular panel

In a prospective study of 248 patients with suspected sepsis, Lee et al. (2026) combined PCT and PSEP with IFN-λ3 and bio-ADM. Several findings stand out. Presepsin discriminated sepsis better than procalcitonin (AUC 0.78 vs. 0.70, p = 0.034), while bio-ADM (AUC 0.74) and IFN-λ3 (AUC 0.61) performed comparably to each other. Patients with all four markers elevated had in-hospital mortality of 58.3%, against 31.3% for those identified by an elevated SOFA score alone; the reported hazard ratios were 14.7 and 4.6, respectively. Among patients aged 75 years or older, four-marker positivity was associated with 66.7% mortality (HR 9.2, 95% CI 1.4–110.1), and no deaths occurred when none of the markers was raised. Adding the panel to SOFA improved net reclassification and integrated discrimination by up to 82% (p = 0.02) (Lee et al., 2026). The very wide confidence interval in the elderly subgroup is a reminder of how few events underpin these estimates.

4.3.3 Viral sepsis and paediatric oncology

In COVID-19-associated sepsis, day-5 serum ferritin above 1,191 ng/mL acted as a fairly specific rule-in marker of more severe organ failure, alongside the PaO₂/FiO₂ ratio, LDH and CRP, whereas procalcitonin added little, plausibly because the primary insult was viral (Aytürk et al., 2026). This echoes earlier paediatric work linking hyperferritinaemic inflammation to mortality (Carcillo et al., 2017). Among children with cancer admitted to intensive care, overall PICU mortality was 19% and organ failure occurred in 39%; the need for mechanical ventilation was the main independent predictor of death (HR 3.02, p < 0.01), while PRISM-IV lost independent significance in multivariable models (p = 0.146) ("Predictors of PICU Mortality," 2026).

4.4 Artificial intelligence, integrative workflows and translational horizons

4.4.1 Machine-learning classifiers

AI and ML architectures have been built into several discovery pipelines to turn high-dimensional omics data into decision support (Baizhanova et al., 2025; Hudu et al., 2026) (Table 4; Figure 4). In SAE, 1,307 quality-controlled plasma features were compared across random forest

support vector machines, logistic regression and XGBoost, with XGBoost performing best (Hu et al., 2026). An optimised 12-metabolite subset, led by elevated succinate, depleted PC(18:2_22:6), loss of ceramide Cer(d24:1/18:0) and carnitine C22:6, reached an independent test accuracy of 93% and AUCs of 0.924 to 0.975 (Hu et al., 2026). A clinical XGBoost model using time-series electronic health record (EHR) variables in older ICU patients predicted SAE onset within 24 h with an AUC of 0.898 (Y. Han et al., 2025).

4.4.2 Microbiome, EHR and explainable models

For microbiome–metabolome integration, random forest models handled non-linear interactions in sparse 16S/mNGS abundance and metabolite data and outperformed logistic regression for 28-day mortality (Hudu et al., 2026). Deep neural networks and transformer models trained on MIMIC-IV time series forecast the need for continuous renal replacement therapy and ICU delirium trajectories (Sheikhalishahi et al., 2023; Song et al., 2025), building on broader deep-learning methods in bioinformatics (Min et al., 2017). To soften the "black box" problem, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) have been used to convert feature matrices into risk scores clinicians can interrogate (Hudu et al., 2026; Yagin et al., 2024).

4.4.3 Causal inference and targeted delivery

Proteogenomic strategies that combine GWAS variants with plasma protein QTLs and eQTLs use Mendelian randomisation to distinguish upstream, potentially druggable causes from reactive markers (Burnham et al., 2024; Gonzalez-Barbuzano et al., 2026; Sun et al., 2018). Nanocarrier and biomimetic encapsulation platforms, meanwhile, aim to deliver antioxidant and anti-inflammatory agents to injured microvascular beds while limiting systemic toxicity (Z. Li et al., 2025).

4.4.4 Key translational bottlenecks

Five hurdles recur across the literature. Turnaround for LC-MS/MS and RNA-seq workflows often exceeds 24 h, which does not fit the resuscitation window of septic shock (Gonzalez-Barbuzano et al., 2026; Na et al., 2023). Inter-laboratory and cross-platform batch effects hinder standardisation (Gonzalez-Barbuzano et al., 2026; Hudu et al., 2026). Discovery cohorts are frequently small, single-centre and adult European, limiting generalisability (Gonzalez-Barbuzano et al., 2026; Gürsoy Koca et al., 2026). High feature-to-sample ratios invite overfitting without prospective validation (Hudu et al., 2026). And real-time ML alerts need human-factors safeguards to avoid alert fatigue (Hudu et al., 2026). These constraints are summarised alongside each modelling framework in Table 4 and depicted in Figure 4.

5. Toward Endotype-Guided Critical Care: Interpreting Multi-Scale Evidence in Sepsis-Induced Organ Failure

5.1 Principal findings

This review set out to trace sepsis-induced organ failure across biological scales, and the main message is, we think, fairly consistent even if the evidence behind it is uneven. Organ failure in sepsis does not look like a single process that simply spreads. It looks more like a shared set of upstream disturbances, chiefly mitochondrial energy failure and dysregulated cell death (Z. Han et al., 2024; L. Zhang et al., 2026), that each organ then translates into its own failure programme (Figure 3; Table 2). Integration across omics layers repeatedly revealed things that single layers missed, the dual TLR4 arms in SALD being the clearest example (Na et al., 2023) (Figure 2). And at the clinical end, combining biomarkers from distinct pathways added prognostic information beyond SOFA (Lee et al., 2026) (Table 3). What is still missing is the bridge: very few studies connect a molecular signature to a treatment decision in real time.

5.2 Bioenergetics as a common denominator

If there is one thread that runs through every scale, it is energy. The PDK4-mediated block at PDH in the heart (Z. Han et al., 2024; Shimada et al., 2022), acylcarnitine accumulation in the kidney (Liu et al., 2024), succinate build-up in the brain (Hu et al., 2026) and the systemic lactate and lipid shifts seen in plasma (Pandey, 2025; White et al., 2023) can all be read as variations on one theme: cells that cannot oxidise fuel efficiently. This fits the second of the cardinal rules proposed by Ghazal et al. (2022), that bioenergetic demand shapes host reactivity, and it links naturally to immunometabolic reprogramming of immune cells (H. Kim et al., 2026). It is tempting to go further and argue that metabolic failure is the primary event. We would be cautious. Much of the mechanistic evidence comes from rodent models, and the direction of causality between inflammation, hypoperfusion and mitochondrial injury is still hard to pin down in patients. Succinate is an instructive exception, because supplementation experiments worsened neurological injury (Hu et al., 2026), but even there, the translation to human SAE remains to be tested.

5.3 Organ-specific programmes and inter-organ crosstalk

Although the upstream drivers overlap, the organ-level readouts differ in ways that probably matter for therapy. Ferroptosis features prominently in the heart and kidney, via VDAC2 malonylation and Hippo/ACSL4 signalling respectively (Liu et al., 2024; She et al., 2023), while pyroptosis and barrier failure seem more central in the brain and lung (Mohsin et al., 2025; Su et al., 2025), and redox and amino acid remodelling dominate in the liver (P. Li et al., 2026) (Table 2). One implication, admittedly speculative, is that a single "anti-sepsis" drug may be less realistic than organ-directed combinations, perhaps delivered by nanocarriers (Z. Li et al., 2025) (Figure 4). The gut adds a further twist. Loss of SCFA-producing taxa and barrier breakdown may act less as a separate organ failure and more as an amplifier that feeds DAMPs and microbial products to the liver, brain and elsewhere (Hudu et al., 2026). We find this idea attractive, but the human data are mostly associative, and prospective studies that sample stool, blood and organ biomarkers together are rare.

5.4 Integration across time: why sampling schedules matter

The TDMI study makes a methodological point that deserves more attention than it usually gets. Different molecular layers change on different clocks: phosphorylation within hours, transcripts a little later, proteins and metabolites later still (Na et al., 2023) (Figure 2). Sampling every layer at one time point, as most clinical studies do, may therefore align signals that were never contemporaneous, or miss early events altogether. Longitudinal plasma proteomics has shown that the proteome itself shifts in ways associated with mortality (Palmowski et al., 2025). In patients, of course, "time zero" is rarely known. Anchoring samples to ICU admission, first antibiotic dose or onset of organ dysfunction is an imperfect compromise, and we suspect some of the inconsistency between cohorts reflects this rather than true biological disagreement.

5.5 From biomarker panels to bedside decisions

The multi-marker data are among the most clinically approachable in this review (Table 3). A four-marker panel identifying a group with 58.3% mortality, compared with 31.3% for elevated SOFA alone, and a group of older patients with no deaths when all markers were negative (Lee et al., 2026), suggests a real capacity both to rule in and to rule out risk. Still, the elderly subgroup confidence interval (1.4–110.1) shows how thin the event counts were, and single-cohort estimates of this kind tend to shrink on external validation. The COVID-19 data add a useful nuance: marker performance depends on the type of infection, and procalcitonin, so useful in bacterial sepsis, contributed little in viral sepsis (Aytürk et al., 2026). The paediatric evidence adds another. Liver dysfunction strongly predicted morbidity but not mortality once global scores were accounted for (Gürsoy Koca et al., 2026), and in children with cancer, the need for ventilation outweighed PRISM-IV ("Predictors of PICU Mortality," 2026). A marker, in other words, may be prognostically useful for one outcome and not another, and reviews that report only mortality may undervalue it.

5.6 Machine learning: promise, performance and pitfalls

The ML results look impressive on paper, with validation AUCs above 0.92 for SAE metabolite classifiers (Baizhanova et al., 2025; Hu et al., 2026) and near-perfect accuracy for integrated immune-metabolic signatures (White et al., 2023) (Table 4). We are less sure how much of this will survive contact with new hospitals. Feature-to-sample ratios in many studies are high, cohorts are often under 100 patients, and external validation is the exception rather than the rule (Hudu et al., 2026). Tree-based methods such as random forest and XGBoost handle non-linear, sparse data well, and explainability tools such as SHAP and LIME help (Yagin et al., 2024), but they do not fix a biased training set. Deep learning on EHR time series (Min et al., 2017; Sheikhalishahi et al., 2023; Song et al., 2025) raises additional concerns about missing data, privacy and computational cost. The most defensible near-term role for AI may be modest: selecting small, measurable marker sets from large omics discovery data, which can then be tested prospectively with ordinary assays.

5.7 Causality, endotypes and therapeutic targeting

Most omics biomarkers are, to be blunt, correlates. Proteogenomic approaches that combine GWAS, pQTL and eQTL data with Mendelian randomisation offer one way to separate causes from consequences (Burnham et al., 2024; Sun et al., 2018), and the identification of consensus transcriptomic signatures (Scicluna et al., 2025) and interferon-dependent endotypes (Fratea et al., 2026) points to the possibility of enrolling patients into trials by biology rather than by syndrome. Complementary evidence from SA-AKI suggests that transcriptomic endotypes and protein biomarkers capture overlapping but distinct risk (Tavris et al., 2025). Whether endotype-guided therapy will actually improve outcomes, though, is an open question that no study in this review answers.

5.8 Strengths and limitations

This review brings together evidence across molecular, cellular, organ and clinical scales, which is, as far as we know, less common than single-organ or single-omics reviews. It has limitations. It is a narrative synthesis, and study selection, although structured, is more open to judgement than in a formal systematic review. Many of the included studies are experimental or small and single-centre, and several rely on the same underlying cohorts or databases, notably MIMIC-IV. Reported performance metrics were taken as published, and we could not correct for optimism or for differences in outcome definitions. Publication bias toward positive biomarker studies is likely. Finally, the field moves quickly, and some 2026 publications cited here may be revised or superseded.

5.9 Future directions

Several priorities follow from this synthesis. Prospective, multicentre cohorts with harmonised sampling times and matrices are needed, ideally including children, neonates and populations outside Europe and North America (Gonzalez-Barbuzano et al., 2026; Gürsoy Koca et al., 2026). Rapid, targeted assays that measure a small number of omics-derived markers within clinically useful timeframes would do more for translation than ever-larger discovery panels. Batch-correction and reporting standards should be adopted across laboratories (Hudu et al., 2026). And, perhaps most importantly, biomarker-stratified trials, in which a molecular endotype determines treatment allocation, are needed to show that any of this changes outcomes (Figure 4).

6. Conclusion

Sepsis-induced multi-organ failure appears to be a dynamic systems failure that spans molecular, cellular, tissue and whole-patient scales. Across the studies reviewed, shared disturbances in mitochondrial energy metabolism and regulated cell death seem to be translated by each organ into distinct failure programmes: glutathione and STING signalling in the liver, VDAC2-linked ferroptosis in the heart, succinate accumulation in the brain, pericyte loss in the kidney, and microbial barrier breakdown in the gut. Time-resolved multi-omics integration and multi-marker panels have exposed biology, and risk, that bedside scores largely miss. Yet much of the evidence, however intriguing, remains preclinical, small in scale or externally unvalidated, and turnaround times are still too slow for septic shock. Multi-scale integration therefore looks promising rather than proven. Prospective, multicentre and endotype-stratified studies, with harmonised sampling and rapid assays, will probably decide whether it can move from explaining sepsis to actually treating it.

References