Plasma mitochondrial-associated proteins scale with illness severity and predict outcome in suspected sepsis

Abstract

Background

Mitochondrial dysfunction has been implicated in sepsis-related organ failure. We investigated whether plasma profiles of mitochondrial-associated proteins are associated with illness severity and short-term outcomes and explored whether they reveal distinct mitochondrial subphenotypes.

Methods

In this post hoc analysis of the Acutelines data-biobank, we included 213 adult patients presenting to the emergency department with a suspicion of sepsis, with infection confirmed by independent adjudication. Blood plasma proteins of patients were quantified with the Olink Explore HT platform; of 5,416 proteins, 232 mitochondrial-associated proteins (MitoCarta3.0) were z-standardized and averaged per patient to derive a continuous mitochondrial-protein score. To explore whether operationally distinct patient groups could be identified along this gradient, we additionally performed principal component analysis with K-means clustering on the same protein set. Differential abundance analysis, pathway enrichment, and least absolute shrinkage and selection operator regression were performed as exploratory analyses.

Results

The continuous mitochondrial-associated protein score increased with illness severity (Spearman ρ 0.18–0.29; p < 0.01) and worse 72-hour outcomes (odds ratio per standard deviation: organ failure-free survival 0.57; intensive care unit admission 1.66; mortality 2.36). Exploratory clustering identified two groups differing mainly in overall mitochondrial-associated protein abundance. Considerable overlap and limited support for a discrete two-cluster structure suggested variation along a continuum rather than clearly distinct mitochondrial subphenotypes. Exploratory pathway analyses implicated mitochondrial metabolism, cellular stress, and injury-related processes.

Conclusions

Higher plasma mitochondrial-associated protein abundance was associated with greater illness severity and worse short-term outcomes in suspected sepsis. These associations are hypothesis-generating, providing a first step toward understanding the role of mitochondrial dysfunction in sepsis severity.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13054-026-06291-4.

Introduction

Sepsis is defined as a life-threatening organ dysfunction caused by a dysregulated host response to infection [1]. It affects approximately 166 million people annually and remains one of the leading causes of mortality worldwide, with an estimated 21.4 million deaths per year [2]. A major challenge in managing sepsis is its profound heterogeneity. Patients differ not only in baseline characteristics and comorbidities, but also in the causative pathogen, site of infection, and underlying biological mechanisms, all of which contribute to highly variable clinical presentations and responses to treatment [3]. Growing evidence suggests that patients with sepsis can be subphenotyped based on shared biological characteristics [4]. Several studies have identified sepsis subphenotypes using gene expression profiles or clinical data [5, 6], revealing groups associated with disease severity and increased mortality. However, the underlying biological mechanisms contributing to organ failure in sepsis remain incompletely understood, limiting the development of targeted therapies [7].

One key biological mechanism contributing to sepsis-induced organ failure is mitochondrial dysfunction, characterized by impaired oxidative phosphorylation, reduced adenosine triphosphate (ATP) production, increased generation of reactive oxygen species, and ultimately cell death [8, 9]. During the early phase of sepsis, mitochondrial function has been shown to be severely impaired, with ATP synthesis reduced by up to 50% and mitochondrial biogenesis markedly diminished [10]. Reduced mitochondrial membrane potential and lower ATP levels are associated with organ failure and death in sepsis, and survivors consistently demonstrate better preservation of mitochondrial function compared to non-survivors [11]. Lactate, a frequently used marker of illness severity in sepsis, reflects several pathophysiological processes including, but not limited to, impaired mitochondrial oxidative metabolism, adrenergic stimulation of aerobic glycolysis, and altered lactate clearance; elevated lactate levels have been shown to predict mortality independently of organ dysfunction and shock [12, 13]. Beyond established clinical markers, circulating mitochondrial DNA and markers reflecting mitochondrial biogenesis and function have emerged as candidate biomarkers for predicting sepsis severity and mortality [8, 14]. However, despite this growing evidence, a comprehensive understanding of the heterogeneity in mitochondrial responses among sepsis patients and how distinct mitochondrial subphenotypes relate to clinical outcomes and mechanisms of organ failure is currently lacking [14, 15]. Characterizing this heterogeneity may provide a basis for more personalized sepsis management.

Therefore, this study aimed to determine the association between plasma abundance of mitochondrial-associated proteins and clinical severity and short-term outcomes in patients with suspected sepsis at emergency department (ED) presentation. As a secondary exploratory analysis, principal component analysis (PCA) and K-means clustering were used to explore potential mitochondrial subphenotypes based on mitochondrial-associated protein abundance.

Methods

Study setting: Acutelines

Data and samples were obtained from Acutelines, a multidisciplinary prospective data-biobank that captures the acute care journey of patients presenting to the ED of the University Medical Center Groningen (UMCG), a tertiary care teaching hospital serving the Northern Netherlands. Trained researchers screen and recruit eligible ED patients, while dedicated software automatically captures clinical data from the electronic health record (EHR) and bedside monitoring systems. Biomaterials are collected from patients presenting in either of the two highest urgency triage categories and, irrespective of triage category, from patients with shock or suspected sepsis. For this purpose, suspected sepsis is defined broadly and pragmatically based on the treating physician’s clinical judgement or fulfilment of Sepsis-2 or Sepsis-3 criteria. Participants provide written informed consent, with proxy consent when applicable. The Acutelines cohort is broadly representative of the regional acute care population, and full details of its design and participant selection have been published previously [16]. Acutelines was approved by the medical ethics committee of the UMCG (METC ID: 2019/589) and registered at ClinicalTrials.gov (NCT04615065) on September 1, 2020 [17].

For the present study, we selected adult patients (≥ 18 years) who presented to the ED with suspected sepsis, had available biomaterial samples, and in whom infection at initial presentation was retrospectively confirmed according to the Calandra criteria [18]. In cases of disagreement between adjudicators, a third reviewer made the final determination. The Calandra criteria are a consensus framework that defines microbiologically documented and clinically probable infection by site, integrating clinical, radiological and microbiological evidence. Throughout this manuscript, ‘early sepsis’ denotes this enrollment framework rather than a biological or temporal stage of sepsis progression.

Patients with a history of hematopoietic stem cell transplantation and patients receiving systemic immunosuppressive therapy for autoimmune or chronic inflammatory disease were not selected, so that the mitochondrial-protein signature reflects the host response to infection rather than the effects of therapeutic immunosuppression. Solid organ transplant recipients were not selected where possible; four were identified after inclusion and retained in the analysis. External validation will be required to establish the applicability of the findings to the broader ED sepsis population, including patients receiving chronic immunosuppression.

The current study was approved by the ethics committee of the UMCG (PaNaMa ID: 17551) and reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines [19].

Primary and secondary outcomes

The primary objective was to assess the association between plasma levels of mitochondrial-associated proteins and both short-term outcomes and clinical severity. As a secondary exploratory analysis, PCA and K-means clustering were used to identify potential mitochondrial subphenotypes. The primary clinical endpoint was organ failure-free survival within 72 h, defined as survival without admission to the intensive care unit (ICU) during this period. ICU admission and mortality within 72 h were additionally reported separately as components of this composite endpoint.

In addition, we evaluated whether the identified clusters were associated with the final diagnosis of sepsis, defined as an increase in the Sequential Organ Failure Assessment (SOFA) score of ≥ 2 points from baseline in the presence of confirmed or suspected infection.

Blood sampling and processing

Blood samples were collected in L-heparin tubes at ED presentation, before treatment initiation. Plasma was separated by centrifugation (1,500 × g, 15 min) and stored at -80 °C. Plasma protein quantification was performed using the Olink Explore HT platform (Olink Proteomics, Uppsala, Sweden) [20, 21], which enabled relative quantification of 5,416 proteins via proximity extension assay technology. Results were reported as Normalized Protein eXpression (NPX) values on a relative log2 scale. NPX values were generated using Olink’s standard pipeline. Technical performance was monitored with Olink’s built-in quality-control system, which comprises internal controls spiked into every sample (incubation, extension and detection controls) and external plate controls (inter-plate control, negative control and pooled sample/plate controls). Sample QC compares the detection and incubation controls of each sample against the plate median, with samples deviating by more than +/-0.3 NPX flagged and excluded; plate QC monitors control variation across each plate. Between-plate technical variation was removed using Olink’s recommended normalization (intensity normalization, centering each assay’s median across plates), and assays were evaluated against their limit of detection. No values were missing in the analyzed dataset [21].

Statistical analysis

Analyses were conducted in RStudio version 4.4.1. Continuous variables are presented as mean ± standard deviation (SD) or median interquartile range (IQR), depending on normality assessed by the Shapiro-Wilk test. Between-group comparisons were performed using the Wilcoxon rank-sum test for continuous variables and Fisher’s exact test for categorical variables. A p-value < 0.05 was considered statistically significant.

Mitochondrial-associated protein profiling: Of the 5,416 proteins detected by Olink Explore HT, mitochondrial-associated proteins were identified by cross-referencing with MitoCarta3.0, a curated, experimentally supported inventory of 1,136 human genes encoding proteins with verified mitochondrial localization, assembled from mass spectrometry across tissues, GFP co-localization and literature curation, and including sub-organelle and pathway annotations [22], yielding 232 mitochondrial-associated proteins for downstream analyses. A continuous mitochondrial-protein score was derived per patient as the mean z-standardized NPX value across these 232 proteins. Associations between the continuous score and clinical severity markers (national early warning score (NEWS), SOFA, lactate) were assessed using Spearman correlation. Associations with 72-hour clinical outcomes were assessed using logistic regression, expressed as odds ratios per 1 standard deviation increase in the score with 95% confidence intervals. To distinguish a mitochondrial-predominant signal from generalized intracellular protein release, all 5,416 proteins were assigned to a subcellular compartment, mitochondrial (Olink × MitoCarta3.0, n = 232), cytosolic (n = 901), nuclear (n = 1,500), or intracellular-membrane (n = 814), using the Human Protein Atlas (HPA) main-location annotation [23]; 1,969 proteins annotated as secreted, plasma-membrane, or without a main-location annotation were excluded. Composite compartment scores were computed as the mean z-standardized NPX within each compartment per patient. The residual mitochondrial enrichment score (RMES) was defined as the residual of the mitochondrial composite score regressed on the non-mitochondrial intracellular composite score, isolating the mitochondrial abundance beyond generalized intracellular protein release.

Exploratory clustering and downstream analyses: As a secondary exploratory aim, PCA with scaling was performed on the 232 mitochondrial-associated proteins. The number of clusters (k) was selected on the retained principal components by evaluating the within-cluster sum of squares (elbow), average silhouette width, and gap statistic for k = 2–10. K-means clustering was subsequently applied to the first three principal components. Differential abundance analysis (DAA) was performed using linear models on NPX values with the Limma package, adjusted for age, sex and illness severity (SOFA, NEWS and lactate), with multiple testing corrected by the Benjamini–Hochberg method; differentially abundant proteins were defined as adjusted P < 0.05 and |log2 fold change| > 0.1. A sensitivity analysis additionally adjusted for body mass index (BMI), heart failure, metformin and sodium-glucose cotransporter 2 inhibitor (SGLT2-inhibitor) use. Gene Ontology (GO) and Reactome pathway enrichment analyses were conducted using hypergeometric testing. Least absolute shrinkage and selection operator (LASSO)-penalized logistic regression was applied to the 232 mitochondrial-associated proteins with cluster membership as the outcome; the penalty was selected by 10-fold cross-validation, selection stability was assessed across 1,000 bootstrap resamples, and discrimination was tested against 1,000 label permutations. Sex-stratified analyses were additionally performed.

Results

Patient characteristics

In total, 213 adult patients with suspected sepsis and adjudicated infection were included at ED presentation; of these, 84% (n = 178) met Sepsis-3 criteria (Table 1). The median age was 67 years, and 35% (n = 74) were female. The most common comorbidities were hypertension (26%, n = 56), asthma/chronic obstructive pulmonary disease (COPD) (24%, n = 50), and ischemic heart disease (12%, n = 26). The median mean arterial pressure was 92 mmHg (IQR 80–105), heart rate 97 bpm (IQR 83–113), and respiratory rate 22/min (IQR 18–26). The median lactate was 1.4 mmol/L (IQR 0.9–2.3). Organ failure-free survival within 72 h was 72% (n = 153), ICU admission within 72 h was 24% (n = 52), and 72-hour mortality was 7% (n = 14).

Table 1.

Characteristics of the study population

	n = 213
Demographics
Female (n (%))	74 (35)
Age (median [IQR])	67 [57, 74]
BMI (median [IQR])	27.2 [23.6, 31.0]
Comorbidities (n (%))
Ischemic heart disease	26 (12)
Heart failure	23 (11)
Hypertension	56 (26)
Asthma/COPD	50 (23)
Kidney disease	17 (8)
Liver disease	11 (5)
Solid organ transplantation	4 (2)
Malignancy	23 (11)
Vital parameters at triage (median [IQR])
Systolic blood pressure [mmHg]	128 [109, 145]
Diastolic blood pressure [mmHg]	73 [63, 85]
Mean arterial blood pressure [mmHg]	92 [80, 105]
Heart rate [beats/min]	97 [83, 113]
Respiratory frequency [/min]	22 [18, 26]
Oxygen Saturation (SpO2) [%]	95 [93, 97]
Temperature (°C)	37.2 [36.6, 37.8]
GCS	15 [15, 15]
NEWS	5 [3, 7]
Lab values at triage (median [IQR])
Blood gas lactate (mmol/L)	1.4 [0.9, 2.3]
Leukocytes (×10⁹/L)	9.9 [7.8, 16.4]
Thrombocytes (×10⁹/L)	237 [160, 292]
C-reactive protein (mg/L)	110 [36, 224]
Creatinine (µmol/L)	89 [68, 123]
Bilirubin (µmol/L)	10 [7, 16]
Clinical endpoints (n (%))
Organ failure-free survival < 72 h	153 (72)
ICU admission < 72 h	52 (24)
Mortality < 72 h	14 (7)
Sepsis according to sepsis-3 criteria	178 (84)

Baseline characteristics of all included patients (n = 213) presenting to the emergency department with suspected infection. Demographics, comorbidities, vital parameters at triage, laboratory values at triage, and clinical endpoints are presented. Continuous variables are expressed as median [interquartile range]; categorical variables as n (%). Abbreviations: BMI, body mass index; COPD, chronic obstructive pulmonary disease; GCS, Glasgow Coma Scale; IQR, interquartile range; NEWS, national early warning score

Higher plasma mitochondrial-associated protein abundance is associated with greater illness severity and worse short-term outcomes

To assess the association between plasma profiles of mitochondrial-associated proteins and clinical severity and short-term outcomes, a continuous mitochondrial-protein score was derived as the mean z-standardized NPX value across the 232 MitoCarta3.0-annotated proteins. The mitochondrial-protein score was associated with clinical short-term outcomes: the primary endpoint, organ failure-free survival within 72 h (defined as the absence of both ICU admission and mortality), was inversely associated with the score (odds ratio per 1 SD increase: 0.57 [95% CI 0.42–0.78]). Its components, ICU admission and mortality within 72 h, showed concordant associations (ICU admission: OR 1.66 [1.22–2.27]; mortality: OR 2.36 [1.49–3.72]). The score was also positively associated with markers of illness severity and organ dysfunction, including NEWS, lactate and the SOFA score (Spearman ρ 0.18–0.29; all p < 0.01; Fig. 1).

Fig. 1.

Overall mitochondrial-protein score in relation to illness severity and clinical outcomes. The overall mitochondrial-protein score, defined as the mean z-score of the 232 mitochondrial proteins for each patient, is shown in relation to illness severity and 72-hour clinical outcomes. (A-C) Association between the mitochondrial protein score and illness severity at ED presentation: (A) NEWS, (B) SOFA, and (C) lactate. Linear regression lines with 95% confidence bands are shown, together with Spearman correlation coefficient (ρ) and corresponding p-value. (D-F) Association between mitochondrial protein score and 72-hour clinical outcomes: (D) organ failure-free survival, (E) ICU admission, and (F) mortality. The odds ratio per 1-SD increase in the mitochondrial-protein score is shown with the corresponding p-value

Exploratory clustering and downstream analyses

To explore whether operationally distinct patient groups could be identified along the mitochondrial-protein gradient, we applied K-means clustering to the first three principal components. This separated patients into two strata: cluster 1 (n = 135, lower abundance) and cluster 2 (n = 78, higher abundance; Fig. 2A–E). Data-driven criteria supported k = 2 as the most defensible operational partition, while the gap statistic favoring k = 1 confirmed that the strata represent a dichotomization of a continuum rather than discrete subphenotypes (Supplementary Figure S1). Cluster 2 patients had higher BMI, greater prevalence of heart failure, higher illness severity scores, and worse 72-hour outcomes than cluster 1, consistent with their position at the high end of the mitochondrial-protein gradient (Table 2; Fig. 2F-L).

Fig. 2.

Identification and clinical characterization of mitochondrial protein clusters. (A) Schematic overview of the protein selection strategy. From the complete Olink proteomics panel (n = 5,416 proteins), mitochondrial proteins were identified by cross-referencing with the MitoCarta3.0 database (n = 1,136), resulting in a final set of Olink mitochondrial proteins (n = 232) used for downstream analyses. (B) Scree plot showing the proportion of variance explained by each of the first six principal components (PCs) derived from the mitochondrial-associated protein profiles. (C-E) PCA plots showing the distribution of patients according to cluster assignment (cluster 1, black circles; cluster 2, grey triangles). Panels show PC1 vs. PC2 (C), PC2 vs. PC3 (D), and PC1 vs. PC3 (E). (F-I) Proportion of patients in each cluster according to 72-hour clinical outcomes and sepsis diagnosis: (F) organ failure-free survival < 72 h (%), (G) ICU admission < 72 h (%) and (H) mortality < 72 h (%), (I) final diagnosis of sepsis (%). P-values were calculated using Fisher’s exact test. (J-L) Distribution of markers of illness severity at ED presentation between clusters: (J) SOFA score, (K) NEWS score, and (L) lactate (mmol/L). P-values were calculated using the Wilcoxon rank-sum test

Table 2.

Clinical characteristics of cluster 1 and cluster 2

n (%)	Cluster 1	Cluster 2	p-value
	135 (63)	78 (37)
Demographics
Female (n (%))	47 (35)	27 (35)	1.000
Age (median [IQR])	67 [57, 74]	68 [59, 76]	0.397
BMI (median [IQR])	26.1 [23.2, 30.6]	28.7 [24.5, 32.6]	0.044
Comorbidities (n (%))
Ischemic heart disease	15 (11)	11 (14)	0.671
Heart failure	9 (7)	14 (18)	0.020
Hypertension	36 (27)	20 (26)	0.998
Asthma/COPD	33 (24)	17 (22)	0.786
Kidney disease	10 (7)	7 (9)	0.885
Liver disease	9 (7)	2 (3)	0.184
Solid organ transplantation	2 (1)	2 (3)	0.971
Malignancy	11 (8)	12 (15)	0.158
Vital parameters at triage (median [IQR])
Systolic blood pressure [mmHg]	129 [115, 148]	119 [103, 140]	0.019
Diastolic blood pressure [mmHg]	75 [65, 87]	71 [61, 83]	0.104
Mean arterial blood pressure [mmHg]	95 [81, 106]	89 [75, 100]	0.066
Heart rate [beats/min]	96 [80, 109]	100 [86, 116]	0.027
Respiratory frequency [/min]	21 [17, 25]	23 [19, 29]	0.014
Oxygen Saturation (SpO2) [%]	96 [93, 98]	95 [92, 97]	0.051
Temperature (°C)	37.2 [36.7, 37.8]	37.2 [36.1, 37.8]	0.454
GCS	15 [15, 15]	15 [15, 15]	0.994
NEWS	4 [3, 6]	6 [4, 8]	0.003
Lab values at triage (median [IQR])
Blood gas lactate (mmol/L)	1.3 [0.9, 1.9]	1.7 [1.0, 3.5]	0.002
Leukocytes (×10⁹/L)	9.9 [7.7, 15.7]	10.6 [8.1, 17.7]	0.216
Thrombocytes (×10⁹/L)	236 [162, 289]	247 [157, 301]	0.460
C-reactive protein (mg/L)	94 [32, 204]	150 [41, 246]	0.066
Creatinine (µmol/L)	80 [64, 102]	99 [78, 171]	< 0.001
Bilirubin (µmol/L)	10 [6, 16]	13 [8, 24]	0.005
Clinical endpoints (n (%))
Organ failure-free survival < 72 h	107 (79)	46 (59)	0.003
ICU admission < 72 h	24 (18)	28 (36)	0.005
Mortality < 72 h	4 (3)	10 (13)	0.008
Sepsis according to sepsis-3 criteria	107 (79)	71 (91)	0.034

Comparison of baseline demographics, comorbidities, vital parameters, laboratory values, and clinical endpoints between patients assigned to cluster 1 and cluster 2. Continuous variables are expressed as median [interquartile range]; categorical variables as n (%). P-values were calculated using the Wilcoxon rank-sum test for continuous variables and Fisher’s exact test for categorical variables. P < 0.05 is shown in bold. Abbreviations: BMI, body mass index; COPD, chronic obstructive pulmonary disease; ED, emergency department; GCS, Glasgow Coma Scale; ICU, intensive care unit; IQR, interquartile range; NEWS, national early warning score; SOFA, sequential organ failure assessment

To examine cluster-associated biology beyond the mitochondrial proteins used for clustering, DAA was repeated on the complete Olink panel (5,416 proteins). In the severity- and sex-adjusted model, 3,284/5,416 proteins were differentially abundant between the clusters, the large majority more abundant in cluster 2 (Fig. 3A). Effect sizes were essentially unchanged after additionally adjusting for BMI, heart failure, metformin and SGLT2-inhibitor use (log fold change correlation r = 0.99), and mitochondrion remained the most strongly enriched cellular component (adjusted p = 2.4 × 10⁻²⁴; Supplementary Figure S2). The 12 proteins most strongly more abundant in cluster 2 were all mitochondrial, including NFU1, MECR, DECR1, LONP1, LETM1, BAX, ACOT13, AK2, AIFM1, METAP1D and MMAB. In contrast, the 12 proteins more abundant in cluster 1 were mainly secreted proteins involved in normal homeostasis: BPIFA2, BPIFB1, ApoA1, CA6, SHBG, PON3 and SCGB3A1, of which ApoA1 and PON3 are known negative acute-phase reactants. Pathway enrichment (Fig. 3B–E) characterized cluster 2 by three themes: mitochondrial localization and metabolism (gene ontology cellular component (GO CC): mitochondrion and mitochondrial matrix as the most enriched cellular locations); innate immune activation and host response (Reactome: toll-like receptor (TLR) cascades, cellular responses to stimuli and stress, and positive regulation of tumor necrosis factor (TNF) production); and cellular reorganization and trafficking (gene ontology molecular function (GO MF): RNA binding, protein–macromolecule adaptor activity, and cytoskeletal/actin binding). This higher abundance was of similar magnitude in both sexes (Supplementary Figure S3).

Fig. 3.

Differential abundance and pathway enrichment across the complete Olink panel, adjusted for severity and sex. (A) Volcano plot showing differentially abundant circulating proteins in cluster 2 as compared to cluster 1, derived from linear models adjusted for age, sex, SOFA, NEWS, and lactate. Proteins with a significant increase are shown in red and proteins with a significant decrease are shown in blue (adjusted p-value < 0.05, |log2 fold change| >0.1); mitochondrial (Olink × MitoCarta3.0) proteins are outlined as black-bordered circles and the most strongly differential proteins are labeled. (B) Reactome pathway analysis and (C-E) Gene Ontology (GO) enrichment analysis of the adjusted-significant proteins for (C) Biological Process (BP), (D) Molecular Function (MF), and (E) Cellular Component (CC) terms. Dot size indicates the number of genes and dot color indicates the adjusted p-value

Within the 232 mitochondrial-associated proteins, severity- and sex-adjusted differential abundance identified the majority as more abundant in cluster 2; the 30 most differential proteins and their pathway enrichment are shown descriptively in Supplementary Figure S4. Mitochondrial biogenesis and energy-metabolism terms were enriched whereas mitophagy-related terms were not, a differential pattern not predetermined by the clustering and relevant to the candidate release mechanisms discussed below.

To identify a sparse set of proteins most strongly discriminating the clusters, LASSO-penalized logistic regression was applied to the 232 mitochondrial-associated proteins with cluster membership as the outcome (Fig. 4). This is interpreted as post-hoc feature reduction rather than independent prediction, as the clusters were derived from the same protein set. DECR1 was the strongest discriminator (coefficient 1.27), a finding that should be interpreted in the context of this being a single-platform, single-cohort derivation. Bootstrap stability and permutation testing confirmed the signature was not driven by overfitting (MRPS27, NGRN, DECR1 and GUK1 selected in > 80% of resamples; permutation p < 0.001). The discriminating proteins span processes previously implicated in mitochondrial biology in sepsis: fatty acid β-oxidation (DECR1, ACOT13), mitochondrial quality control and biogenesis (MRPS27, NGRN, FIS1), apoptosis and cellular injury (BAX, GADD45GIP1), and oxidative stress (SOD2).

Fig. 4.

Sparse, internally validated mitochondrial-protein signature discriminating cluster 1 and cluster 2. (A) LASSO regression coefficients of the proteins retained in the sparse signature, expressed as log-odds for cluster 2 membership and ordered by effect size. (B) Bootstrap stability of protein selection, showing the frequency with which each protein was selected across 1,000 resamples; dashed lines mark the 0.5 and 0.8 selection-frequency thresholds. (C) Permutation analysis showing the distribution of cross-validated area under the curve (CV-AUC) values obtained under 1,000 randomly permuted cluster labels (grey histogram), with the observed CV-AUC of the true model indicated by the red line and the corresponding permutation p-value

To assess whether the mitochondrial signal was disproportionate relative to generalized intracellular protein release, all measured proteins were assigned to subcellular compartments and their between-cluster differences compared (Fig. 5A). Proteins from all intracellular compartments were more abundant in cluster 2, but the difference was greatest for mitochondrial-associated proteins (Cohen’s d 2.57, versus 2.30 cytosolic, 2.15 nuclear and 1.82 intracellular-membrane), and mitochondrial-associated proteins were the most prominent component among the most abundant proteins across the full panel (4.9-fold, p ≈ 1 × 10⁻²⁵; Fig. 5C). The RMES, isolating the mitochondrial abundance beyond generalized intracellular protein release, remained higher in cluster 2 (d = 0.61, p < 0.001; Fig. 5B), indicating that the mitochondrial component is disproportionate rather than a simple reflection of overall cell injury. The RMES did not, however, track illness severity or short-term outcomes (all Spearman ρ ≤ 0.11; all OR CIs crossing 1). Given the strong correlation between the mitochondrial and non-mitochondrial intracellular scores (r = 0.91), residualization removes approximately 82% of the variance in the mitochondrial score, including the variance shared with illness severity; a null association is therefore expected by construction. The RMES is presented as a specificity measure only, not as a prognostic variable.

Fig. 5.

Specificity of the mitochondrial-protein signal distinguishing cluster 1 and cluster 2. (A) Subcellular-compartment analysis showing the standardized between-cluster difference (Cohen’s d, cluster 2 vs. cluster 1) in mean protein abundance for each compartment, after assigning every measured protein to the mitochondrial (Olink × MitoCarta3.0, n = 232), cytosolic (n = 901), nuclear (n = 1,500), or intracellular-membrane (n = 814) compartment using the HPA; 1,969 proteins annotated as secreted, plasma-membrane, or without a main-location annotation were excluded. Horizontal bars denote 95% confidence intervals and the mitochondrial compartment is highlighted in red. (B) Residual mitochondrial enrichment score (RMES) per cluster, defined as the mitochondrial composite score after regressing out the non-mitochondrial composite score, thereby isolating the mitochondrial abundance beyond generalized intracellular protein release; the standardized difference (Cohen’s d) and Wilcoxon rank-sum p-value are shown (cluster 1, black circles; cluster 2, grey triangles). The RMES is presented as a measure of the specificity of the mitochondrial signal and is not intended as a prognostic variable. (C) Density distribution of the per-protein cluster 2 vs. cluster 1 effect size (Cohen’s d) for mitochondrial (red) versus non-mitochondrial (grey) proteins, showing that mitochondrial-associated proteins are the most prominent component among the proteins with the highest abundance in cluster 2

Discussion

In this study, higher plasma abundance of mitochondrial-associated proteins was consistently associated with greater illness severity and worse short-term outcomes in patients presenting to the ED with suspected sepsis. This relationship was evident across a composite mitochondrial protein score and multiple clinical measures, suggesting that mitochondrial associated protein profiles capture clinically relevant variation early in the course of sepsis. Exploratory PCA and K-means clustering identified two groups that differed primarily in overall mitochondrial-associated protein abundance; however, the considerable overlap between clusters and limited support for a discrete two-cluster structure suggest that these groups may represent different positions along a biological continuum rather than clearly distinct mitochondrial subphenotypes.

Patients at the higher end of this mitochondrial protein continuum had greater illness severity at presentation, were more frequently diagnosed with sepsis, and experienced higher rates of ICU admission and mortality and lower organ failure-free survival. This continuous relationship resembles that of lactate, which increases across a spectrum of physiological disturbance and is strongly associated with sepsis severity and short-term outcome [3, 4] In the group with higher plasma abundance of mitochondrial-associated proteins exhibited worse vital signs upon presentation, were more likely to have sepsis, and had poorer 72-hour outcomes, including significantly higher ICU admissions, increased mortality, and reduced organ failure-free survival.

Differences in mitochondrial-associated protein abundance may reflect variation in the magnitude of mitochondrial and cellular stress during acute infection. Patients with higher abundance also had a higher BMI and more frequently had pre-existing heart failure, conditions associated with mitochondrial stress, altered substrate metabolism, and reduced bioenergetic reserve [11, 24–26]. These factors may increase susceptibility to mitochondrial dysfunction during sepsis, although the present study cannot distinguish pre-existing mitochondrial vulnerability from changes induced by acute illness. Exploratory differential-abundance, LASSO and pathway analysis further identified proteins and processes related to energy metabolism (including DECR1, ACOT13, MECR), mitochondrial integrity and quality control (such as LONP1, LETM1, NFU1, MRPS27, NGRN, FIS1), nucleotide metabolism and oxidative stress (including AK2, GUK1, SOD2), and apoptosis and cellular injury (like BAX, AIFM1, GADD45GIP1) [9, 11, 12, 15, 27]. These findings support the biological relevance of the observed mitochondrial protein signal but should be considered hypothesis-generating.

The mechanisms underlying increased plasma abundance of mitochondrial-associated proteins remain uncertain. Mitochondrial components may enter the circulation through cellular injury and regulated cell death, extracellular-vesicle release, mitochondrial-derived particles, or other processes involved in mitochondrial turnover and intercellular signaling [28–31]. Once extracellular, mitochondrial components may also function as damage-associated molecular patterns (DAMPs) and amplify innate immune activation, providing a potential link with TLR- and TNF-related pathways observed in the broader proteomic analysis [28–32]. The enrichment of pathways related to mitochondrial biogenesis and energy metabolism, without corresponding enrichment of mitophagy-related pathways, may indicate altered mitochondrial homeostasis; however, the present data cannot distinguish changes in mitochondrial production, release, or clearance. Given the observational and cross-sectional nature of this human study and the absence of direct mechanistic experiments, these interpretations remain hypothesis-generating.

Our findings also have implications for biological phenotyping in sepsis. Previous studies have identified subphenotypes based on transcriptomic and broader host-response profiles [33, 34], whereas we specifically investigated variation in mitochondrial-associated proteins. The limited separation between clusters in our study suggests that mitochondrial biology may not readily divide patients into distinct categories at a single time point. Instead, mitochondrial-associated protein profiles may represent a graded biological response that varies across patients. Whether clinically meaningful mitochondrial subphenotypes exist within this continuum may depend on additional dimensions, such as the temporal trajectory of mitochondrial responses, pre-existing metabolic vulnerability, or the underlying type and severity of infection. Longitudinal studies are therefore needed to determine whether mitochondrial-associated protein profiles change dynamically during sepsis and whether these trajectories provide information beyond established markers of illness severity.

Pathway enrichment analysis using the complete Olink panel placed the circulating mitochondrial-associated protein signal within a wider cellular and immune framework. This includes cellular responses to stress and stimuli, infectious diseases, TLR cascades, viral infection signaling, intracellular and protein transport, and TNF production [15]. GO CC enrichment highlighted mitochondrion and mitochondrial matrix as the most enriched cellular locations. The proteins elevated in the subgroup of patients with relatively low mitochondrial-protein abundance (APOA1, SHBG, PON3, BPIFA2, BPIFB1, CA6) are mainly secretory plasma proteins involved in homeostasis. Several are negative acute-phase reactants, which usually decrease during systemic inflammation; their relative stability supports a less inflamed state [35]. Overall, these findings suggest that the circulating mitochondrial-associated protein response in sepsis reflects a coordinated process combining mitochondrial biology, innate immune activation, and broader cellular stress, rather than a purely metabolic one [14, 15].

This study has several limitations. Plasma mitochondrial-associated protein abundance is an indirect measure of mitochondrial biology and does not directly reflect mitochondrial mass or function [22, 36]; however, plasma profiling offers a clinically accessible window into mitochondrial-associated processes at the time of ED presentation. The exploratory clusters showed substantial overlap and should therefore not be interpreted as discrete mitochondrial subphenotypes, although this finding itself supports the concept of a continuous mitochondrial response across patients. Olink provides relative rather than absolute protein abundance, and external validation in an independent cohort was not available. The relatively small number of ICU admissions and deaths limited the precision of prognostic analysis. Finally, this was a post-hoc, observational, single-center study, which limits causal inference and generalizability but allowed detailed phenotyping and high-dimensional proteomic characterization early in the clinical course.

Future studies should validate the continuous mitochondrial protein signal in independent more heterogeneous cohorts and examine its evolution through longitudinal sampling. Combining plasma proteomics with direct measures of mitochondrial function and targeted quantitative assays may further clarify its biological significance and determine whether mitochondrial protein profiles provide information beyond established markers or identify clinically relevant subphenotypes within this continuum.

Conclusion

Higher plasma mitochondrial-associated protein abundance was associated with greater illness severity and worse short-term outcomes in patients with suspected sepsis. Exploratory clustering suggested that mitochondrial-associated protein profiles vary predominantly along a continuum rather than forming clearly distinct mitochondrial subphenotypes. These findings support further investigation of mitochondrial-associated protein profiles as biomarkers of sepsis severity and may inform future studies of mitochondrial-targeted approaches in sepsis.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ATP

Adenosine triphosphate

BMI

Body mass index

COPD

Chronic obstructive pulmonary disease

CV-AUC

Cross-validated area under the curve

DAA

Differential abundance analysis

DAMPs

Damage-associated molecular patterns

ED

Emergency department

EHR

Electronic health record

GO

Gene Ontology

GO BP

Gene Ontology Biological Process

GO CC

Gene Ontology Cellular Component

GO MF

Gene Ontology Molecular Function

HPA

Human protein atlas

ICU

Intensive care unit

IQR

Interquartile range

LASSO

Least absolute shrinkage and selection operator

NEWS

National early warning score

NPX

Normalized Protein eXpression

PCA

Principal component analysis

RMES

Residual mitochondrial enrichment score

SD

Standard deviation

SGLT2-inhibitor

Sodium-glucose cotransporter 2 inhibitor

SOFA

Sequential organ failure assessment

STROBE

Strengthening the Reporting of Observational Studies in Epidemiology

TLR

Toll-like receptor

TNF

Tumor necrosis factor

UMCG

University Medical Center Groningen

Author contributions

CJM: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing (original draft), Visualization.  FEvB: Investigation, Writing (review & editing). BAC: Investigation, Writing (review & editing). JL: Investigation, Data curation, Writing (review & editing). STH: Conceptualization, Methodology, Data curation, Writing (review & editing), Supervision. HRB: Conceptualization, Methodology, Supervision, Writing (review & editing).

Funding

Not applicable.

Data availability

Data is available upon reasonable request from the corresponding author.

Declarations

Ethics approval and consent to participate

Acutelines is approved by the medical ethics board of the UMCG and registered at ClinicalTrials.gov under trial registration number NCT04615065 on 2020-09-01 [17]. The study protocol was approved by the institutional review board (IRB) of the University Medical Center Groningen. A deferred consent procedure (by proxy) was in place to enable the collection of data and biomaterials before written consent was obtained.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.