Multiomic analyses of longitudinal plasma samples identify thromboinflammation endotypes and trajectories in patients with trauma

Abstract

Understanding the complexity of trauma-induced thromboinflammation necessitates data-driven approaches. We hypothesized that longitudinal plasma profiling could reveal underlying differences in patients with injury who present with similar clinical characteristics but ultimately have different outcomes. Here, we performed multiomic analyses of longitudinal plasma samples from a clinical trial of patients with traumatic injury to identify molecular endotypes and trajectories that were associated with patient outcomes. The pathophysiologic states of patients with trauma were defined by the longitudinal proteomic and metabolomic plasma profiles from a diverse cohort. Then, patients were endotyped according to their longitudinal trajectories through trauma omic states, and injury patterns and outcomes were compared. We identified endotypes associated with divergent clinical outcomes despite similar injury patterns at presentation. Organ failure and time spent in the intensive care unit (ICU) were predicted with high accuracy using omic markers. Patients who presented with evidence of elevated proteasome activation, catabolism, and superoxide formation were vulnerable to heart failure, lung failure, and acute lung injury, respectively. In addition, omic markers of increased hypoxia, RBC lysis, and hydrolase activity better fit mortality and ICU time compared with injury covariates, while providing biological insight. Injury and outcome patterns persisted in a validation trauma cohort after endotype assignment at a single, early time point. These data align with the understanding that patients with trauma may experience markedly different biological responses and outcomes despite similar clinical presentations. We suggest that mapping patient trajectories through biological injury states could provide a framework for personalized patient treatment after trauma.

Editor’s summary

Traumatic injury is complicated by thromboinflammation, a combination of coagulation, inflammation, and immune dysregulation, that leads to substantial morbidity and mortality but is poorly understood. Here, Cohen and colleagues performed multiomic analyses of longitudinal samples from the Control of Major Bleeding After Trauma clinical trial to identify molecularly defined thromboinflammatory endotypes associated with divergent outcomes despite similar injury patterns at presentation. By tracking patient trajectories through the omic trauma space, the authors could predict clinical outcomes like organ failure with high accuracy. These findings suggest that mapping patient trajectories through biological injury states could potentially assist in the personalized treatment of patients after trauma as well as provide insights into the pathophysiology of thromboinflammation. —Melissa L. Norton

INTRODUCTION

Trauma and associated hemorrhage remain a leading cause of death worldwide (1, 2). Resuscitation and care of individual patients is confounded by variation in both response to injury and in final outcomes such as mortality and tissue dysfunction across populations of patients with trauma (3–5). Injury patterns such as degree of sustained tissue injury or shock severity, demographics, and preexisting conditions explain some of the variation in patient response to trauma and their outcomes (6–9). However, patients with similar injury patterns may follow different biological trajectories toward recovery or may have divergent outcomes despite having similar injuries. Further complicating care is the degree to which a patient may or may not respond to or develop complications from hemostatic treatments (10, 11). For example, blood products transfused during early treatment intended to reverse shock and resuscitate patients are independently associated with their own complications, including dysregulated thromboinflammation, organ failure, and even mortality (10, 12, 13). This complex physiologic and biologic milieu manifests at the bedside as the all too common clinical conundrum where two seemingly identical patients have markedly divergent responses to injury and treatment and, therefore, outcomes.

Thromboinflammation, a poorly characterized combination of coagulation, inflammation, and immune dysregulation, underlies mortality in patients with trauma and morbidity in survivors and is highly variable (11, 14). Traumatic injury releases damage-associated molecular patterns (DAMPs) including tissue components, nucleic acids, proteins, and metabolites into circulation, which activate an inflammatory response that is synergistically compounded by shock and subsequent reperfusion (11, 15–17). Even when clinically resuscitated and stabilized, patients suffer from sustained endothelial injury and increased immune signaling and inflammation that can contribute to organ failure and late mortality (10, 18–20). In addition, these patients are at risk of infection, in part because of a dysregulated inflammatory state that exhibits an attenuated response to invading pathogens (11, 15). Compounding this complexity is traumatic brain injury that has its own thromboinflammatory signaling (21). It is clear that thromboinflammation is central to the biology of trauma, and it has been the focus of much research (22–25).

Current clinical practices are largely informed by univariate mediators that lack comprehensive phenotyping of complex trauma pathobiology and fail to offer definitive guidance in patient care (26–30). Even large, multicenter transfusion- and treatment-focused clinical trials have failed to provide sufficient biological insight and treatment guidance to improve patient care (28, 29, 31–34). However, post hoc analyses of these trials identified patient endotypes with specific inflammatory profiles that responded to treatment (35–39). This suggests that clinical decision-making that relies largely on clinical experience could benefit from a better understanding of postinjury pathology that links underlying biological dimensionality with treatment response and outcomes (40–43). Although this approach has revolutionized patient care in other fields, it has yet to provide breakthroughs in trauma medicine (4, 35, 44, 45). High-dimensional profiling of tissue injury and hemorrhagic shock biomarkers offers a promising approach to establish trauma-specific endotypes by resolving the nonlinear trajectories of these markers that drive patient outcomes.

Previously, our group conducted a randomized trial comparing prehospital plasma with saline control for resuscitation in patients with hemorrhagic shock due to trauma in the Control of Major Bleeding After Trauma (COMBAT) trial, which was terminated early because of a lack of difference in 28-day mortality between treatment arms (30). This afforded a unique opportunity for controlled longitudinal blood sampling with matched clinical care and outcome data from patients with trauma. We hypothesized that multiomic plasma profiling could identify thromboinflammatory endotypes and distinct biological trajectories that are associated with differences in outcome measures. Here, we used unsupervised clustering of metabolomic and proteomic data obtained from longitudinal patient plasma samples to characterize molecular signatures after trauma and to determine molecularly defined trauma trajectories and endotypes.

RESULTS

Characteristics of the study patients and data analysis workflow

A total of 97 healthy control participants and 118 patients who met enrollment criteria of severe injury by mechanism or physiologic criteria were enrolled in this study (Fig. 1A). For patients, blood was collected at specified time points [field, emergency department arrival (ED), and postinjury at hours 2, 4, 6, 12, 24, 48, 72, 120, and 168] or until the patient’s discharge (Fig. 1B), for a total of 856 samples (759 patient samples and 97 healthy control samples from a blood donation center). From patient plasma, 1012 proteins and 472 metabolites were quantified using untargeted proteomics and metabolomics by liquid-chromatography coupled with tandem mass spectrometry (LC-MS/MS). In addition, standard clinical measurements, blood labs (blood count, chemistry, and hematology), viscoelastic hemostatic parameters, and outcomes were recorded for all patients throughout the study (Fig. 1C).

Fig. 1. Plasma omics and clinical measurements were obtained from a diverse population of patients with trauma with varying injuries and outcomes.

(A) Schematic showing overall timeline and number of plasma samples collected per time point. (B) Bar graph of patient enrollment per time point and reasons for discharge. (C) Number of parameters per assay measured. (D) Patient characteristics and outcomes per shock and tissue injury group (LSLT, LSHT, HSLT, and HSHT) showing median values with IQR (one-way ANOVA with Tukey’s post hoc test) or percent (χ² proportion test) *P < 0.05. (E) Scatterplot of tissue injury and shock; diamonds indicate deaths. (F) Schematic of unsupervised omics-based patient endotyping. Trajectories were defined by cluster assignment upon ED arrival and sequence length. Then trauma severity [NISS, BE (base excess), and INR] and clinical outcomes were compared to identify differences in clinical outcomes (mortality, organ failure, ICU time, and ventilator time) among patients with similar injuries (*red star) versus outcome differences due to trauma severity (*black star). Validation was performed on a separate trauma plasma dataset (TAP), where TAP trajectories were assigned using a COMBAT-trained model followed by injury and outcome comparison. HC, healthy control; BMI, body mass index; M/F, male/female; TBI, traumatic brain injury; HR BPM, heart rate beats per minute; SBP, systolic blood pressure; HGB, hemoglobin; MTX, massive transfusion; PS, patient state; T, trajectory.

Patient characteristics for the COMBAT study have been previously detailed (30). Here, they are first presented by tissue injury and shock severity to determine clinical and outcome differences stratified by these groups (Fig. 1D). Patients were grouped according to their shock severity [base excess at ED < −10: high shock (HS)] and new injury severity score [NISS >25: high trauma (HT)] into HSHT, high shock–low trauma (HSLT), low shock–high trauma (LSHT), and low shock–low trauma (LSLT) (Fig. 1E). For 20 patients with missing base excess at ED, omic data from 333 patients from the trauma activation protocol (TAP) cohort were used to train and test an imputation model for value assignment (fig. S1). As expected, patients with more severe injuries had worse outcomes. Patients with higher NISS and lower base excess had higher needs for massive transfusion [>10 units of red blood cells (RBCs) within 24 hours], longer hospital stay, fewer ICU- and ventilator-free days, and a higher incidence of mortality and lung organ failure among the injury severity groups (Fig. 1D).

We first determined the longitudinal omic trends underlying these differences stratified by injury severity. Then, we investigated how underlying biological states were associated with thromboinflammatory outcomes using unsupervised approaches (Fig. 1F). Omic patient states were created to identify groups of patients with similar biological trajectories, and then clinical measurements and outcomes among these endotypes were compared. These results were then validated in the TAP dataset, which was quantified upon ED arrival only.

Longitudinal omics and clinical trends of the “average” patient with trauma

The longitudinal trends of plasma omic markers after trauma were clustered using C-means to identify groups of proteins and metabolites with similar expression patterns. For metabolomics, there was early activation of the tricarboxylic acid cycle and arginine biosynthesis and sustained amino acid and thiamine metabolism (Fig. 2A). Hypotaurine metabolism peaked at 2 to 6 hours, and sphingolipid and histidine metabolism remained elevated until 24 hours. From 24 to 48 hours, increases in nucleic acids, glutathione, hormone, and fatty acid biosynthesis were observed, with bile acid and proteolysis metabolites peaking 5 to 7 days postinjury. Analysis of plasma proteomic markers revealed high expression of detoxification, fluid management, and neutrophil/platelet degranulation proteins in the first 6 hours after trauma. This was followed by activation of immune responses, complement activation, and coagulation processes up to about 24 hours (Fig. 2B). Accompanying this were elevations in proteins involved in lipid remodeling and metabolism and antimicrobial activity. Between 24 and 48 hours, there was a second spike in neutrophil degranulation and glycolysis proteins, followed by a strong peak of proteins involved in the acute phase response (APR), leukocyte activity, complement regulation, and hypoxia.

Fig. 2. C-means clustering reflects metabolomics, proteomics, and clinical trends of the “average” patient with trauma.

(A) Heatmap of median values of metabolites sharing kinetic patterns (left side) and enrichment of each cluster (right side, scores from MetaboAnalyst). (B) Heatmap of median values of proteins sharing kinetic patterns and Metascape Gene Ontology (GO) terms of each cluster. (C) Clinical laboratory tests showing median values at each time point. (D and E) Strength of association between clinical measurements and associated metabolites (D) and proteins (E), with line thickness showing relative partial slope. cn, citrated native; pa_CO₂, partial pressure of arterial carbon dioxide; HCT, hematocrit; Plt, platelet; DBP, diastolic blood pressure; cryo, cryprecipitate; WBC, white blood cell; k, kinetics time; Bicarb, bicarbonate; VCO₂, carbon dioxide production.

Patient vitals and blood hematology, chemistry, and count data were also clustered by their longitudinal patterns. There were early increases in heart rate, lactate, and base excess, alongside decreased blood pressure, which normalized postresuscitation (Fig. 2C). Patients presented with coagulopathy [low maximum amplitude (MA) and fibrinogen, elevated clot lysis at 30 min (LY30)] and initial hypocalcemia that worsened within 2 hours before calcium administration normalized ionized calcium. Linear mixed modeling identified associations between omic data and clinical metrics: Carnitines and phosphates correlated with platelet count, partial pressure of arterial oxygen (PaO₂), and fibrinogen, whereas tissue injury metabolites were primarily linked to clotting (Fig. 2D). Proteins involved in coagulation and complement tracked with clotting assays and platelet counts, and APR proteins correlated with blood gas measures (Fig. 2E).

Given that the COMBAT trial was terminated early because of a lack of difference in 28-day mortality between prehospital plasma and saline control arms (30), we next tested for injury and omic differences between these groups. There were no differences in ED arrival base excess (P = 0.92), NISS (P = 0.32), international normalized ratio (INR) (P = 0.86), sex (P = 0.38), age (P = 0.21), or body mass index (P = 0.07) between treatment groups (Student’s t test). Similarly, we found no differences in omic features between groups per time point, neither with all patients combined nor when separated by injury severity.

Metabolomic biological states of patients with trauma

Next, we sought to define the metabolomic, proteomic, and integrated multiomic biological states that underlie trauma endotypes and determine their ability to discriminate otherwise clinically similar patients. First, a uniform manifold approximation and projection (UMAP) embedding of metabolomic plasma data was generated and clustered into eight states that reflected different metabolomic states of patients with trauma (mPSs) (Fig. 3A). These mPSs were defined by signatures of elevated or decreased abundances of specific metabolites relative to other states (Fig. 3B and fig. S2; P values and enrichment results in tables S1 and S2). The different mPSs aligned with time since injury, where metabolic states were enriched at different time points, including acute injury (initial injury through about 6 hours; mPSs 1, 5, and 8), postinjury (about 6 to 24 hours postinjury; mPSs 2, 3, and 4), and recovery (>24 hours postinjury; mPSs 6 and 7) (Fig. 3C).

Fig. 3. Metabolomic-based clustering identifies eight PSs linked to acute injury, postinjury, and recovery.

(A) mPS UMAP colored by PS. (B) Top five metabolites representative of each PS. (C) mPS enrichment and density of data points per time point. (D) Number of data points, median (IQR) of clinical values per PS, and statistical comparison (one-way ANOVA with Tukey’s post hoc test). Time series graph of (E) hematocrit (HCT), (F) platelet count, (G) clotting factor 7 (F7), (H) clotting factor 8 (F8), and (I) lysis percentage of clot 30 min after max amplitude (MA) (LY30) (linear mixed model with Šidák’s multiple comparisons test). *P < 0.05. See fig. S4 for full comparisons of clinical laboratory tests. ETC, electron transport chain; AA, amino acids; FA, fatty acids; m, metabolism; bs, biosynthesis; TH, thyroid hormone.

After the generation of mPSs, clinical and outcome data were overlaid to visualize their distributions across the trauma metabolomic map (fig. S2). Data points associated with death, low ICU-free days, and severe injury (HSHT) were dispersed across several mPSs, demonstrating the heterogeneous nature of metabolic profiles across patients after trauma (mPSs 3, 5, and 8; fig. S2). Classification of mPSs using cross-validation (CV) ensemble learning within a nested framework differentiated between mPS classes with 87.5% overall accuracy and identified 42 noncorrelated significant metabolites (fig. S3).

Two main metabolic meta-states were generally enriched for catabolism and tissue injury (mPSs 1 to 4) and anabolism including energy production and biosynthesis (mPSs 5 to 8) (Fig. 3A and fig. S2). For anabolic states, mPS1 was enriched for key energy production pathways including the citric acid cycle, whereas mPS2 was enriched for protein biosynthesis (alanine and cysteine metabolism), acetyl–coenzyme A generation from various macromolecules (thiamine and pantothenate metabolism), and redox homeostasis via riboflavin metabolism. mPS3 was enriched for purine and amino acid metabolism, autophagy (trehalose degradation), and anabolic glycolysis (pentose phosphate), and mPS4 was enriched for bile, antioxidant, and neurotransmitter (glutamate) synthesis. For catabolic states, mPS5 was enriched for the arachidonic acid cycle and thyroid hormone synthesis. mPSs 6 and 8 were both enriched for methylhistidine (skeletal muscle breakdown) and tryptophan (kynurenine pathway) metabolism and for carnitine (fatty acid pathway), phosphatidylcholine (cell membrane), and adrenal catecholamine biosynthesis. However, they reflected different timescales, with mPSs 8 and 6 reflecting early and later injury, respectively. Similar to mPS2, mPS7 was enriched for thiamine, alanine, and riboflavin metabolism although with a different metabolic signature.

mPSs correlated with static and temporal clinical profiles (Fig. 3, D to I, fig. S4, and tables S3 and S4). mPS8 exhibited the most perturbed vital signs (lowest body temperature and blood pressure), reflected the most shock (lowest base excess and bicarbonate and highest lactate), and was the most coagulopathic (highest INR and lowest clotting factors). mPS8 was further characterized by persistent brain injury [low Glasgow coma scale (GCS)] and a longitudinal fibrinogenemic and thrombocytopenic coagulopathy. Other notable differences included those in blood chemistry (low P_aO₂ in mPS4; low pH in mPSs 3 and 8), blood count (high white blood cells, platelets, hemoglobin, and hematocrit in mPSs 1, 2, and 5), clotting [lowest PTT (partial thromboplastin time), INR, D-dimer, and von Clauss fibrinogen in mPS5], and transfusion of blood products [high tranexamic acid in mPS4; high RBCs, fresh frozen plasma (FFP), and crystalloids in mPSs 4 and 8]. Several mPSs were clinically similar reflecting our hypothesis that metabolomic differences would distinguish otherwise clinically indistinguishable states.

Patient trajectories through metabolomic trauma space

Because the clinical situations and physiological and biological processes in patients with trauma are dynamic, we next grouped them by their ED starting mPS and sequence lengths (number of time points from ED to final time point) into seven trajectories, which took different paths through metabolomic trauma space (Fig. 4A). Each trajectory (T) was associated with a distinct longitudinal metabolomic signature, defined by their path through mPSs (Fig. 4, B and C, and table S5). Patients who began with antioxidant (mT3), catecholamine (mT5), or polyamine (mT7) biosynthesis remained strictly in these metabolomic states, whereas other mTs showed increased mPS diversity over time (fig. S5). mT classification using ensemble learning in a nested CV framework was performed at 86.6% overall accuracy (fig.S5). Further, testing of mPS trajectory and sequence length found significant divergence among all trajectories by permutation testing of their dynamic time warping (DTW) distances (P < 0.05; table S6).

Fig. 4. Metabolomic trajectories reveal pathways related to outcomes.

(A) Metabolomic trauma trajectory UMAP, colored by trajectory. (B) Density plot of data points per time point and number of patients (N) per trajectory. (C) Two-dimensional (2D) plot of patient trajectories through mPSs (y axis) over time (x axis) ± 95% CI (confidence interval). (D to F) Injury severity median ± IQR of (D) NISS, (E) INR, and (F) BE upon ED arrival of trajectories (one-way ANOVA with Tukey’s post hoc test). Proportions of (G) mortality, (H) heart failure, (I) lung failure, (J) ALI per trajectory (Fisher’s exact test; *red asterisk indicates significant differences among similarly injured trajectories, and *black asterisk indicates significant differences among trajectories with different injury patterns). (K) Median ± IQR ICU-free days per trajectory (one-way ANOVA with Tukey’s post hoc test). AUROC ± 95% CI metabolomic classification curves using metabolomics versus injury covariates of (L) heart failure (n = 9/75 Y/N), (M) lung failure (n = 19/53 Y/N), and (N) ALI (n = 17/15 Y/N). (O) Regression of ICU-free days among patients with similar injuries (n = 66). (P) NISS, (Q) INR, and (R) BE median ± IQR of the TAP validation dataset after trajectory assignment using the COMBAT-trained model. (S and T) Classification accuracy of (S) ICU-free days and (T) ventilator-free days in the TAP validation set using COMBAT-trained models. *P < 0.05, **P < 0.01, and ***P < 0.001.

Patients in mT3 and mT7 had higher NISS and INR (P < 0.05), and there were no significant differences in base excess (Fig. 4, D to F). mT7 had higher mortality than other trajectories (P < 0.05 versus mTs 1 and 3 to 6), which coincided with its more severe injury patterns (Fig. 4G). Although patients in mT6 and mT7 began in the same state, metabolomics distinguished between these patients at higher rates [area under the receiver operating characteristic curve (AUROC) = 0.97] than injury covariates (AUROC = 0.65, P = 0.035). Between patients in these trajectories, mT6 had elevated l-glutamate and sphinganine and decreased citrate and carnitines at ED as compared with patients in mT7 (P < 0.05) (fig. S5).

Beyond mortality, there were differences in clinical outcomes that were not explained by injury patterns. mT6 had significantly higher incidences of acute lung injury (ALI) and lung and heart failure among patients with similar injuries, and mT3 had higher incidence of lung failure than mT2 (P < 0.05; Fig. 4, H, I, and J). mT6 was predominately Hispanic or Latino (P < 0.05), but there were no other significant differences in preexisting conditions or morbidities among trajectories (fig. S5). There were also differences in ICU-free and ventilator-free days among patients with similar injuries, with mT6 having fewer days than mT5 and mT1 (P < 0.05; Fig. 4K and fig. S6).

Among patients with similar injuries, penalized logistic regression determined a significant increase in log odds of heart failure in mT6 compared with other trajectories (mTs 1, 2, 4, and 5) (P = 0.0075) (fig. S7). The classification accuracy of heart failure among patients with similar injuries in these trajectories was higher using metabolomics (AUROC = 0.96) than using injury covariates (AUROC = 0.57, P < 0.0001) (Fig. 4L and fig. S7). Those patients who developed heart failure had elevated psychosine (a cytotoxic sphingolipid), α-tocopherol-succinate (a form of vitamin E), maltotriose (an oligosaccharide), and hydroxy-kynurenine (fig. S7). A similar analysis revealed that mT6 was positively and significantly associated with an increased likelihood of lung failure (P = 0.0036) (fig. S8). Patients who developed lung failure had higher expression of psychosine, gerberinol, and glutamyl-taurine than those who did not (Fig. 4M and fig. S8). mT6 was also associated with increased likelihood of ALI versus T5 (P = 0.005), with elevations in prostaglandin E2, sphinganine, and ferric gluconate and decreases in tryptophan, picolinic acid, and decanoic acid (fig. S9). Classification of ALI was improved using metabolites (AUROC = 0.93) over injury covariates (AUROC = 0.65, P = 0.017) (Fig. 4N and fig. S9).

Using the continuous variable nature of ICU time, we then compared regression analysis results using clinical covariates versus metabolites. Regression analysis of ICU-free days yielded similar performance using injury covariates (R² = 0.42) or metabolites (R² = 0.57 and P = 0.173), with hydroxykynurenine, glycerol-3-phosphate, and spermidine contributing strongly to regression (Fig. 4O and fig. S6). Similar results were observed for ventilator-free days with metabolomics (R² = 0.503) and injury covariates (R² = 0.342) regressing with equal performance (P = 0.172) and identifying many of the same metabolites that contributed to ICU-free days (fig. S6).

For the validation dataset, plasma omic data from TAP were acquired with orthogonal methods (results were generated under slightly different LC-MS/MS conditions, resulting in metabolite identifications with about 50% overlap with the COMBAT dataset). After data transformation (using Yeo-Johnson), and despite missingness of metabolites peripheral to our model, trajectories were assigned to patients in the TAP dataset using a COMBAT-trained model (Fig. 1F). Injury patterns of well-represented trajectories (mTs 1, 2, 4, and 6) were similar to COMBAT injury patterns with no significant differences in NISS or base excess and with mT2 having high INR (Fig. 4, P to R). There were no differences in ICU- or ventilator-free days, but patients who died were conserved to mTs 1, 2, and 6 (fig. S10). The classification of the highest versus lowest quartile ICU- and ventilator-free days in TAP using a COMBAT-trained LASSO model yielded overall accuracies of 0.82 and 0.74, respectively, with highest quartiles predicted at higher accuracies than the lowest quartile (Fig. 4, S and T). Other outcomes were unavailable for comparison.

Proteomic biological states of patients with trauma

Dimension reduction with UMAP followed by hierarchical clustering was again used to create proteomic patient states (pPSs) (Fig. 5A and fig. S11). In proteomic trauma space, data points reflective of severe injury and mortality were localized (pPS1) and defined by high concentrations of both intracellular and extracellular DAMPs and metabolic derangement by ED arrival (Fig. 5, B and C, fig. S11, and table S7). pPS4 also represented severe injury, but at times from 2 to 12 hours after injury, and was enriched for glycolysis, gluconeogenesis, and high detoxification response (figs. S11 and S12 and table S7). Other pPSs were enriched for humoral immune response (pPS2) or clot formation and proteolysis (pPS6). pPSs 3, 5, and 7 from around 24 to 168 hours postinjury were enriched for various degrees of inflammatory response (APR, complement activation, platelet, and neutrophil degranulation) and coagulation. Classification using nested CV ensemble learning differentiated between pPSs with 86.7% overall accuracy and identified 82 uncorrelated significant features by permutation testing (fig. S12).

Fig. 5. Proteomic clustering reveals injury severity and recovery states with distinct tissue injury and coagulation profiles.

(A) pPS trauma UMAP colored by PS. (B) Top five proteins representative of each PS. (C) pPS enrichment and density of data points per time point. (D) Number of data points, median (IQR) of clinical values per pPS, and statistical comparison (one-way ANOVA with Tukey’s post hoc test). (E to I) Time series graphs of (E) platelet count, (F) clotting factor 5 (F5), (G) D-dimer, (H) INR, and (I) MA of clot formation (linear mixed model with Šidák’s multiple comparisons test). *P < 0.05. See fig. S13 for full comparisons of laboratory tests.

Clinically, pPSs differed predominantly in clotting components and dynamics and vital signs (Fig. 5, D to I, fig. S13, and tables S8 and S9). pPSs 1 and 4 were associated with the most perturbed vitals (lowest temperature, blood pressure, and GCS and highest heart rate) and blood chemistry (lowest VCO₂, base excess, and bicarbonate; high lactate). These states also showed evidence of coagulopathy and fibrinolysis [high PTT, INR, D-dimer, thromboelastography (TEG) parameters; low clotting factors (F2, F5, F8, and F9) and fibrinogen] and received high amounts of transfused blood products (RBCs, platelets, cryoprecipitate, and FFP) (fig. S13). pPSs 6 and 8 also followed these patterns, although they were not as pronounced. pPSs 2 (acute) and 3 (postinjury) were the least perturbed. Postinjury pPSs 5 and 7 showed normalization of vitals (raised temperature and normal SBP and HR) and lactate; robust fibrinogen production, clot strength (MA- and K-TEG), and formation rate (angle-TEG); and decreased P_aO₂.

Patient trajectories through proteomic trauma space

Similar to our metabolomic methodology, proteomic trajectories were assigned to patients by initial pPS and sequence length (Fig. 6A). Trajectories had distinct proteomic signatures and significant differences among all trajectories in PS vector and sequence length by permutation testing (P < 0.05; fig. S14, A and B, and tables S10 and S11). Further, nested CV ensemble learning classified trajectories at 85.5% accuracy (fig. S14, C and D). Although most trajectories followed narrow proteomic expression patterns, pT1 and pT5 had high within-group variance reflective of the proteomic change in these trajectories over time (fig. S14, A and B). This was evident by the convergence among these trajectories across multiple patient states toward a uniform postinjury recovery state (APR and hemostasis) (Fig. 6, B and C).

Fig. 6. Proteomic trajectories highlight recovery patterns.

(A) Proteomic trauma trajectory UMAP, colored by trajectory. (B) Density plot of data points per time point and number of patients per trajectory. (C) 2D plot of patient trajectories through pPS enrichment (y axis) over time (x axis) ± 95% CI. (D to F) Injury severity median ± IQR of (D) NISS, (E) INR, and (F) BE upon ED arrival (one-way ANOVA with Tukey’s post hoc test). (G to J) Proportions of (G) mortality, (H) lung failure, (I) heart failure, and (J) ALI per trajectory (Fisher’s exact test). (K) Median ± IQR ICU-free days per trajectory (one-way ANOVA with Tukey’s post hoc test; *red asterisk indicates significant differences among similarly injured trajectories, and *black asterisk indicates significant differences among trajectories with different injury patterns). (L to N) AUROC curves ± 95% CI of (L) lung failure (n = 26/74 Y/N), (M) heart failure (n = 10/90 Y/N), and (N) ALI (n = 46/54 Y/N). (O) Regression of ICU-free days among all patients with trauma (n = 100). (P to R) TAP validation dataset of (P) NISS, (Q) INR, and (R) BE after trajectory assignment using a COMBAT-trained model. (S and T) Prediction accuracy of (S) ICU-free days and (T) ventilator-free days in the TAP validation set using COMBAT-trained models. *P < 0.05, **P < 0.01, and ***P < 0.001. Degran, degranulation; ECM, extracellular matrix; IGF, insulin like growth factor; Comp, complement; Coag, Coagulation; ROS, reactive oxygen species; Detox, detoxification.

Patients in pT2 were most severely injured with significantly higher NISS and INR and lower base excess as compared with most other pTs (P < 0.05; Fig. 6, D to F). These patients also had the highest mortality rate (Fig. 6G). Patients in both pT1 and pT2 began in the same pPS of severe tissue injury and high metabolic stress (Fig. 6, A and C). Of these, pT2 had high mortality (90%) soon after injury, because injuries in these patients did not resolve. Patients in pT1 shifted toward postinjury recovery and had lower mortality (10%). Classification analysis between these trajectories found that survivors from high tissue injury pPSs 1 and 4 had decreased CD93 and elevated plasma proteins involved in fatty acid metabolism, oxidation, and wound healing (fig. S14G). In other trajectories, less injured patients arrived in a state of relatively mild humoral response and were discharged early (pT4) or shifted to APR and complement activation (pT3). pT5 (high tissue injury), pT6 (clotting and proteolysis), and pT7 (degranulation and extracellular proteins) began in various degrees of inflammation and coagulopathy before postinjury recovery.

There were significant differences in mortality, ALI, and lung and heart failure among pTs, but these could be explained by injury patterns with patients with more severe injuries having worse outcomes (Fig. 6, H to J). Among all pTs, however, there was improved classification using proteomics over injury covariates for lung failure (AUROC = 0.92 versus 0.70, P = 0.002), heart failure (AUROC = 0.75 versus 0.53, P = 0.006), and ALI (AUROC = 0.91 versus 0.70, P = 0.0008) (Fig. 6, L to N). Protein signatures contributing to these classifications included humoral and metabolic response to severe injury for lung failure; proteasome and complement activation, RBC lysis, and histone presence for heart failure; and detoxification, catabolic processes, and RBC gas exchange for ALI (fig. S15, A to C, and table S12).

Significant differences in ICU- and ventilator-free days could also be attributed to injury patterns (except pT5 versus pT7) (Fig. 6K and fig. S15, D to F). However, regression of ICU-free days was improved using proteomics (R² = 0.851) over injury covariates (R² = 0.462 and P < 0.001), with proteins involved in inflammatory, detoxification, and metabolic processes having high feature importance (Fig. 6O and fig. S15D). Similarly, proteins involved in coagulation, oxidation, and proteolysis improved regression of ventilator-free days (R² = 0.877) over injury covariates (R² = 0.396 and P < 0.001) (fig. S15, E and F).

As with metabolomics, for validation, trajectories were predicted and assigned to the single ED time point from the TAP validation dataset using a random forest model trained on the COMBAT dataset. Injury and outcome patterns were similar between the COMBAT and TAP validation sets, although some trajectories had low patient numbers or were not assigned (Fig. 6, P to R). In TAP, pTs 1 and 2 had relatively worse injury patterns and fewer ICU- and ventilator-free days (fig. S15, G to I). Classification of highest versus lowest quartile ICU- and ventilator-free days in TAP using a model trained on COMBAT yielded overall accuracies of 0.72 and 0.67, respectively, with the highest quartile predicted at higher accuracies than the lowest quartile (Fig. 5, S and T).

Multiomic biological states of patients with trauma

Having determined the underlying trauma biological states and trajectories from metabolomics and proteomics separately, we next integrated these layers to evaluate how multiomic patient states (oPSs) categorized trauma trajectories. As previously, oPSs were defined by enrichment of distinct multiomic marker signatures and time point densities (Fig. 7, A to C, and table S13). Data points reflective of mortality, low ICU-free days, and high injury severity were centralized and defined by metabolic derangement and high tissue damage (oPS1) (fig. S16 and table S13). Other early injury states were characterized by strong humoral response (IgG release, oPS3), APR and carnitine synthesis (FA metabolism, oPS5), regulation of coagulation and alpha-linolenic acid metabolism (oPS7), and lipid transport (apolipoproteins) and complement activation (oPS8). Recovery states were characterized by lipid transport and fatty acid metabolism (oPS2) and strong APR, response to infection, and catecholamine biosynthesis (oPS6) (Fig. 7C). Classification using nested CV ensemble learning differentiated between oPS classes with 90.1% overall accuracy and identified 167 significant uncorrelated features by permutation testing (fig. S17).

Fig. 7. Multiomic clustering defines states linking metabolic, immune, and recovery profiles to injury severity and outcomes.

(A) oPS trauma UMAP colored by PS. (B) Top five omic markers representative of each PS. (C) oPS enrichment and density of data points per time point. (D) Number of data points, median (IQR) of clinical values per oPS, and statistical comparison (one-way ANOVA with Tukey’s post hoc test). (E to I) Time series graphs of (E) lactate, (F) platelet count, (G) HGB, (H) clotting factor 9 (F9), and (I) MA of clot formation (linear mixed model with Šidák’s multiple comparisons test). *P < 0.05. See fig. S18 for full comparisons.

Clinically, oPSs differed in static and dynamic changes in vitals, hemostasis profiles, and blood chemistry (Fig. 7, D to I). oPS1 was the most perturbed with lowest GCS, SBP, and fibrinogen and highest HR, INR, and calcium and sustained elevated lactate (Fig. 7D, fig. S18, and tables S14 and S15). Other early injury states, oPSs 3 and 5, had more normal vitals and blood chemistry but lower platelet count and elevated INR. Between these, oPS3, with elevated humoral response had lower clotting factors and maximum amplitude–thromboelastography (MA-TEG). oPSs 7 (regulation of coagulation) and 8 (elevated complement) had lower PTT, INR, and fibrinogen with relatively normal clotting factor concentration. The remaining long-lived (oPSs 2 and 4) and recovery (oPS6) states were associated with lower P_aO₂, base excess, and INR with elevated fibrinogen.

Patient trajectories through multiomic trauma space

As before, multiomic trajectories (oT) were assigned to patients on the basis of their start ED oPS and sequence length (Fig. 8A). oT1 and oT4 were short-lived because of home discharge (oT4) or death (oT1), whereas long-lived trajectories remained in their initial oPS (oT3 energy production and oT4 hypotaurine metabolism) or converged toward APR and regulation of coagulation (Fig. 8, A to C). High within-group variance in oT2 and oT8 reflected this convergence across different oPSs (fig. S19). oTs were significantly different by permutation testing (P < 0.05), and oT classification had an overall accuracy of 93.1% (fig. S19, table S16). oTs had distinct injury patterns (Fig. 8, D to F). oT1 consisted of patients with severe injuries (high NISS and INR and low BE) with early mortality, whereas patients in oT2 (high NISS) began in the same omic state and had later (>48 hours) mortalities (14%) (Fig. 8G). There were no differences in mortality not explained by injury patterns.

Fig. 8. Multiomic trajectories reveal distinct profiles and integrate features from both layers.

(A) Multiomic trauma trajectory UMAP, colored by trajectory. (B) Density plot of data points per time point and number of patients per trajectory. (C) 2D plot of patient trajectories through oPS enrichment (y axis) over time (x axis) ± 95% CI. (D to F) Injury severity median ± IQR of (D) NISS, (E) INR, and (F) BE upon ED arrival (one-way ANOVA with Tukey’s post hoc test). (G to J) Proportions of (G) mortality, (H) ALI, (I) lung failure, and (J) heart failure per trajectory (Fisher’s exact test; *red asterisk indicates significant differences among similarly injured trajectories, and *black asterisk indicates significant differences among trajectories with different injury patterns). (K) Median ± IQR ICU-free days per trajectory (one-way ANOVA with Tukey’s post hoc test). (L to N) AUROC omic curves of (L) ALI (n = 19/35 Y/N), (M) lung failure (n = 20/20 Y/N), and (N) heart failure (n = 7/24 Y/N). (O) Regression of ICU-free days among patients with similar injuries (n = 40). (P to R) TAP validation dataset (P) NISS, (Q) INR, and (R) BE after trajectory assignment from COMBAT model. Prediction accuracy of (S) ICU-free days and (T) ventilator-free days in the TAP validation set using COMBAT-trained models. *P < 0.05, **P < 0.01, and ***P < 0.001.

However, there were differences in ALI and lung and heart failure not attributable to injury severity. oT8 had higher incidence of ALI with a significant increase in log odds compared with oT6 and oT3 (P = 0.037) despite no significant differences in injuries at presentation (Fig. 8H and fig. S20, A and B). Ensemble learning based on multiomic data classified ALI among these patients with similar injuries at higher accuracy than with injury covariates (AUROC = 0.86 versus 0.57, P = 0.005) (Fig. 8L and fig. S20). In addition, oT2 developed lung failure (P < 0.05 versus oT5 and oT8) and heart failure (P < 0.05 versus oT8) at higher rates (90 and 50%, respectively) despite similar injury patterns (Fig. 8, I and J). Classification of lung and heart failure was also improved using multiomics over injury covariates (AUROC 0.94 versus 0.68, P = 0.01; 0.91 versus 0.54, P = 0.03) (Fig. 8, M and N, and fig. S20). Further, oT2 had fewer ICU- and ventilator-free days than oT5 and oT8 despite similar injury patterns (Fig. 8K). Regression of ICU-free days was improved with multiomic markers (AUROC = 0.767) as compared with injury covariates (AUROC = 0.158, P < 0.001), as were ventilator-free days (AUROC = 0.923 versus 0.042, P < 0.001) (Fig. 8O and fig. S20). In other trajectories, oT4, which had the most comorbidities (67%; P < 0.05 versus oTs 3 to 8), had good outcomes from relatively mild injuries (fig. S21).

Injury severity patterns were similar in the TAP validation set, with TAP oT2 having higher NISS and INR and lower base excess compared with other TAP trajectories (Fig. 8, P to R). Further, TAP oT2 had fewer ICU- and vent-free days compared with TAP oTs 3, 6, and 8, which was consistent with the COMBAT trajectory findings (fig. S22). COMBAT-trained models classified the highest versus the lowest quartile ICU-free and vent-free days at 0.75 and 0.68, respectively, in the TAP validation set (Fig. 8, S and T).

DISCUSSION

Under the current trauma care paradigm, clinicians largely rely on experience to assess their patients’ current state and likely outcomes to guide treatment decisions. Improvements in care protocols have saved lives; however, survivors suffer prolonged critical illness, and seemingly identical patients have divergent outcomes for reasons that remain unclear. Given the heterogeneity of the individual patient response to trauma and treatment, data-driven approaches that capture the underlying biology of trauma hold promise to improve outcomes by resolving these complex, dynamic relationships. Although traditional predictors (NISS, base excess, lactate, INR, etc.) illuminate trauma phenotypes, the lack of multivariate patterns of these variables may explain clinicians’ inability to accurately distinguish patient outcomes and response to treatment. Recent research has identified trauma endotypes that reflect patient phenotypes and has further provided insight into why specific subgroups respond to treatment differently (35–37, 44, 46–50). However, the biological states that reflect the diverse injury patterns and the nonlinear trajectories of patients through injury and treatment toward final outcomes remain undefined.

Here, to address this gap, we defined trajectories as dynamic paths through distinct trauma omic states, which identified specific biological processes that drive thromboinflammatory outcomes. Patients were grouped into trajectories according to their underlying longitudinal molecular profiles, which allowed for an agnostic characterization. The longitudinal endotyping enabled comparison of patient trajectories that began in similar biological states and revealed a divergence in outcomes that could not be discerned by demographic, injury, or clinical data.

Metabolomic and proteomic layers were analyzed separately and then in combination to distinguish how circulating biomarkers from the metabolic derangements and tissue injury patterns of trauma underlie distinct biological states and could classify outcomes. Metabolomic trajectories displayed high variance in injury patterns, for example, high NISS variance in mT5 and high BE variance in mT2. This could reflect the short-lived nature of these metabolites in plasma or highlight the marked differences in patient metabolic response to injury, which remain to be determined. mPSs identified trajectories with a strong difference in INR and, to a lesser extent NISS, but not base excess, which typically reflects the degree of shock. Future research is needed to understand the relationships between clinically accepted measurements of injury and shock and circulating metabolomic profiles. Compared with patients with similar injuries, mT6 was susceptible to adverse outcomes, with high rates of heart failure, lung failure, ALI, and extended critical care times. These patients had elevated markers of fatty acid oxidation, lipid biosynthesis, polyamines, and succinate, which have been implicated as markers in trauma (17, 51, 52). Collectively, metabolomic endotypes offered insight into the energetic pathways elevated or perturbed early in trauma that underlie adverse outcomes.

Unlike metabolomics, a distinct shift in proteomic response over time was observed, with patients converging toward the classical APR after 24 hours. Before convergence, patients originated from very different proteomic start states, ranging from DAMP overload and detoxification to bacterial and humoral response. Injury severity was highly reflected in the proteomic states with patients with similar injuries originating from pPSs 1 and 4. Although injury covariates were highly prognostic, proteomic endotypes identified trauma states and markers that improved classification accuracy and further provided biological insight into the development of thromboinflammatory outcomes. This information may be clinically valuable for informing patient care, especially when outcome prediction is difficult (for instance, for patients with less severe injuries). As with metabolomics, future research is needed to determine whether these markers mediate outcomes and underlie treatment response.

Multiomic analysis was performed to gain additional insight into functional trauma states and trajectories. The multiomic trauma UMAP exhibited elements of both metabolomic and proteomic UMAPs, with the concentration of severely injured patients (oT1), convergence toward recovery states over time (oTs 2, 6, 7, and 8), and trajectories that remained in their start state (oTs 3 to 5). In general, proteins had higher feature importance in outcome classification as compared with metabolites, suggesting stronger prognostic accuracy with tissue injury markers in this study. All analyses identified trauma states of severe injury and metabolic crisis (mPS8, pPS1, and oPS1), which captured patients with early mortality (mT7, pT2, and oT1). Additional trajectories also began in these same start states (mT6, pT1, and oT2), and most of these patients survived. However, they suffered worse outcomes as compared with patients in other trajectories despite similar injury patterns, suggesting that the starting biological state and trajectory through trauma space influences thromboinflammatory outcomes.

Limitations include the small sample size and sparsity in some outcomes, like mortality. Although sample collection was overall very good, there is natural missingness of measurements based on censoring (death or discharge) and random missingness. These limitations have been mitigated in part by validation with the TAP cohort, in which outcomes and injury patterns persisted in assigned trajectories. The single ED time point TAP dataset was chosen for validation, given the critical need to endotype patients upon arrival. Future efforts will investigate endotyping by single (ED) and multiple (trajectory) time point sampling to further forecast patient trajectory and outcomes. Because the TAP dataset was used for base excess imputation for 20 of 118 (17%) patients missing this value at ED in COMBAT, the potential circularity introduced by using models trained on the validation cohort is an acknowledged limitation in the comparison of base excess in TAP after trajectory assignment. Efforts are ongoing to address these limitations and investigate and validate these findings in independent and external trauma datasets.

By design, this study focused on trauma endotyping by plasma molecular profiles alone. Although we highlight correlations with clinical data, future efforts will integrate clinical data to facilitate their utility in endotyping and identifying treatable clinical features that associate with plasma biomarkers. Further, much work is needed to determine whether omic markers directly mediate thromboinflammatory outcomes or reflect other disease processes. Efforts for direct clinical applications include a rapid means of state classification for trauma endotyping to inform clinical decision-making and provide a means for rapid discernment of precision medicine care for the patient with trauma. In this way, trauma endotyping by trajectory could dynamically guide and therefore improve outcomes of patients with trauma by enabling personalized care that targets distinct biologically and physiologically actionable disease states.

MATERIALS AND METHODS

Study design

This was an observational study using data gathered during the COMBAT clinical trial (28). n = 150 patients were calculated to be sufficient for statistical power in the COMBAT trial. For the present study, all patients enrolled in COMBAT with both metabolomic and proteomic data available were included, totaling 118 patients and 97 healthy controls. Because of the heterogeneity of baseline physiology and response to trauma, no patients were treated as outliers. Plasma underwent LC-MS/MS proteomics and metabolomics. Standard blood biochemical and coagulation data were collected including blood chemistry, counts, and assessment of coagulation factor concentrations, as well as patient demographics. End points for the study were 168 hours postinjury or the final time point of hospital care regardless of discharge status (home, death, long-term care, and other) if earlier. For biological and technical replicates, n = 1 plasma sample was collected from each patient at each time point. Blood was drawn with the first clinical samples collected in the field, upon arrival to the emergency department, or within 1 hour of hospitalization before critical care or resuscitation was given. Blood was processed to create platelet-poor plasma and frozen at −80°C for later analysis. Comprehensive demographic, injury, treatment, and outcome data were collected on all patients. For the TAP validation dataset, TAP patients were prospectively enrolled from all highest level trauma activations (based on trauma program physiologic and injury mechanism criteria) at a Major Level 1 Trauma Center. Patients under 18 years old, obviously pregnant, prisoners, and those known to be on anticoagulant medications were excluded. Plasma samples for COMBAT and TAP were collected under Colorado Multiple Institutional Review Board protocol nos. 12–1349 and 13–3087.

Statistical analysis

Omic features with constant values and with >95% missing were removed. Omic features were log2-transformed and scaled by mean centering and dividing by the SD for each feature to approximate and assume normal distribution for statistical analyses (fig. S24). Significant differences were determined as P < 0.05 using one-way analysis of variance (ANOVA) or Dunn’s test with Tukey post hoc analysis, Student’s t test, Fisher’s exact test, or Pearson χ² test. Omic comparisons were corrected using the Benjamini-Hochberg FDR method, and shown with median ± interquartile range (IQR) in box plots or median ± SEM in time series line graphs. Continuous clinical data with P < 0.05 were considered statistically significant, and shown using median ± IQR. For longitudinal analysis of omic markers and clinical measurements, linear mixed models (LMMs) were fitted to each variable with cluster, time point, and their interaction as fixed effects and a random intercept per patient, followed by post hoc pairwise comparisons of cluster levels using estimated marginal means (adjusted for multiple comparisons with Šidák’s method). For the circos plot, LMMs were also used to determine associations between omic markers and clinical measurements (fixed effects of age, sex, and time; random effects of injury group and individual patient), again with Šidák’s multiple comparison correction on the estimated marginal means. The partial slope was used for comparison and visualization in the circos plots.

To quantify trajectory divergence, DTW within a permutation testing (×1000) framework was used to compare distances (53). DTW measured the pairwise similarities, which were used to quantify the within- (mean pairwise DTW distances between all patients within each trajectory) and between-group (mean DTW distance between trajectories) distances. P values were computed as the proportion of permuted distances that were greater than or equal to the observed distance.

Nested 10-fold CV ensemble learning with LASSO glmnet, random forest, and xgboost was used for classification of proteomic, metabolomic, and multiomic patient states and trajectories within the CV.SuperLearner framework (54). Hyperparameters were tuned to achieve maximum classification accuracy. Model performance was evaluated using a confusion matrix to assess classification accuracy and error rates. To identify omic signatures, permutation testing was conducted to identify significant features by comparing observed feature importance to a null distribution generated from permuted class labels and isolating features with importance exceeding a threshold of two SDs above the permuted mean. Highly correlated variables among significant features were removed (|r| > 0.75) based on pairwise correlation analysis, resulting in a final set of omic features for each patient state and trajectory.

For binary outcomes classification, nested 20-fold CV ensemble learning with linear model, LASSO glmnet, random forest, and xgboost was used within CV.SuperLearner. Feature selection was performed using nested 20-fold CV LASSO glmnet, and features with positive coefficients were displayed in feature importance frequency plots. Classification performance was evaluated using AUROC, and model comparisons were performed using Delong’s method. Regression analysis of ICU-free days and ventilator-free days was performed using nested 20-fold CV ensemble learning in SuperLearner. To compare the regression models statistically, we performed a t test on the squared residuals. For the TAP validation dataset, overlapping features between COMBAT and TAP datasets were selected and used. Both datasets underwent Yeo-Johnson data transformation before comparison.

Supplementary Material

The PDF file includes:

Materials and Methods

Figs. S1 to S23

Legends for tables S1 to S16

Legend for data file S1

References (56–81)

Other Supplementary Material for this manuscript includes the following:

Tables S1 to S16

Data file S1

MDAR Reproducibility Checklist

Data, code, and materials availability:

All data are available in the paper or the Supplementary Materials. Protein and metabolite expression data are available in data file S1. Patient-level data and any additional information required to reanalyze the data reported in this work are available from the corresponding authors upon request. Raw proteomic MS files have been deposited in the PRIDE database under accession number PXD053799, and raw metabolomic MS files are available on Metabolomics Workbench (https://dx.doi.org/10.21228/M85R82, project PR002069). The code used for analysis has been deposited to Dryad (55) and is available at https://doi.org/10.5061/dryad.d51c5b0dn. All materials used or generated in this study are commercially available or will be supplied upon reasonable request.