Serum Levels of Branched Chain Amino Acids Predict Duration of Cardiovascular Organ Failure in Septic Shock

Abstract

Background:

Sepsis shifts cardiac metabolic fuel preference and this disruption may have implications for cardiovascular function. A greater understanding of the role of metabolism in the development and persistence of cardiovascular failure in sepsis could serve to identify novel pharmacotherapeutic approaches.

Methods:

Secondary analysis of prospective quantitative ¹H-NMR metabolomic data from patients enrolled in a phase II randomized control trial of L-carnitine in septic shock. Participants with a SOFA score of >=5, lactate >=2, and requiring vasopressor support for at least 4 hours were eligible for enrollment. The independent prognostic value of metabolites to predict a) survival with shock resolution within 48 hours and b) vasopressor free days were assessed. Concentrations of predictive metabolites were compared between participants with and without shock resolution at 48 hours.

Results:

Serum ¹H-NMR metabolomics data from 228 patients were analyzed. 81 (36%) patients met the primary outcome; 33 (14%) died prior to 48 hours. The branched chain amino acids (BCAA), valine, leucine, and isoleucine were univariate predictors of the primary outcome after adjusting for multiple hypothesis testing, while valine remained significant after controlling for SOFA score. Similar results were observed when analyzed based on vasopressor free days, and persisted after controlling for confounding variables and excluding non-survivors. BCAA concentrations at 48 hours significantly discriminated between those with shock resolution versus persistent shock.

Conclusions:

Among patients with septic shock, BCAA concentrations independently predict time to shock resolution. This study provides hypothesis generating data into the potential contribution of BCAAs to the pathophysiology of cardiovascular failure in sepsis, opening areas for future investigations.

Introduction

Sepsis remains a major public health priority, representing one of the leading causes of both inpatient death and spending (1–3). The mortality rate increases dramatically when cardiovascular failure and hemodynamic shock are present, reaching approximately 40% (4). Despite these grim statistics, novel pharmacotherapies for the condition are few, and the cornerstones of treatment rely on early identification, hemodynamic stabilization and prevention of ongoing hypoperfusion, and the administration of prompt antimicrobials. Critically, a recent analysis suggests the majority of deaths from sepsis are not preventable through improvements in hospital-based care (2). As such, novel approaches to the syndrome are highly desirable.

Among the many pathophysiologic disturbances of sepsis is a significant disruption and dysregulation of metabolism. Elevations in lactate are well known to occur and highly predictive of outcome, particularly when measured early in the course of sepsis (5, 6). However, hyperglycemia, hyperlactatemia, lipolysis, and protein catabolism are also common and similarly predict outcome (7–9). In addition to serving as prognostic markers, there exists evidence that sepsis affects metabolic fuel, the function of vital organs (10–15), and may contribute to multisystem organ failure. In the heart, sepsis causes decreased fatty acid utilization in favor of glycolytic substrates such as glucose and lactate, which are inherently associated with less efficient energy expenditure (10). Similar changes, as well as loss of metabolic flexibility, can be observed in the setting of chronic heart failure, a condition characterized by chronic rather than acute inflammation (16–18). While maladaptive in the chronic state, it remains unclear whether this shift represents an adaptive or maladaptive response in the setting of septic shock.

The traditional treatment of septic shock often involves the application of vasopressors, many of which manifest their effects via stimulation of adrenergic receptors and generation of cyclic adenosine monophosphate (cAMP). These agents affect numerous metabolic pathways, including those involved in the generation of lactate (19), that affect the underlying metabolic disruptions that contribute to cardiac failure in sepsis. A better understanding of these processes, as well as the role of metabolic flexibility, may promote the development of novel metabolic therapies to address cardiac dysfunction in sepsis.

In order to better discern the relationship between metabolism and cardiovascular failure in septic shock, we designed a secondary analysis of prospectively collected metabolomic and clinical data of over 200 patients with septic shock. The objective of this study was twofold. First, we sought to determine if concentrations of peripherally measured metabolites of participants with septic shock independently predict shock resolution using either an early (<48 hour) binary definition of septic shock resolution, or a continuous (vasopressor free days alive at 28 days) definition. Second, we aimed to test if these predictive metabolites normalize among patients with early shock reversal.

Methods

Study design:

Serum samples specifically for quantitative proton (¹H) nuclear magnetic resonance (NMR) metabolomics were collected as part of a previously described 16-center clinical trial comparing 3 doses of L-carnitine to saline placebo (20). The parent trial was approved by each site’s institutional review board, all patients or their surrogate gave written informed consent, and the trial was registered at clinicaltrials.gov prior to initiation (NCT01665092). Samples were drawn at the time of enrollment and at 48 hours after treatment, were allowed to clot at room temperature and were centrifuged within 30 minutes following standard operating procedures. Serum samples were stored (−80° C) until the time of analysis. At the time of assay, samples were thawed on ice and analyzed as previously described (13).

Study groups and outcomes:

Since all patients enrolled in the trial met criteria for septic shock, all patients with available samples were assessed for the primary outcome of shock resolution within 48 hours. Short-term shock resolution was defined as the complete cessation of vasopressor administration that was sustained for at least 24 hours in participants who survived at least 48 hours. All other patients (including those that died) were assigned to the persistent shock group to account for competing risks. This time point was chosen because the primary outcome of the clinical trial was change in SOFA score from enrollment to 48 hours. Time to shock resolution was also evaluated as a continuous variable, defined as the number of vasopressor free days over 28 days (21, 22). Consistent with the calculation of ventilator-free days (23), patients who died within 28 days of any cause were assigned 0 shock-free days, whereas survivors were assigned a value of (28 - days requiring vasopressors). Among survivors to 48 hours, the concentrations of metabolites of patients with persistent shock were compared with those with shock resolution. As a post-hoc analysis, the prediction of 28-day mortality of BCAAs was investigated. A correlation network analysis was constructed in Metscape (24) and as a post-hoc exploratory analysis, a pathway analysis was conducted in Metaboanalyst (https://metaboanalyst.ca) (25).

Statistical analyses:

Descriptive data are reported as means and standard deviations, medians with interquartile ranges, or proportions as appropriate. Consistent with our prior work, missing values were imputed as one-half the lowest measured metabolite concentration for each metabolite if present in at least 70% of the cohort, under the assumption the actual level was below the level of detection of the assay. Metabolite concentration data were natural-log transformed, median normalized, and scaled. For the primary analysis, metabolite concentrations at enrollment were compared between the shock resolution and persistent shock groups at 48 hours. A receiver operator characteristic (ROC) curve of the best performing metabolite was constructed and the area under the ROC (AUROC) was calculated. To determine the predictive value of individual metabolites on the primary outcome, multivariable logistic regression models were constructed, with significant differences between groups as well as treatment with L-camitine considered as candidate variables which were assessed by backwards stepwise elimination while maintaining variables with a p<0.10. To assess the robustness of our findings, a secondary analysis was performed using vasopressor free days in all patients through the use of multivariate Cox proportional hazards using the same approach. In both analyses, preplanned subgroup analyses were performed only among survivors in case metabolite concentrations only predicted early death and not shock resolution. As the study was a randomized controlled trial of a metabolic therapy, a subgroup analysis of only patients enrolled into the placebo arm was also planned. Finally, to assess whether metabolites changed concurrently with resolution of cardiovascular failure (defined as persistent shock), metabolite concentrations at 48 hours were assessed among those with and without shock resolution. For all comparisons, tests were two-sided and p-values of <0.05 were considered significant. P values were corrected for multiple comparisons using the false discovery rate (FDR) method of Storey, et al (26); metabolites with an FDR below 5% were considered as differentiating. For all multivariate analyses, post-estimation jackknife and bootstrap analyses were performed to assess for the effects of potential individuals or groups of outliers that disproportionately affected the findings. Data were analyzed using STATA v15.1 (College Station, TX).

Results

Of the 250 participants randomized in the parent trial, quantitative ¹H-NMR metabolomics identified 27 serum metabolites from the serum samples of 228 participants. Metabolomics data sets are available at the Metabolomics Workbench (https://www.metabolomicsworkbench.org/; accession numbers 10.21228/M8VX0Z (T0), 10.21228/M8810Q (T48)). Shock resolution at 48 hours was observed in 81/228 (36%) participants, leaving 147/228 (64%) with persistent shock; 33 of these patients died prior to 48 hours and represented the subgroup analysis of only survivors to 48 hours. Patients enrolled into the placebo arm (69/228; 30%) represented the second subgroup analysis. Clinical and demographic variables of the cohort stratified by the primary outcome are summarized in Table 1. The median (interquartile range) number of vasopressor free days was 19 (0, 26), and 103 (45%) of patients died prior to 28 days and were considered not to have achieved the clinical outcome of shock resolution and survival. Univariate analysis identified 6/27 metabolites (valine, leucine, isoleucine, methionine, phenylalanine, and propylene glycol, Figure 1A–F) that significantly differentiated patients with shock resolution and persistent shock, with 2 of them (valine and leucine; q < 0.001, 0.03, respectively) remaining significant after the FDR correction (Figure 1A and 1B). The AUROC of the best performing metabolite (valine) was 0.67 (95% CI 0.60-0.74; Figure 2). In the subgroup analysis, excluding those who died before 48 hours, 3 metabolites (valine, leucine, and phenylalanine) were significant predictors after FDR adjustment (q < 0.001, 0.01, 0.03). In the subgroup analysis of 69 placebo treated patients, valine and leucine were the strongest predictors but did not reach statistical significance (p = 0.05 and 0.06, respectively). In the multivariable logistic model, valine remained a significant predictor of shock resolution after controlling for severity of illness through the use of the sequential organ failure assessment (SOFA), as illustrated in Table 2. The final model included SOFA score, heart rate, as well as methionine concentration. Methinonine did not reach statistical significance (p = 0.052) but was nevertheless retained in the final model based on our a priori methods. L-carnitine treatment, cumulative vasopressor index (CVI), and lactate were not independent predictors in the model. Forcing lactate into the final model did not significantly affect the conclusions. Jackknife and bootstrap modeling yielded similar results and had minimal impacts on the odds ratios and did not affect the significance of the results (data not shown). Subgroup analysis confined to the smaller cohort of only placebo treated patients demonstrated similar findings, but did not reach statistical significance (p = 0.07).

Table 1:

Demographics and clinical characteristics of patients with and without shock resolution by 48 hours from enrollment.

Variable	Shock resolution (n = 81)	Persistent shock (n = 147)	P-value
Demographics
Age, years (IQR)	62 (52, 70)	64 (55, 73)	0.26
Male, n (%)	51 (63)	77 (52)	0.12
Female, n (%)	30 (37)	70 (48)
Race
Black, n (%)	24 (30)	45 (31)	0.91
Asian, n (%)	2 (2)	3 (2)
White, n (%)	48 (59)	89 (61)
Other, n (%)	7 (9)	9 (6)
Ethnicity
Hispanic, n (%)	5 (6)	7 (5)	0.64
Past medical history
Diabetes, n (%)	26 (32)	54 (37)	0.59
Liver disease, n (%)	8 (10)	26 (18)	0.12
Renal disease, n (%)	8 (10)	26 (18)	0.12
Physiologic variables
Heart rate, beats per minute (IQR)	94 (82, 107)	102 (88, 115)	0.01
Respiratory rate, breaths per minute (IQR)	20 (16, 24)	21 (18, 26)	0.12
Cumulative vasopressor index (IQR)	4 (3, 7.5)	4 (4, 8)	0.001
Body mass index (IQR)	28 (22, 36)	28 (24, 34)	0.22
Laboratory values
White blood count, cells/mm3 (IQR)	17 (12, 26)	19 (11, 28)	0.69
Platelet count, cells/mm3 (IQR)	178 (98, 246)	132 (70, 204)	0.01
Creatinine, mg/dL (IQR)	1.8 (1.2, 2.7)	2.0 (1.3, 2.9)	0.31
Total Bilirubin, mg/dL (IQR)	0.9 (0.5, 1.7)	1.4 (0.6, 3.3)	0.007
Lactate, mmol/L (IQR)	3.1 (2.0, 5)	4.1 (2.7, 7)	0.002
Severity of Illness
SOFA score	9 (7, 12)	12 (10, 14)	<0.001
Co-interventions
L-carnitine treatment	55 (68)	104 (71)	0.65

Figure 1A-1F:

Dot plots with bar graphs of the transformed, normalized metabolite concentrations between patients with shock resolution versus those with persistent shock. Valine (A), leucine (B), isoleucine (C), methionine (D), phenylalanine (E) and propylene glycol (F) were chosen based on an initial p < 0.05. Valine (A) and leucine (B) remained significant after correction for a false discovery rate of 0.05 (*). Bar graphed data are the median (interquartile range) of patients with persistent (n=81) and resolved shock (n=147).

Figure 2:

Receiver operator characteristic curve of the best performing metabolite in univariate analysis (valine) to predict survival with resolution of shock within 48 hours (AUROC 0.67; p <0.0001).

Table 2:

Univariate and multivariate logistic regression model of metabolite concentrations at enrollment to predict the primary outcome

Variable+	Univariate odds ratio	Lower 95% CI	Upper 95% CI	Multivariate odds ratio	Lower 95% CI	Upper 95% CI
Valine*	1.93	1.40	2.67	1.49	1.05	2.11
Leucine*	1.61	1.18	2.19	--	--	--
Isoleucine	1.44	1.06	1.96	--	--	--
Propylene glycol	0.70	0.52	0.94	--	--	--
Methionine#	0.73	0.56	0.96	0.74	0.55	1.00
Phenylalanine	1.33	1.00	1.76	--	--	--

CI - Confidence interval; SOFA - Sequential Organ Failure Assessment score (at enrollment)

⁺Other significant variables retained in the multivariate model were SOFA score and heart rate.

^*FDR corrected p value ≤ 0.05.

^#Methionine was not a significant independent predictor but was maintained in the model due to p<0.10 per the a priori methods.

When shock resolution was analyzed as a continuous rather than binary variable using vasopressor free days alive, 10 metabolites were significant, 8 of which withstood FDR correction. After excluding non-survivors, 3 of these (isoleucine, valine, and leucine) were significant before and only 1 (isoleucine) after FDR correction. In the multivariate Cox proportional hazard model, isoleucine, 2-oxo isocaproate (ketoleucine), and tyrosine remained significant independent predictors after controlling for baseline differences between groups. The final model also included SOFA score and body mass index. Isoleucine was not a significant predictor in the smaller subgroup of only placebo treated patients (p = 0.27). To visually assess the predictive capacity of the best performing metabolite (isoleucine) for shock resolution in the entire cohort, isoleucine was dichotomized into low or high values at enrollment based on the median transformed and normalized value. Time to event curves are presented in Figure 3, and demonstrates significant uncorrected univariate differentiation between groups in vasopressor free days (p = 0.005).

Figure 3:

Time to event curve for achieving shock resolution based on serum isoleucine concentration at study enrollment. The x-axis represents vasopressor free days and the y-axis represents the proportion of patients achieving shock resolution. Note that the mortality rate at 28 days was 45%, so nearly half the cohort did not achieve the outcome, consistent with other studies of septic shock. Patients were dichotomized into low and high isoleucine based on the median normalized, log transformed value (p=0.005).

Univariate analysis of metabolite concentrations at 48 hours among patients with or without shock resolution identified 8/27 metabolites, 7 of which remained significant after correction for multiple comparisons (Table 3). Two of these (valine and tyrosine) remained significant independent predictors in multivariate logistic regression. Jackknife and bootstrap postestimation odds were consistent with the initial model. In addition to the multivariate modeling, we explicitly tested if BCAA levels differed between patients treated with placebo or L-carnitine. None of the three differed significantly by L-carnitine treatment, (p = 0.10-0. 64), even prior to application of a correction for multiple hypothesis testing.

Table 3:

Univariate and multivariate logistic regression model of metabolite concentrations at 48 hours that differentiate patients with shock resolution versus persistent shock

Variable	Univariate odds ratio	Lower 95% CI	Upper 95% CI	Multivariate odds ratio	Lower 95% CI	Upper 95% CI
Isoleucine*	2.20	1.41	3.44	2.11	1.34	3.32
Creatine*	0.54	0.38	0.75	0.54	0.38	0.78
Oxoisocaproate*	1.94	1.33	2.81	--	--	--
Leucine	1.74	1.24	2.45	--	--	--
Tyrosine	1.81	1.27	2.59	1.87	1.28	2.73
Valine	1.81	1.27	2.56	--	--	--
Glucose	1.60	1.15	2.23	--	--	--
Ornithine	1.38	1.00	1.91	--	--	--

CI - Confidence interval; SOFA - Sequential Organ Failure Assessment score (at enrollment)

^*FDR corrected p value ≤ 0.05.

As a post-hoc hypothesis generating exercise, we tested if BCAAs merely predict cardiovascular outcomes or also mortality. All BCAAs were predictive of mortality in univariate modeling (OR 1.6–1.8, p<=0.001). We further wanted to begin to assess which metabolic profile and pathways are associated with BCAAs to begin to glean insights that might be useful for the design of subsequent experiments. To achieve this, we constructed a correlation network using Metscape (http://metscape.ncibi.org/). For this, the Pearson correlation coefficients across all the measured metabolites were uploaded into Metscape (Figure 4). Valine was highly correlated with leucine and leucine with isoleucine. Both isoleucine and valine were negatively correlated with betaine, a direct oxidation product of choline and an osmolyte. The data used to create the correlation network can be found in the supplemental digital content, http://links.lww.com/SHK/B177.

Figure 4:

The correlation network of serum metabolites detected by quantitative ¹H-NMR metabolomics. Valine and isoleucine strongly correlate with the branched chain amino acid, leucine. Both valine and isoleucine negatively correlate with betaine. Pearson correlation coefficients were generated in Metaboanalyst and uploaded into Metscape (see supplemental digital content, http://links.lww.com/SHK/B177). Red designates a positive correlation and blue, negative. Line thickness represents the strength of the association.

Discussion

In this study, we investigated the association between cardiovascular failure, defined as the presence of hemodynamic shock, and peripherally measured metabolic disruption in the setting of sepsis. The primary finding of the study was that BCAAs (valine, isoleucine, and leucine) were predictors of shock resolution. These findings persisted regardless of whether an early (<48 hour) binary outcome or an intermediate continuous vasopressor free day outcome was used. Furthermore, these findings persisted after controlling for initial severity of illness. These findings provide novel insights into metabolic pathways associated with cardiovascular failure in sepsis.

Metabolic dysregulation in sepsis is a widespread hallmark of the syndrome, and is well documented to predict clinical outcomes. Mitochondrial and metabolic dysfunction have further been hypothesized to contribute to and may even be responsible for the multiple organ dysfunction syndrome (MODS) that frequently accompanies sepsis (27–29). As such, further understanding of the underlying pathophysiologic changes of metabolism in sepsis may provide insights into the disease process that could lead to the identification of novel drug targets and ultimately new pharmacotherapeutic approaches. There is an increased recognition that chronic inflammation, metabolic syndrome, and maladaptive organ function (such as chronic heart failure) are tightly linked. As such, metabolically informed therapies may prove efficacious in treating these conditions in the chronic state. As sepsis induces profound acute inflammation associated with similar metabolic changes as observed in the chronic setting, we hypothesize that similar metabolically informed therapies may prove advantageous.

Previous studies demonstrate that sepsis profoundly shifts metabolic fuel preference in organs such as the heart, away from fatty acids towards glycolytic substrates, including lactate (10). This shift is associated with inefficient myocardial work that is not fully understood. Shifts in myocardial metabolism are associated with changes in cardiac function, and could be partially responsible for sepsis cardiomyopathy. Indeed, recent data suggest sepsis cardiomyopathy as identified by left ventricular global longitudinal strain is associated with severity of lactate elevation, though whether these changes are simple associations or represent causal pathways remains untested (30). Our work confirms that cardiovascular dysfunction is associated with changes in multiple metabolites, and may provide insights into specific pathways involved in its resolution.

The most striking and consistent observation from this study was the association between BCAAs and shock resolution. The role of BCAAs in sepsis is incompletely understood. Studies have variably reported decreased or normal levels in sepsis (31, 32), with evidence suggesting increased clearance in the liver from the periphery (33). Furthermore, BCAAs have previously been reported to be associated with sepsis survival (32). Our observations suggest this might vary based on when samples were acquired in the course of the disease process and / or the relative severity of shock physiology, which could explain the variability in BCAA levels reported in the literature. Furthermore, high BCAA concentrations may represent an adaptive response to sepsis that predisposes patients to shock resolution and survival. Indeed, we previously reported increased valine in sepsis survivors (34).

While the role of BCAAs in sepsis is still relatively poorly understood, there is increasing acknowledgement of their importance in cardiovascular health (35). As a therapeutic, infusion of valine and leucine have been demonstrated to reduce arrhythmias and decrease blood pressure when administered in a left anterior descending occlusion ischemia reperfusion model (36). On the other hand, recent data from patients with heart failure demonstrates that chronic accumulation of BCAAs selectively inhibit mitochondrial pyruvate (and therefore lactate and glucose) metabolism, and sensitizes the heart to damage from ischemia reperfusion injury (37). BCAAs, particularly leucine, are also known to activate the mammalian target of rapamycin (mTOR) (38). mTOR plays a role in inflammation and mitochondrial biogenesis, and may be directly related to sepsis pathophysiology. It may also influence the Warburg-like effect observed in sepsis. We speculate these disparate findings may suggest an acutely adaptive but chronically maladaptive mechanism of cardiac metabolic fuel choice, though this hypothesis remains to be tested.

Apart from the changes of BCAAs and their breakdown products, we also observed changes in tyrosine and phenylalanine associated with shock resolution. As these amino acids serve as precursors to endogenous catecholamine biosynthesis, it is not surprising that the levels of these metabolites might differ between groups with more or less rapid resolution of shock. Propylene glycol was also noted to be different between groups. While the source of this difference is unknown at this time, it is possible this could reflect differences in benzodiazapine usage, which was not tracked for this study. This has the potential to be of clinical relevance, as at least one previous report suggests it can contribute to lactate generation and metabolic acidosis in patients receiving prolonged infusions in the ICU (39).

In addition to the expected association between valine and isoleucine, the correlation analysis highlights a negative association between valine and leucine and the osmolyte betaine. While not directly metabolically related, this finding suggests that patients with persistent shock and higher levels of betaine, may have significant disruption in cell (hydration) homeostasis. As an important osmolyte (40) regulating cellular hydration, the disruption of which can lead to cell apoptosis (41), this pathway may have relevance for the development of MODS. However, from the current work it is unclear whether betaine is a marker of shock severity, or whether it simply reflects differences in clinical practice related to the use of cystalloid fluid.

There are several important limitations to consider in this study. First, metabolomics data are not available in real time, and this analysis was performed on normalized rather than raw concentrations. As such, the clinical utility of these measurements for risk prognosis in practice would need to be completed on real clinical samples rather than the highly processed metabolomics samples prior to implementation. Rather, this study lends hypothesis generating insights into the role of metabolism in sepsis-induced shock states. Second, it is possible that confounders, including vasopressors or L-carnitine treatment, could have affected our results. However, we took steps to evaluate this possibility and did not observe evidence of such an effect, though the possibility of unmeasured confounders affecting our results remains as in any assosiative study. Third, we used a ¹H-NMR metabolomics platform that generally identifies relatively high abundant, small polar molecules, many of which are amino acids. An untargeted liquid chromatography mass spectrometry approach may have yielded different results and/or insights. Fourth, we performed multiple hypothesis testing, raising the possibility of false positive results. However, we applied a conservative FDR rate, and found that the BCAAs and their breakdown products are consistently evident across multiple different methods of analysis and multivariable models, increasing the likelihood these findings represent a true physiologic phenomenon. Finally, this study was observational in nature, and due to the aforementioned limitations, all conclusions regarding pathways and mechanisms remain hypothesis generating.

Conclusion

Among patients with septic shock, serum concentrations of the BCAA (isoleucine, leucine and valine) at enrollment independently predict time to shock resolution. This study provides hypothesis generating data into the potential contribution of BCAAs to the pathophysiology of cardiovascular failure in sepsis, opening areas for future investigations.