Metabolomics discloses potential biomarkers to predict the acute HVPG response to propranolol in patients with cirrhosis

In cirrhosis, a decrease in hepatic venous pressure gradient (HVPG) > 10% after acute iv propranolol (HVPG response) is associated with a lower risk of decompensation and death. Only a part of patients are HVPG responders and there are no accurate non‐invasive markers to identify them. We aimed at discovering metabolomic biomarkers of HVPG responders to propranolol.


| INTRODUC TI ON
Portal hypertension (PHT) is the driving force for decompensations occurring in cirrhosis. A portal pressure gradient ≥10 mm Hg as measured by the hepatic venous pressure gradient (HVPG) defines clinically significant PHT. Beyond this value the presence of oesophageal varices, ascites, hepatic encephalopathy and hepato-renal syndrome can appear; an HVPG ≥12 mm Hg is necessary for varices to bleed. 1 A decrease in HVPG ≥20% from baseline values or below 12 mm Hg spontaneously or under treatment with non-selective β-blockers (NSBB) is associated with less incidence of PHT-related complications and of death. [2][3][4] The HVPG response to β-blocker therapy can also be determined with a single haemodynamic study, 5,6 where responders (about 50%-60%) are identified by decrease ≥10% from baseline after the acute administration of iv propranolol (0.15 mg/kg). As for the chronic response of HVPG, these patients have a lower incidence of decompensation and death. [5][6][7] Several factors can influence the HVPG response to NSBB, including the degree of liver failure, the dose of β-blockers, the extent of portal-systemic collaterals and varices. 3,[8][9][10] However, different parameters assessed to predict the HVPG response (heart rate, femoral/portal blood flow changes, β-adrenoceptors polymorphisms, antrum mucosa vasoactive proteins) [11][12][13][14][15] have not been accurate enough. Therefore, the invasive measurement of HVPG is necessary and, despite minimally invasive, some patients are reluctant to it and the technique is not universally available. Thus, it would be relevant to develop non-invasive methods for recognizing the response to β-blockers for medical therapy optimization.
Metabolomics is an "omics" discipline that has gained interest in biomedical research. The application of high-throughput techniques such as liquid chromatography coupled to mass spectrometry (LC-MS) allows measuring simultaneously thousands of metabolites from a variety of complex samples (biological fluids or tissue extracts) in a short-time period. These techniques perform a semi-quantitatively analysis of a wide range of molecules, such as glycerophospholipids, glycerolipids, sphingolipids, fatty acids, bile acids or amino acids. Recently, potential applications for metabolomics have been proposed by reporting specific profiles in several liver disorders such as drug-induced liver injury, NASH or idiopathic portal hypertension or to stratify degrees of cirrhosis severity. [16][17][18][19][20][21][22] Most of these early metabolomic analysis were untargeted and metabolites, defined based on their retention time and mass to charge ratio, could not always be identified. However, in the recent years, noteworthy advances in metabolomic field allow targeted analysis that identify specific metabolites, generating hypothesis and developing prognostic models to be used in clinical practice.
The aim of this pilot study was to identify a metabolomic serum profile in patients with cirrhosis and PHT that allows a non-invasive prediction of the HVPG response to acute iv propranolol.

| PATIENTS AND ME THODS
Sixty six patients with cirrhosis and clinically significant PHT (HVPG ≥ 10 mm Hg), in whom HVPG response to iv propranolol was assessed, were prospectively included between September 2010 and June 2015. All patients had oesophageal varices with or without a previous variceal bleeding and were to initiate primary or secondary prophylaxis with non-selective β-blockers. Inclusion criteria were diagnosis of cirrhosis (liver biopsy or unequivocal clinical data and compatible findings on imaging techniques); age between 18 and 80 years; HVPG ≥ 10 mm Hg; presence of oesophageal varices (with or without previous bleeding episode); and indication of β-blockers. Exclusion criteria were severe liver failure (Child-Pugh score>12 points); recent blood-derived product transfusion (patients with bleeding); hepatocellular carcinoma; acute alcoholic hepatitis; portal vein thrombosis; contraindications to β-blockers; pregnancy; or refusal to participate in this study. Hepatocellular carcinoma and acute alcoholic hepatitis were excluded because of their unknown effects on metabolome. All patients were on stable clinical conditions and patients with recent bleeding, this study was done at day 5 of admission and they were on clinical and haemodynamic stable conditions to initiate with β-blockers. This study was conducted following the principles of the Declaration of Helsinki (revised in Seoul in 2008). This study was approved by the Ethics Committee for Clinical Investigation of Hospital Clinic (registry number 2010/6008, approval 9/IX/10) and all patients gave their written informed consent.
Baseline clinical characteristics and laboratory tests were collected, as well the treatments received. Treatments were grouped in major families to control potential effects on metabolites profiles.
The groups were: antibiotics, diuretics, antihypertensive drugs, insulin/oral antidiabetics and other hormones.

| Haemodynamic studies
The measurement of HVPG and its response to iv propranolol was performed as previously reported. 23 Briefly, after an overnight fast, under local anaesthesia (mepivacaine 1%, subcutaneously) with ultrasonographic guidance, an 8F venous catheter introducer (Axcess; Maxxim Medical, Athens, TX) was placed in the right jugular vein by the Seldinger technique. Under fluoroscopy, a 7F balloon-tipped catheter ("Fogarty" Edwards Lifesciences LLC, CA) was guided into the main right or middle hepatic vein for measurements of wedge (occluded,

Key Points
• Two serum metabolites help at identifying patients with cirrhosis and a good response to β-blockers for portal hypertension. It may help avoiding invasive studies to assess this response and may facilitate a better individualization of therapy. WHVP) and free hepatic venous pressures (FHVP). HVPG results from the difference between WHVP and FHVP. The adequacy of occlusion was checked by gentle injection of a small amount of radiologic contrast medium after balloon inflation. After baseline measurements, iv propranolol (0.15 mg/kg) was administered over 10 minutes. HVPG response was assessed at minute 15-20 as previously described. 5,6 A positive HVPG response was defined as a decrease equal or greater than 10% from baseline value. All measurements were taken in triplicate and permanent tracings were obtained in each case in a multichannel recorder (GE Healthcare, Milwaukee, WI), and were reviewed specifically for this study by experienced investigators (JB, and JCGP).

| Blood sample details and metabolomic profiling
Blood samples were obtained prior to the haemodynamic studies. All data were processed using the TargetLynx application manager for MassLynx 4.1 (Waters Corp.). The peak detection process included 389 metabolic features, identified prior to the analysis.
Intrabatch normalization followed the procedure described by Martinez-Arranz et al (see Data S1). 24

| Metabolomic variables
We initially explored differences in metabolic profile of responders and non-responders after a Log2 fold-change transformation for each metabolite (Log2 FC = Log2(Average responders) -Log2(Average non-responders)). Log2 conversion makes data to become more normally distributed. 25 Thereafter, a univariate analysis by unpaired Student's t test (or Welch´s t test when unequal variances) was applied to assess differences among responders and non-responders.
Heatmaps were created to show individual metabolite differences between groups.
After the metabolomic broad profiling analysis, only metabolites with a baseline chromatographic resolution and a good signal to noise ratio were considered for analysis and prognostic model development. This selection was done to build a robust model, easily reproducible and transferable to other laboratories worldwide.
These metabolites underwent univariate logistic regression to assess HVPG response; those with a P ≤ 0.05 underwent standard stepwise logistic regression to find predictive combinations of response. Based on the sample size, the proportion of responders and to avoid overfitting, we studied combinations of two or three metabolites. The performance of the models was assessed by means of AUROC curves and Akaike information criterion (AIC), which gives a relative value among potential models so that the best one has the lowest value. 26 The best metabolite prognostic model was studied for potential cut-off points to identify HVPG responders and non-responders.
Youden Index, sensitivity (SN), specificity (SP) and predictive values were assessed. The selected model underwent internal validation using "leave-one-out" cross-validation computing. 27 The potential association of the significant metabolites and its potential influence by relevant clinical variables (aetiology, Child-Pugh class, prophylaxis type or medications) was further assessed.

| Clinical variables
Baseline variables were compared between HVPG responders and non-responders. Those clinical variables associated with the HVPG response were introduced in the metabolomic models attempting to improve their performance. Quantitative variables are expressed as mean ± standard deviation, and qualitative variables as absolute and relative frequencies. Categorical variables were compared using the chi-square test. Continuous variables were compared with Student's t test. Logistic regression was used for multivariable analysis.
Statistical analysis of clinical, laboratory and haemodynamic data was performed with the statistical package spss 20.0 (SPSS Inc, Chicago). Metabolic statistical analysis was performed with r statistical package v.3.1.0. Statistical significance was established at a P < 0.05.

| Patients and haemodynamics
The clinical and laboratory characteristics of the 66 patients included in this study are summarized in Table 1. No baseline clinical, biochemical or haemodynamic variables were significantly different between HVPG responders and non-responders, except for a greater proportion of non-responders among patients to begin secondary prophylaxis. Forty-one (62%) patients were responders and 25 (38%) were non-responders. The mean HVPG was 16.9 ± 3.6 mm Hg with a mean decrease in responders of 21 ± 12%. Table 2 summarizes hepatic haemodynamic characteristics of patients.  TA B L E 2 Hepatic haemodynamics of the patients and response to iv propranolol non-responders and were included for logistic regression analysis.

| Metabolomic analysis and metabolite predictive models for HVPG response
These individual different metabolites are represented in Figure 1.
Logistic regression to develop a prognostic model for HVPG response was performed with combinations of two metabolites from the 18 finally selected. Several combinations of these metabolites showed a good performance (AUROC around 0.8), most of them including the NEFA 20:2(n-6) (eicosadienoic acid) ( Table 3).
Adding a third metabolite did not significantly improve the per- Similar results were also obtained with the other potential combinations of metabolites (Table 3).

| Association of selected metabolites with clinical variables: primary/secondary prophylaxis, Child-Pugh class, aetiology and concomitant medications
The 18 metabolites associated with the HVPG response to propranolol were further analysed for association with a priori clinically relevant variables: type of prophylaxis, Child-Pugh class, aetiology and concomitant medications. Secondary prophylaxis was associated with a worse response ( and predefined group of medications. Table S1 shows the effects of Child, prophylaxis and aetiology on metabolite model's coefficients and performance.

| D ISCUSS I ON
In this pilot study, we provide a simple predictive model to identify HVPG responders to acute iv propranolol based on metabolomic serum analysis. The current study reveals several lipid substances at significantly different concentrations between HVPG responders and non-responders, most of them non-esterified fatty acids and glycerophospholipids (plasmalogens). Several combinations of these metabolites showed a good discrimination for HVPG response. In

F I G U R E 1 Heatmap of individual metabolites at different concentrations between responders and non-responders.
Most of these metabolites were NEFA and glycerophospholipids (plasmalogens), which were at higher concentrations in responders agreement with previous studies, 2,3,5,6,12,14,28,29 no association between HVPG response and baseline variables, except for primary/ secondary prophylaxis patients was observed. However, this variable (or any clinical variable) did not add to the prediction of the metabolite model.
In a pragmatic approach, we decided to evaluate the model including a plasmalogen (PC(P-16:0/22:6)) and eicosadienoic acid, which showed a slightly (despite non-significant) better AUROC curve (AUROC 0.801). Using the Youden approach, 0.629 was the best cut-off value with a good overall performance maintaining a similar proportion of responders/non-responders well classified and misclassified ( Figure 3A). However, from a clinical perspective, misclassifying non-responders as responders may prevent these patients to be shifted to more effective (usually also more invasive) therapeutic alternatives. By contrast, responders who would be misclassified as non-responders could be "overtreated" and potentially exposed to therapeutic secondary effects. Taking all these considerations in mind, we decided to also propose a two cut-off approach with a higher capacity to identify HVPG (mostly non-responders) could be directly included. Patients falling in the "grey zone" (one-quarter of the cohort), would require the HVPG study. However, it would be possible to avoid the remaining 75% HVPG studies ( Figure 3B). This dual approach would also be interesting for centres with limited availability for HVPG measurements. There were many other significantly different metabolites according to response and our model was finally selected from a data-driven analysis though probably other potential models could also be useful ( Table 2). The present model may offer    31,32 This effect is mediated through β-adrenoceptors and the hormone-sensitive lipase. 33,34 The fact that non-responders showed lower NEFA levels, may reflect a lesser β-adrenergic stimulation or a metabolic resistance of the β-adrenoceptor that might explain that NSBB did not reach the desired effect. Glycerophospholipids have been proposed as protective agents in animal models of NASH and declining levels have been described in parallel to liver fibrosis progression. 35,36 Therefore, it might be hypothesized that lower levels of glycerophospholids observed in HVPG non-responders might be related to a more fibrogenic phenotype of cirrhosis, which could be less dependent and modifiable by vasoactive systems such as β-ad- In conclusion, the combination of two serum metabolites might help at identifying the HVPG response to acute iv propranolol in Seventeen patients fell in the grey zone patients with cirrhosis. The analysis of these metabolites could be a useful non-invasive tool to identify these patients though further validation of the model would be desirable.

ACK N OWLED G EM ENTS
We thank Rosa Sáez, Lara Orts, Àngels Baringo and Laura Rocabert for nursing support and expert technical haemodynamic assistance, and Clara Esteva for administrative assistance.

CO N FLI C T O F I NTE R E S T
The authors disclose no conflicts.