Patient details

Recruitment of excessive drinkers from alcohol treatment centres and liver clinics was conducted at the four major hospitals (Royal Prince Alfred, Concord Repatriation General, Liverpool and Fairfield) in Sydney, Australia. Written informed consent was obtained from all participants. The study was approved by the Royal Prince Alfred Hospital Ethics Committee (HREC/11/RPAH/88).

Excessive drinkers are defined as having daily alcohol consumption of ≥80 g per day (men) and 50 g per day (women) for ten years or more. Cases (n = 31) have clinical and/or histopathology evidence of ALC, with exclusion of hepatitis B or C, autoimmune liver disease, haemochromatosis and Wilson’s Disease (S1 Table). Controls (n = 28) have normal results for AST, ALT, total bilirubin, albumin, platelet count and INR while actively drinking or within 7 days of stopping the most recent episode of heavy alcohol use and no clinical and/or biopsy evidence of significant liver disease (S1 Table). HIV infection is an exclusion criterion for both Cases and Controls.

Clinical data from consenting participants included alcohol and tobacco history (amount, type, duration, abstinence, parental history), other drug use, demographics (age, gender, ethnicity, education, marital status, medication), blood biochemistry (liver function tests, bilirubin, blood clotting by international normalised ratio (INR), creatinine), clinical evidence of liver injury (ascites, encephalopathy, variceal bleeding), anthropometrics and Alcohol Use Disorder Identification Test [17]. Patient characteristics for the two groups are presented in Table 1. Serum was prepared from non-fasting blood and stored at -80°C. The majority of the samples were non-fasting due to the unpredictable nature of patients presenting at the drug treatment centre, that did not allow strict regimens of blood collection times and fasting status.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Patient characteristics.

https://doi.org/10.1371/journal.pone.0130346.t001

Lipid extraction

Lipids were extracted from serum as described previously [16]. In brief, 200μL of a 2:1 choloroform:methanol mix, together with 10 μL of the internal standard mix was added to 10μL serum. The samples were vortexed for 10mins, incubated in an ultrasonic water bath at 20°C for 30mins and rested for 20mins at ambient temperature. After centrifuging the samples (13,000 g, 10mins) supernatants containing extracted lipids were dried under N₂. The dried lipids were resuspended in 50μL water-saturated butanol and 50μL 10mM NH₄COOH in methanol. After centrifuging (4000 g, 5mins) supernatants were transferred into HPLC glass vials with 0.2mL micro-inserts and Teflon insert caps.

Lipidomic analysis

Liquid chromatography electrospray ionisation tandem mass spectrometry was conducted as previously described [16]. Briefly, online LC-MS/MS was performed using a Zorbax C18, 1.8 μm, 50 x 2.1 mm column (Agilent Technologies) connected to the ionization source of an API4000 QTRAP mass spectrometer (Applied Biosystems). Liquid chromatography was performed using the following gradient conditions; 0–100% buffer B over 8.0 min, 2.5 min at 100% B, a return to 0% B over 0.5 min then 3.0 min at 0% B prior to the next injection. Diacylglycerols and triacylglycerols were separated using the same solvent system with an isocratic flow (100 μL/min) of 85% B. Solvents A and B consisted of tetrahydrofuran:methanol:water (30:20:50 and 75:20:5 respectively), both containing 10 mM NH₄COOH. Measurement of individual lipid species was performed using scheduled multiple-reaction monitoring in positive ion mode. Lipid measures were calculated by relating the peak area of each species to the peak area of the corresponding internal standard. Cholesteryl ester species were corrected for response factors determined for each species. Total measured lipids of each class were calculated by summing the individual lipid species.

Statistical Analysis

Prior to regression analysis, data for each lipid species, class or subclass was standardised to the interquartile range (IQR). This facilitates the comparison of odds ratios or beta coefficients between lipid species with varying concentration ranges. Binary logistic regression analysis adjusted for age and BMI was performed to determine the association of lipid classes, subclasses and individual lipid species with ALC case-control status. The resultant IQR odds ratio for a given lipid measurement represents the number of times an individual with a lipid measurement at the 75^th percentile is more likely to have ALC than an individual with a lipid measurement at the 25^th percentile. Linear regression adjusted for age and BMI was used to determine the linear relationship between lipids and the model for end stage liver disease (MELD) score or bilirubin level. The beta-coefficients obtained represent the change in outcome measure (MELD score or bilirubin level) associated with an interquartile range increase in the lipid measurement. For all analyses, p-values were corrected for multiple comparisons using the Benjamini-Hochberg method [18] and statistical significance was determined by a corrected p-value <0.05.

The classification performance of lipid species and clinical characteristics were evaluated using machine learning methods. Along with lipid species, the following clinical characteristics were used: sex, BMI, systolic blood pressure, diastolic blood pressure, white blood count, haemoglobin, platelet count, albumin, bilirubin, creatinine, INR, aspartate aminotransferase to alanine aminotransferase ratio, gamma-glutamyl transferase, cholesterol, triglyceride, LDL-C, HDL-C, daily, yearly, and life time alcohol consumption. As age was significantly different between the ALC and alcoholic without liver cirrhosis (AwLC) groups, all the lipid species measurements and clinical characteristic values were adjusted for age based on the residuals from linear regression. After thresholding all the negative values to one, all the adjusted values were log-transformed.

To classify ALC from AwLC, we performed feature selection, model training and testing within a 3-fold stratified cross-validation framework (repeated 200 times). In each iteration, the order of inclusion of features into the model was determined by the reliefF algorithm [19]. Model development involved the implementation of a linear kernel-based support vector machine classifier obtained from the libraries of LIBSVM 2.84 [20]. The performance the model was assessed on the test dataset by computing the C-statistic (Area Under the Curve (AUC)), accuracy, specificity, sensitivity and net reclassification index for the first 30 features included in the model. Finally, the mean and 95% confidence intervals of each performance measure over 600 iterations were calculated.

Biological relevance of lipids associated with alcoholic liver cirrhosis

It is long known that alcohol consumption is associated with changes in all lipid components of plasma [21]. Alcohol enhances fatty acid synthesis, driving triglyceride synthesis, and hinders VLDL secretion leading to the accumulation of triglycerides in lipid droplets within the hepatocytes [22]. Alcohol induced fat accumulation in the hepatocytes can trigger an inflammatory response giving rise to steatohepatitis and increases the risk of progression to fibrosis and cirrhosis in some individuals, leading to serious health consequences. Identification of these individuals in the early stages of the disease would enable intervention prior to irreparable liver damage. Lipidomic technologies, now allow for the detailed analysis of individual lipid species associated with disease and their evaluation as potential biomarkers.

There are two dimensions to lipid biomarkers associated with liver disease; how far the disease has progressed (normal liver/steatosis/various stages of fibrosis/cirrhosis); and the cause of the disease (viral/NAFLD/alcohol etc.). In our study, the comparison was between individuals with ALC and AwLC, thus we are examining the more advanced stage of disease after the onset of alcoholic steatosis and establishment of cirrhosis. Our observations of several lipid classes and species showing positive associations with ALC, relative to AwLC, highlights the extensive metabolic abnormalities associated with cirrhosis. The strong positive association of both dihexosyl- and trihexosylceramide with cirrhosis contrasts with our previous observations that these glycosphingolipids are negatively associated with obesity [16] as well as with prediabetes and type 2 diabetes [15] all situations where steatosis is likely to be a contributing factor. A similar effect was observed for the alkylphosphatidylcholine and lysoalkylphosphatidylcholine that also showed positive associations with ALC (relative to AwLC) but were previously observed to have negative associations with obesity, prediabetes and type 2 diabetes [15,16]. However, while early studies showed chronic alcohol increased hepatic activity of the enzymes choline phosphotransferse and phosphatidylethanolamine methyltransferase involved in phosphatidylcholine synthesis [23] we observed an odds ratio of only 0.60 for phosphatidylcholine with ALC (non-significant) indicating that a general increase in phosphatidylcholine synthesis is not associated with progression to ALC and that this was restricted to only the ether linked phosphatidylcholine species.

We also observed negative associations of diacyl- and triacylglycerol species with ALC, with IQR odds ratios of 0.32 (95% CI 0.11–0.88) and 0.33 (95% CI 0.13–0.88) respectively, although these were non-significant after correction for multiple comparisons. Similar observations in a baboon model of alcoholic liver disease, showed that microsomal diacylglycerol acyltransferase and cytosolic phosphatidate phosphohydrolase activities increased with early stages of fatty liver which disappeared with progression of liver injury [24]. The clinical measure of total triglycerides also showed the same trend (Table 1). These multiple associations with ALC that appear to oppose those associations with steatosis and early liver disease, may relate to a decrease in free fatty acids resulting from a decrease in lipogenesis as the liver progresses to cirrhosis and the subsequent damage reduces the synthetic capacity of the liver [25] including fatty acid synthesis. These differences may also be a consequence of additional interactions of alcohol metabolites, acetaldehyde and acetate, that can form adducts thereby interfering with lipoprotein catabolism. In the present study, it should also be noted that these metabolic differences may arise as a result of some ALC patients undergoing treatment/awaiting transplant and so may not be actively drinking at the time of sample collection.

Interestingly, phosphatidylinositol, was also positively associated with cirrhosis demonstrating that not all lipid metabolic pathways are influenced in the same way. Phosphatidylinositol, is a major source of arachidonic acid, a precursor to the proinflammatory eicosanoids and so the positive association of this lipid class with both the early indicators of steatosis previously reported [15,16] and with the progression to cirrhosis, shown here, supports the role of the inflammatory state contributing to the disease progression.

Notwithstanding the opposing associations of several metabolic pathways in the ALC group, the elevated levels of dihexosylceramide may have specific effects on gene regulation and lipid metabolism leading to the increase in free cholesterol associated with cirrhosis. A recent paper by Chatterjee et al. examined the effect of inhibiting glycosphingolipid synthesis on atherosclerosis and arterial stiffness [26]. They observed that inhibition of glycosphingolipids synthesis resulted in a decrease in monohexosyl- and dihexosylceramide in the liver and this was associated with a decrease in serum cholesterol and triglycerides via recruitment of multiple genes/pathways of lipid metabolism. Specifically, inhibition of glycosphingolipids synthesis increased LDL receptor and SREBP2 gene expression and decreased HMGCoA reductase gene expression, suggesting inhibition of cholesterol biosynthesis and increased LDL uptake. The decreased monohexosyl- and dihexosylceramide were also associated with increased expression of genes responsible for cholesterol efflux including ABCA1, ABCG5 and ABCG8. This inhibitory effect of lactosylceramide (a dihexosylceramide) in cholesterol efflux via the ABCA1/Apolipiprotein A-1 pathway has also been demonstrated in cultured fibroblasts [27]. The decreased levels of lactosylceramide were also associated with the up regulation of the Cyp7A1 gene which encodes a 7-hydroxylase enzyme that converts cholesterol to bile acid for excretion [26]. Thus, lactosylceramide appears to influence multiple lipid metabolic pathways leading to elevated cholesterol within the liver. Elevated cholesterol within hepatic stellate cells has been shown to increase fibrosis in ACAT-1 deficient mice [28] further supporting a causal role for lipid metabolism in ALC. Dihexosylceramide then, is emerging as a key mediator of fibrosis and may play an important role in the regulation of cholesterol metabolism leading to cirrhosis. This raises the dual possibility of utilising these lipids as early indicators of cirrhosis and of targeting this metabolic pathway to prevent progression to cirrhosis. It should be noted that while the underlying dysregulation of lipid metabolism likely plays an important role in ALC, whether these metabolic changes are also present in other forms of cirrhosis remains to be determined.

Classification of alcoholic liver cirrhosis

Despite our incomplete knowledge of the metabolic dysregulation leading to the serum lipid profiles associated with ALC we were able to evaluate these lipids as potential biomarkers to identify those individuals with established liver disease. We were clearly able to show that serum lipid species can improve on traditional clinical characteristics in classification models. The model using only clinical characteristics gave a maximum AUC and accuracy with four clinical characteristics, with platelet count, albumin, INR and haemoglobin being most often included in these models. While models developed with lipid species alone gave similar results, the combination of lipid species with the clinical characteristics resulted in models where the maximum performance was obtained with a model containing 27 features, with platelet count, INR and haemoglobin being incorporated into these models along with multiple species of lipids (Table 4). The net reclassification index (NRI) of the mixed model relative to the model using clinical characteristics only was 4.1% which while modest, highlights the potential of lipids to add to clinical characteristics. Although many lipids were associated with bilirubin, the lipids were preferentially incorporated to the model, over bilirubin, demonstrating their independent value to the model.

A recent report on a new classification model for hepatitis C virus related cirrhosis included plasma lathosterol in a multivariate model with BMI, platelet count and prothrombin index to provide a similar AUC of 0.91 (95% CI 0.82–1.0) [29] compared to our value of 0.892 (95% CI 0.887, 0.897). Interestingly the negative association of lathosterol in this study suggests a down regulation of cholesterol synthesis in hepatitis C virus related cirrhosis as opposed to the apparent increase in ALC.

Study Limitations

Our long term goal is to identify early biomarkers of alcoholic liver disease, which could allow intervention before development of cirrhosis. We compared lipid profiles from AwLC patients (controls) to those with advanced alcoholic cirrhosis to determine lipid biomarkers for the identification of ALC within the high risk alcohol consuming group to provide the clearest lipid signature characteristic of disease progression. The current cohort is of moderate size and consists of patients with advanced ALC compared to AwLC. The large clinical difference between these groups may have resulted in inflated odds ratios because of spectrum bias [30], for some lipid species. We believe the inclusion of early stage alcoholic liver disease patients in future studies and the analysis of prospective cohorts will further strengthen our results. Validation of lipid species identified in this study in patients with early stages of ALC is warranted as are studies to assess other forms of liver disease with different etiologies. For classification modelling, we have employed a 3-fold cross-validation framework repeated 200 times to resample the data to overcome the limited sample size available. We have assessed the performance of the classification models by computing a number of performance measures and their confidence intervals. The gain in AUC and NRI values demonstrate the value of lipid species as features in the classification model over clinical characteristics. We were however, unable to obtain an independent validation cohort to confirm the performance of the classification model.