Biomarkers of Oncogenesis, Adipose Tissue Dysfunction and Systemic Inflammation for the Detection of Hepatocellular Carcinoma in Patients with Nonalcoholic Fatty Liver Disease

Simple Summary Circulating biomarkers for the detection of hepatocellular carcinoma in patients with dysmetabolic liver disease are an unmet need. In the present study, we observed that serum values of five biomarkers (namely, AFP, PIVKA-II, GPC-3, adiponectin and IL-6) were significantly different between patients with and without hepatocellular carcinoma; the best accuracy for the detection of tumor was achieved by PIVKA-II. Furthermore, we developed a model combining age, gender, PIVKA-II, GPC-3 and adiponectin that showed an excellent performance for the identification of patients with hepatocellular carcinoma. This model may be useful for the surveillance of patients with dysmetabolic liver disease at risk of hepatocellular carcinoma development. Abstract Current surveillance strategy for patients with nonalcoholic fatty liver disease (NAFLD) at risk of hepatocellular carcinoma (HCC) development is unsatisfactory. We aimed to investigate the diagnostic accuracy of alpha-fetoprotein (AFP), protein induced by vitamin K absence or antagonist-II (PIVKA-II), glypican-3 (GPC-3), adiponectin, leptin and interleukin-6 (IL-6), alone or in combination, for the discrimination between NAFLD patients with or without HCC. The biomarkers were investigated in a cohort of 191 NAFLD patients (median age 62, 54–68 years; 121 males and 70 females) with advanced fibrosis/cirrhosis, 72 of whom had a diagnosis of HCC. PIVKA-II showed the best performance for the detection of HCC with an area under the curve (AUC) of 0.853, followed by adiponectin (AUC = 0.770), AFP (AUC = 0.763), GPC-3 (AUC = 0.759) and by IL-6 (AUC = 0.731), while the leptin values were not different between patients with and without HCC. The accuracy of the biomarkers’ combination was assessed by a stratified cross-validation approach. The combination of age, gender, PIVKA-II, GPC-3 and adiponectin further improved the diagnostic accuracy (AUC = 0.948); the model correctly identified the 87% of the patients. In conclusion, we developed a model with excellent accuracy for the detection of HCC that may be useful to improve the surveillance of NAFLD patients at risk of tumor development.


Introduction
The epidemiological burden of nonalcoholic fatty liver disease (NAFLD) is rapidly increasing worldwide, with an estimated global prevalence of 25% in the general population [1].
This retrospective case-control study included patients with dysmetabolic induced-HCC and patients with NAFLD/NASH without HCC, recruited at the outpatient clinic of the Unit of Gastroenterology of A.O.U. Città della Salute e della Scienza di Torino-Molinette Hospital, Turin, Italy between November 2012 and January 2020.
For all patients, the inclusion criteria were age ≥18 years, histological diagnosis of NASH with advanced fibrosis/cirrhosis or clinical/radiological evidence of cryptogenic cirrhosis [22], which dysmetabolic etiology was assessed by the presence of metabolic risk factors (central obesity, T2DM, dyslipidemia and hypertension) [23] in the absence of other known causes of liver damage. For patients without HCC, a minimum of 1-year US follow-up after the collection of the serum sample was required. All patients signed written informed consent.
We excluded patients with a liver disease of other etiology, such as drug-induced liver disease, viral hepatitis and autoimmune, cholestatic and metabolic/genetic liver disease. Alcohol-induced liver disease was excluded by selecting patients with a negative history of alcohol abuse (weekly ethanol consumption <140 g for women and <210 g for men) [24].
The presence of advanced fibrosis/cirrhosis was histologically assessed and scored as described by Kleiner et al. [25] or clinically determined by means of a liver elastography (FibroScan ® , Echosens™, Paris, France) or hepatic US features and endoscopic signs of portal hypertension [26,27]. The diagnosis of HCC was achieved by histological examination or by contrast-enhanced imaging methods showing the radiological hallmark of HCC (i.e., the combination of hypervascularity in the late arterial phase and washout on portal venous and/or delayed phases), following the international guidelines [11]. HCC was classified according to the BCLC staging system (0 = very early, A = early, B = intermediate, C = advanced and D = terminal stage) [11].
Serum GPC-3 and plasma leptin values were measured by enzyme-linked immunosorbent assay (ELISA) using CanAg Glypican-3 EIA (Fujirebio Diagnostics AB, Gothenburg, Sweden) and Human Leptin Quantikine ® ELISA (R&D Systems, Minneapolis, MN, USA), according to the manufacturer's instruction. The GPC-3 serum levels were reported in pg/mL, while the leptin plasma levels were reported in ng/mL.

Statistical Analysis
Continuous variables were expressed as median and interquartile ranges (IQR), while categorical variables as number and percentages (%). The D'Agostino-Pearson test was used to test the data normality. The Mann-Whitney test and Kruskal-Wallis test were used to compare continuous variables between two or more groups, respectively. The correlation between continuous variables was assessed by Spearman's rank correlation coefficient (r s ). To evaluate the diagnostic accuracy of the circulating AFP, PIVKA-II, GPC-3, adiponectin, leptin and IL-6 alone, the AUC was assessed by receiver operating characteristic (ROC) curves analysis. A multivariate logistic regression analysis was performed to combine independent variables for the prediction of HCC. This analysis has been repeated using a cross-validation approach to compute its performances in predicting the HCC status, using the scikit-learn package in the Python environment. Specifically, the RepeatedStratifiedKFold with 5 splits and shuffling samples 20 times and the LogisticRegression functions were used to evaluate the performance of the classifier using a cross-validation approach, both with the option random_state = 0. The final performance of the model were computed, averaging the AUC values over the 100 test sets obtained with the above-described crossvalidated approach. Confidence intervals at a 95% confidence level were computed with a bootstrap approach by resampling with a replacement 1000 times.
A two-tailed p < 0.05 was considered statistically significant. The statistical analyses were performed using MedCalc software, version 18.9.1 (MedCalc bvba, Ostend, Belgium) and in-house scripts in Python programming language.

Circulating Biomarkers Values in the Study Cohort
The median serum levels of AFP, PIVKA-II, GPC-3, adiponectin and IL-6 were significantly higher in patients with HCC compared to those without tumor (all p < 0.001); only the plasma leptin values were not different between the two groups of patients (p = 0.649) ( Table 2 and Figure 1). p-values were calculated by the Mann-Whitney test. Abbreviations: alpha-fetoprotein (AFP), glypican 3 (GPC3), hepatocellular carcinoma (HCC) and interleukin-6 (IL-6).

Circulating Biomarkers Values in the Study Cohort
The median serum levels of AFP, PIVKA-II, GPC-3, adiponectin and IL-6 were significantly higher in patients with HCC compared to those without tumor (all p < 0.001); only the plasma leptin values were not different between the two groups of patients (p = 0.649) (Table 2 and Figure 1). p-values were calculated by the Mann-Whitney test. Abbreviations: alpha-fetoprotein (AFP), glypican 3 (GPC3), hepatocellular carcinoma (HCC) and interleukin-6 (IL-6). The AFP serum values showed a significant stepwise increase from patients with advanced fibrosis without HCC to patients with advanced tumor (BCLC = B, C and D) (p < 0.001). The AFP and PIVKA-II levels were significantly different between the patients with early HCC (BCLC = 0, A) and those with advanced tumor (p < 0.050), while among patients without HCC, the AFP, GPC3 and IL-6 serum levels were significantly different between The AFP serum values showed a significant stepwise increase from patients with advanced fibrosis without HCC to patients with advanced tumor (BCLC = B, C and D) (p < 0.001). The AFP and PIVKA-II levels were significantly different between the patients with early HCC (BCLC = 0, A) and those with advanced tumor (p < 0.050), while among patients without HCC, the AFP, GPC3 and IL-6 serum levels were significantly different between the patients with advanced fibrosis and those with cirrhosis (p < 0.050) (Table S1 and Figure S1).

Predictors of HCC and Model Development
Since the two groups of patients (i.e., patients with and without HCC) showed significant differences regarding the demographic, biochemical and clinical features, we performed a multivariate logistic regression analysis to assess the strength of the association with HCC. Age, gender, BMI, ALT, γGT, platelet count, albumin, total bilirubin, INR, cholesterol, triglycerides, AFP, PIVKA-II, GPC-3, adiponectin and IL-6 were considered for inclusion in the multivariate analysis. A logistic regression analysis was based on a stepwise approach keeping the variables at a significance level below 0.01 [28]. A Log transformation was made to AFP, PIVKA-II, GPC-3 and IL-6 due to data skewness. The variables retained in the model are reported in Table 4.

Predictors of HCC and Model Development
Since the two groups of patients (i.e., patients with and without HCC) showed significant differences regarding the demographic, biochemical and clinical features, we performed a multivariate logistic regression analysis to assess the strength of the association with HCC. Age, gender, BMI, ALT, γGT, platelet count, albumin, total bilirubin, INR, cholesterol, triglycerides, AFP, PIVKA-II, GPC-3, adiponectin and IL-6 were considered for inclusion in the multivariate analysis. A logistic regression analysis was based on a stepwise approach keeping the variables at a significance level below 0.01 [28]. A Log transformation was made to AFP, PIVKA-II, GPC-3 and IL-6 due to data skewness. The variables retained in the model are reported in Table 4. where age in years, 1 for males and 0 for females; the probability (p) of HCC is given by: p = 1 / (1 + e −y ).
diagnostic accuracy for the detection of HCC (AUC = 0.948), with a percentage of the patients correctly classified as 87% in the cross-validation (Figure 4). At the cut-off pHCC = 50%, the model showed Sp = 88.1%, Se = 86.9%, +LR = 9.00 and −LR = 0.15 for the detection of HCC.

Discussion
In the present study, we investigated the performances of different biomarkers involved in the oncogenic mechanisms of HCC in patients with NAFLD. Indeed, the biomarkers studied were selected on the premise that tumor development in such patients is driven by the concurrent activation of different oncogenic signaling pathways in accordance with the "multiple hits hypothesis" [29,30], whereby comorbidities, genetic determinants and environmental factors simultaneously contribute to NAFLD/NASH-HCC progression [31]. Interestingly, we observed that circulating biomarkers such as PIVKA-II and adiponectin displayed a good performance for the discrimination between patients with HCC and those without tumors; remarkably, the combination of demographic features (i.e., age and gender) together with oncogenic markers (i.e., PIVKA-II and GPC-3) and markers of adipose tissue dysfunction (i.e., adiponectin) allowed the development of a model showing an excellent performance for the detection of HCC.
Several different biomarkers have been studied in the last decades in order to improve and even personalize the surveillance of patients at high risk for HCC development [32][33][34]. Promising results have derived from comprehensive approaches that have allowed the detection of a wide spectrum of circulating molecules, including tumor proteins and nucleic acids (i.e., circulating tumor DNA/RNA) derived from the primary tumor [35], epigenetic biomarkers such as DNA methylation profiles and noncoding RNAs [36,37] and genetic variants recapitulated in polygenic risk scores [38]. On the other hand, the study of serologic biomarkers such as AFP and PIVKA-II has been pursued over time due to their inexpensiveness, analytical standardization, and acceptable performances [17].
In agreement with previous studies performed in patients with NAFLD [39,40], we observed a good diagnostic accuracy for PIVKA-II (AUC = 0.853) and a moderate performance for AFP (AUC = 0.763). Interestingly, GPC-3 showed a higher accuracy compared to the results of our previous study carried out on a cohort of patients with viral related-HCC [18]. Possibly, both the diverse etiology and clinical characteristics of the patients included may have accounted for the discrepancy observed. In particular, the different prevalence of cirrhosis in the control groups could have affected the performance of GPC-

Discussion
In the present study, we investigated the performances of different biomarkers involved in the oncogenic mechanisms of HCC in patients with NAFLD. Indeed, the biomarkers studied were selected on the premise that tumor development in such patients is driven by the concurrent activation of different oncogenic signaling pathways in accordance with the "multiple hits hypothesis" [29,30], whereby comorbidities, genetic determinants and environmental factors simultaneously contribute to NAFLD/NASH-HCC progression [31]. Interestingly, we observed that circulating biomarkers such as PIVKA-II and adiponectin displayed a good performance for the discrimination between patients with HCC and those without tumors; remarkably, the combination of demographic features (i.e., age and gender) together with oncogenic markers (i.e., PIVKA-II and GPC-3) and markers of adipose tissue dysfunction (i.e., adiponectin) allowed the development of a model showing an excellent performance for the detection of HCC.
Several different biomarkers have been studied in the last decades in order to improve and even personalize the surveillance of patients at high risk for HCC development [32][33][34]. Promising results have derived from comprehensive approaches that have allowed the detection of a wide spectrum of circulating molecules, including tumor proteins and nucleic acids (i.e., circulating tumor DNA/RNA) derived from the primary tumor [35], epigenetic biomarkers such as DNA methylation profiles and noncoding RNAs [36,37] and genetic variants recapitulated in polygenic risk scores [38]. On the other hand, the study of serologic biomarkers such as AFP and PIVKA-II has been pursued over time due to their inexpensiveness, analytical standardization, and acceptable performances [17].
In agreement with previous studies performed in patients with NAFLD [39,40], we observed a good diagnostic accuracy for PIVKA-II (AUC = 0.853) and a moderate performance for AFP (AUC = 0.763). Interestingly, GPC-3 showed a higher accuracy compared to the results of our previous study carried out on a cohort of patients with viral related-HCC [18]. Possibly, both the diverse etiology and clinical characteristics of the patients included may have accounted for the discrepancy observed. In particular, the different prevalence of cirrhosis in the control groups could have affected the performance of GPC-3. Indeed, cirrhosis is a preneoplastic condition characterized by genetic, epigenetic and molecular alterations not yet established in patients with chronic hepatitis but frequently observed in HCC [41][42][43]. As a matter of fact, the more we reduce the clinical differences between two groups of patients, the more the performance of the biomarker decreases.
Noteworthy, adiponectin showed a performance similar to AFP for the detection of HCC. A recent meta-analysis including nine studies and a total of 705 HCC patients and 1390 healthy controls showed that higher adiponectin levels were significantly associated with liver cancer (standard mean difference = 0.97, 95 %CI 0.02∼1.93, p < 0.05) [44]. Furthermore, in patients with HCV-related cirrhosis, higher serum adiponectin levels resulted in a predictor of HCC development [45,46] and liver-related mortality [47]. However, the mechanism by which adiponectin is involved in HCC development is not fully clear. Reduced adiponectin levels have been associated with metabolic syndrome [4,48]; nevertheless, studies in vitro and in the animal model have revealed an antiproliferative activity for adiponectin [49,50], suggesting that increased adiponectin levels might have a protective role in HCC. Consonant with this hypothesis, the administration of adiponectin (5 µg/kg weekly) blocked tumor progression in a thioacetamide-induced rat HCC model, resulting in an 80% increased survival rate, 73% reduced average number of nodules and 46% decreased serum AFP [51]. Further studies are needed to confirm the hepatoprotective and chemoprotective effects of adiponectin against HCC.
Finally, the combination of different biomarkers with clinical and/or demographic characteristics into simple prediction models allowed the further improvement of the diagnostic accuracy for HCC detection [52]. However, the majority of these models have been developed and validated in cohorts of patients chronically infected with HBV or HCV [28,[53][54][55]; to the best of our knowledge, only the GALAD score was tested in the setting of NAFLD, showing a high performance for the identification of patients with HCC [56]. Here, we developed a model including age, gender, PIVKA-II, GPC-3 and adiponectin that showed a high diagnostic accuracy in cross-validation (AUC = 0.948) for the detection of HCC in patients with dysmetabolic liver disease; the model allowed to correctly identify 87% of the patients included in the study. Given the relatively small number of patients enrolled and the lack of a validation cohort, we applied a machine learning approach based on a stratified cross-validation to assess the performance of the model; accordingly, the original sample was partitioned into a training set to train the model and a test set to evaluate its performance, and the procedure was repeated multiple times. As a result, the model denoted a high accuracy, with a low risk of overfitting, and a generalizability of the independent datasets. However, further multicenter studies are needed to independently validate these findings. Furthermore, a cost-effective analysis may be useful to assess the benefits produced by the implementation of the model in the clinical setting with respect to its cost.

Conclusions
Our data confirmed the good diagnostic accuracy of PIVKA-II for the detection of HCC in patients with NAFLD. Furthermore, the combination of age, gender, PIVKA-II, GPC-3 and adiponectin allowed a noticeable improvement in the detection of HCC compared to each single biomarker used alone. These findings need to be validated in a prospective surveillance setting in order to assess the ability of the model to predict the HCC occurrence in patients with dysmetabolic liver disease at risk of tumor development.

Supplementary Materials:
The following are available online at https://www.mdpi.com/article/ 10.3390/cancers13102305/s1: Figure S1: Median values of AFP (A), PIVKA-II (B), GPC-3 (C), adiponectin (D), leptin (E) and IL-6 (F), according to the different stages of liver disease, Figure S2. Diagnostic accuracy of AFP, PIVKA-II, GPC-3, adiponectin and IL-6 for the detection of HCC in lean (A) and obese patients (B), and in normal-glucose tolerant (C) and diabetic patients (D), Figure  S3. Diagnostic accuracy of AFP, PIVKA-II, GPC-3, adiponectin and IL-6 for the detection of early HCC, Table S1: Median biomarker values according to the different stages of liver disease, Table  S2: Comparison of the diagnostic accuracy of AFP, PIVKA-II, GPC-3, adiponectin and IL-6 for the detection of HCC.