bs_bs_banner
doi:10.1111/jgh.12545
H E PAT O L O G Y
Serum peptide pattern that differentially diagnoses hepatitis B virus-related hepatocellular carcinoma from liver cirrhosis Na Wang,*,1 Yuan Cao,*,†,1 Wei Song,*,‡,1 Kun He,* Tao Li,* Jie Wang,* Bin Xu,* Hai-yan Si,† Cheng-Jin Hu† and Ai-Ling Li* *Institute of Basic Medical Sciences, National Center of Biomedical Analysis, Beijing, †Department of Laboratory Medicine, 90th General Hospital of Jinan, Jinan, Shandong, and ‡Department of Hematology and Oncology, The First Hospital of Jilin University, Changchun, Jilin, China
Key words differential diagnosis, hepatocellular carcinoma, liver cirrhosis, serum peptide. Accepted for publication 18 January 2014. Correspondence Ailing Li, Institute of Basic Medical Sciences, National Center of Biomedical Analysis, 27 Tai-Ping Road, Beijing 100850, China. Email: [email protected] Cheng-Jin Hu, Department of Laboratory Medicine, 90th General Hospital of Jinan, 25 Shifan Road, Jinan, Shandong 250031, China. Email: [email protected] 1
These authors contributed equally to this work. Conflict of interest: No.
Abstract Background: Although alpha-fetoprotein (AFP) is a useful serologic marker of hepatocellular carcinoma (HCC), it is not sufficiently sensitive to differentiate HCC and liver cirrhosis (LC) caused by hepatitis B virus (HBV) infection. Aims: The aim is to discover novel noninvasive specific serum biomarkers for the differential diagnosis of HBV-related HCC and LC. Methods: With a highly optimized peptide extraction and matrix-assisted laser desorption/ionization time of flight/time of flight mass spectrometric approach, we investigated serum peptide profiles of 80 HCC and 67 LC patients. Three supervised machine learning methods were employed to construct classifiers. Receiver operator curves were plotted to evaluate the performance of classifiers. Results: With a support vector machine-based strategy, we picked nine peaks with m/z ratios of 819.49, 1076.14, 1341.72, 2551.44, 3156.44, 3812.88, 4184.26, 4465.92, and 4776.41 to construct the classifier. We proposed a novel method for distinguishing HCC from cirrhosis, based on a multilayer perceptron (MLP) method. We obtained a sensitivity of 90.0%, specificity of 79.4%, and overall accuracy of 85.1% on an independent test set. The combination of the MLP model and serum AFP level outperformed serum AFP marker alone in distinguishing HCC patients from LC patients. In this experience, sensitivity increased from 62.5% to 87.5%, and specificity increased from 79.4% to 88.2%. Conclusions: Our results indicate that the MLP model is a novel and useful serum peptide pattern for distinguishing HCC and LC. The peptidome signature alone or together with serum AFP determination may be a more effective method for early diagnosis of HCC in patients with HBV-related LC.
Introduction Hepatocellular carcinoma (HCC) is one of the most common malignancies worldwide, accounting for approximately 600 000 deaths per year.1–4 The incidence of HCC is rising over the past two decades. Over 80% of HCC develop from liver cirrhosis (LC), mainly related to chronic hepatitis B virus infection. Early diagnosis of HCC can greatly improve the outcome, with better longterm survival and reduced recurrence risk for operative treatment. Because of the lack of effective means of early diagnosis, only 30–40% of patients with HCC are candidates for potentially curative treatments at the time of diagnosis.5–7 Therefore, the differential diagnosis between HBV-related HCC and LC is very important. At present, serum alpha-fetoprotein (AFP) is available for the diagnosis and monitoring of HCC. However, a considerable proportion of HCC patients do not have a significant elevation of serum AFP. Because of its low sensitivity, AFP is not an ideal 1544
tumor marker. In addition, imaging techniques (ultrasound, computed tomography, and magnetic resonance imaging) are widely used for HCC diagnosis. At earlier stages, when the tumor is less than 2 cm in size, the imaging techniques are seldom diagnostic for HCC.6,8,9 Accordingly, discovery of novel noninvasive, effective, and reliable serum biomarkers for the differential diagnosis of HBV-related HCC and LC would play a pivotal role in improving the prognosis of patients with HCC. The search for cancer biomarkers has recently turned, at least partially, from cancer-specific proteins, which are rare and usually disappointing in their sensitivity, to circulatory peptides. The development of mass spectrometry (MS) technology now permits the display of hundreds of small- to medium-sized peptides from several microliters of serum in a high-throughput manner.10 Using a precision instrument of matrix-assisted laser desorption/ ionization time of flight MS (MALDI-TOF MS), recent studies have identified cancer-specific peptide markers.11–14 However, the
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
N Wang et al.
serum peptidome study for distinguishing HCC and LC by MALDI-TOF/TOF MS analysis has not been performed. In this study, we used a MALDI-TOF/TOF MS strategy to investigate serum peptide profiles of 80 HCC and 67 LC patients. Three supervised machine learning methods were employed to construct classifiers. Using the multilayered perceptron (MLP) model, we obtained a sensitivity of 90.0%, specificity of 79.4%, and overall accuracy of 85.1% on an independent test set. This model alone, or together with the serum AFP, was more effective than serum AFP detection for the early diagnosis of HCC in patients with HBV-related LC.
Materials and methods Study overview. Study overview was summarized in Figure 1. Serum samples of 80 HCC and 67 LC patients were analyzed by MALDI-TOF-MS for peptide expression. Independent training and test sets were created, with similar representation of age and gender in each set. A support vector machine (SVM)based strategy was used to select peaks to differentially diagnose HCC from LC. With the selected peaks, three supervised machine learning methods were employed to construct classifiers based on the training set to distinguish HCC from LC. The test set was used to evaluate the three classifiers. We also evaluated the performance of the candidate classifier combined with serum AFP. Serum samples. Serum samples of 147 patients, including 80 HCC patients and 67 LC patients, were collected from the 90th hospital of Jinan in China. Tumor stage was defined according to the Barcelona Clinic Liver Cancer (BCLC) staging system;15 for the purpose of this study, we classified tumors with BCLC stage 0 + A as early-stage HCC.16 The clinical characteristics of HCC and LC were summarized in Supplemental Table S1. All patients were HBV infected, with a Child–Pugh class A to C. Blood biochemistry, AFP assay, computed tomography, and liver biopsy were performed on all patients. The study was approved by the
Specific serum peptides of liver cancer
local institutional review board and ethical committee. Informed consent was obtained from all patients. All samples were collected using a standard clinical protocol. The blood samples were collected from patients in the morning, allowed to clot for 2 h at room temperature, and centrifuged for 15 min at 1500–2000 rpm. Sera were then frozen at −80°C. For the reproducibility experiments, sera from 10 HCC patients were pooled. Six within-run assays were then performed on the MALDI-TOF/TOF MS instrument (Supplemental Table S2). The mean CV of within-run assays was 18.7% (range: 16.5–25.7%). Peptide extraction. Copper-chelated magnetic beads were used for extracting peptides from serum. For each sample, 5 μL serum was mixed with 5 μL beads and 20 μL of binding solution. Samples were washed three times using 100 μL of washing solution, then 20 μL of eluent buffer was added. A 1 μL of the eluent and 1 μL α-cyano-4-hydroxycinnamic acid matrix solution were mixed and spotted onto a 600-um-diameter spot size 384 MTP target plate (Bruker Daltonik, Billerica, MA, USA). The peptide calibration standard with the same matrix was applied to target spots for external calibration of the instrument. The processed samples were analyzed using MALDI-TOF/TOF MS (ultraflex III, Bruker Daltonics), equipped with a pulsed ion extraction ion source. MALDI-TOF/TOF-MS data preprocessing and peak selection. After curve smoothing in all sample spectra and subtracting the baseline, the peaks with m/z between 800 Da and 10 000 Da and with a signal-to-noise ratio over 5, were labeled using FlexAnalysis 2.4 software (Bruker Daltonik). A peak cluster was created including m/z value and signal intensities and then exported together into excel file format. An in-house MatchPeaks software was used to normalize and align all sample peak intensities. These peaks were fit in a spreadsheet for further statistical analysis. The SVM-based strategy was used to select peaks to identify their discriminatory power.
Figure 1 Study overview. This diagram shows the route to select distinguishing peptides and processed models for HCC and LC diagnosis. The numbers indicate total number of selected features at that stage of the study.
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
1545
Specific serum peptides of liver cancer
N Wang et al.
Classifier construction and evaluation. To distinguish LC and HCC, three supervised machine learning methods (MLP, sequential minimal optimization [SMO], and classification and regression tree [CART]) were employed to construct classifiers in Waikato Environment for Knowledge Analysis.17 In this study, independent training (n = 73) and test sets (n = 74) were created, with similar numbers of LC and HCC, and similar representation of age and gender. With these sets, we assessed the performance of the three classifiers by considering the number of correctly classified (true positives and true negatives) and incorrectly classified (false positives and false negatives) samples in the test set. Accuracy, sensitivity, and specificity were calculated. Receiver operator curves (ROC) were plotted to evaluate the performance of a range of classifiers. Area under curve (AUC) was given for each curve. StAR18 (Molecular Bioinformatics Laboratory, Pontificial Catholic University of Chile, Santiago, Chile) was used for statistical comparison of AUCs.
Results Dataset and feature selection. To screen serum peptides for HCC and LC, we used MALDI-TOF-MS to analyze serum samples from 80 HCC and 67 LC patients. About 70% of subjects serum spectra were used as training set (n = 73), and the rest were used as test set (n = 74; Fig. 1, Table 1). A total of 235 peaks with m/z between 800 Da and 10 000 Da were obtained from the 147 patients. After all peaks intensity were normalized using house MatchPeaks software (MS2PCA), feature variables were evaluTable 1
Construction and evaluation of the three classifiers using supervised machine learning methods. To discriminate HCC from LC, we constructed three classifiers using supervised machine learning methods MLP, SMO, and CART. To evaluate the performance of the three classifiers, we applied a preliminary test using 10-fold cross-validation on the training set. The accuracy, sensitivity, and specificity of MLP were 85.1%, 90%, and 79.4%, respectively. SMO gave comparable results with an accuracy of 75.7%, sensitivity of 82.5%, and specificity of 67.6%. CART gave the worst results (accuracy of 50%, sensitivity of 62.5%, and specificity of 35.3%). The ability of a classifier to discriminate data correctly in the test set was known as its generalization performance.19 We compared the generalization performance of a series of classifiers by plotting their performance on the test set in ROC space (Fig. 2, Supplemental Table S3). The MLP classifier exhibited the best results on the test set, with a sensitivity of 90.0% and specificity of 79.4% (overall accuracy of 85.1%). The SMO classifier showed similar sensitivity (82.5%) and poor specificity (67.6%) on the test set. CART gave the worst results (accuracy of 50%, sensitivity of 62.5%, and specificity of 35.3%). The AUC was 0.909, 0.751, and 0.582 for MLP, SMO, and CART, respectively. To evaluate the performance of MLP model, ROC of each peptide was plotted as follows (Fig. 3). AUC
Training set and test set
Number of HCC patients Number of LC patients Gender (male : female) Age (years) AFP (ng/mL) ALT (U/L) AST (U/L) AKP (U/L) Child–Pugh A B C Tumor number and size Single nodule ≤ 2 cm Single nodule 2–3 cm Three nodules ≤ 3 cm AJCC stage I II BCLC score 0 A
Training set (n = 73)
Test set (n = 74)
P value
40 33 48:25 54.2 ± 11.8 104.9 ± 164.3 75.4 ± 123.6 71.8 ± 68.8 146.7 ± 87.2
40 34 54:20 53.8 ± 11.4 90.6 ± 168.4 75.1 ± 103.5 94.8 ± 178.4 128.4 ± 90.5
0.928
51 16 6
50 18 6
23 15 2
25 15 0
22 18
24 16
18 22
19 21
0.441 0.557 0.630 0.607 0.120 0.461 0.860
0.617
0.651
0.823
AFP, alpha-fetoprotein; AJCC, American Joint Committee on Cancer; AKP, alkaline; ALT, alanine aminotransferase; BCLC, Barcelona Clinic Liver Cancer; HCC, hepatocellular carcinoma; LC, liver cirrhosis; NA, not available.
1546
ated using the SVM-based strategy and ranked by the square of the weight assigned by the SVM. Nine peptides were selected for classifier construction, which had m/z ratios of 819, 1076, 1341, 2551, 3156, 3812, 4184, 4465, and 4776.
Figure 2 Classifier performance in receiver operator curves (ROC) space. Red line shows ROC curve of multilayered perceptron (MLP). Blue line represents ROC curve of sequential minimal optimization (SMO). Green line represents ROC curve of and regression tree (CART). The area under curve is 0.900, 0.748 and 0.559 for MLP, SMO and , MLP; , SMO; , CART. CART, respectively.
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
N Wang et al.
Specific serum peptides of liver cancer
Figure 3 Performance of single feature peptide in receiver operator curves (ROC) space. The lines show ROC curve of single peptide of the nine peptides. The area under curve is 0.705, 0.629, 0.519, 0.674, 0.739, 0.696, 0.684, 0.675, and 0.661 for the peptide m/z 819, 1076, 1341, 2551, 3156, 3812, 4184, 4465, and 4776, respectively.
of each feature peptide was also calculated. None of AUCs of these proteins were better than that of MLP, which proven the performance of the MLP classifier.
The combination of MLP model analysis and AFP for the differential diagnosis of HCC and LC. Serum levels of AFP are currently applied for screening HCC. When the
level of AFP is over 20 ng/mL, the patient is considered to have HCC. We assayed the serum levels of AFP in all the HCC and LC patients from the training set and test set. When the cutoff value of serum AFP was 20 ng/mL, the outcome indicated the sensitivity of 62.5% and specificity of 79.4% on the test set. The MLP model gave a sensitivity of 90% and specificity of 79.4%. The sensitivity in MLP model was much higher than AFP detection. Importantly, combining MLP model and AFP detection, a sensitivity of 87.5%
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
1547
Specific serum peptides of liver cancer
N Wang et al.
was achieved, with a specificity of 88.6% (Table 2). To compare the accuracy among MLP model, AFP level, and combination of MLP model and AFP detection (MLP + AFP), we plotted ROC of the three methods. AUC was also calculated for each curve (see Fig. 4 and Table 2). Based on our test set, MLP + AFP gave the best result with AUC of 0.943, while MLP showed the comparable AUC of 0.909. However, there is no significant difference (P = 0.0561) on AUCs between the two algorithms by using StAR.18 The AUC of AFP is 0.751, which is similar to the report by Attallah.20 Obviously, the AUC of AFP is less than those of MLP or MLP + AFP (P < 0.05). Identification of feature peptides. Peptide sequencing was performed to identify the nine feature peptides in the sera of HCC patients and LC patients. Five peptides were positively identified by FT MS/MS, and the Fourier transform (FT) sequencing result is shown in Table 3. For reference, m/z 1076 and 2551 stand for peptides of Fibrinogen α, m/z 3156 for a peptide of Transthyretin, m/z 4465 for a peptide of Antichymotrypsin (ACT), and m/z 4776 for a peptide of α-1 antitrypsin.
Figure 4 Performances of multilayered perceptron (MLP), alphafetoprotein (AFP) and MLP + AFP in receiver operator curves (ROC) space. Blue line shows ROC curve of MLP. Red line represents ROC curve of MLP + AFP. Green line represents ROC curve of AFP. The area under curve is 0.943, 0.909, and 0.751 for MLP + AFP, MLP, and AFP, , MLP + AFP; , MLP; , AFP. respectively.
Table 2 Diagnostic performances of AFP, MLP and AFP + MLP on the test set Classifiers
TP
TN
FP
FN
ACC (%)
SE (%)
SP (%)
AUC
MLP MLP + AFP AFP
36 35 25
27 29 27
7 5 7
4 5 15
85.1 86.5 71.6
90.0 87.5 62.5
79.4 88.2 79.4
0.909 0.943 0.751
ACC, accuracy; AUC, area under curve; AFP, alpha-fetoprotein; FN, false negatives; FP, false positives; MLP, multilayered perceptron; SE, sensitivity; SP, specificity; TN, true negatives; TP, true positives.
Table 3
Discussion In this study, we established an MLP model to differentiate HCC and LC patients with a sensitivity of 90.0% and specificity of 79.4%. When the cutoff value of serum AFP was 20 ng/mL, the sensitivity was 62.5%, and the specificity was 79.4%. The combination of the MLP model and serum AFP level outperformed serum AFP alone in distinguishing HCC from LC. Using this combination, the sensitivity increased from 62.5% to 87.5%. To construct classifiers for distinguishing HCC from cirrhosis, three supervised machine learning methods were used. The classification accuracy was 85.1%, 75.7%, and 50% for MLP, SMO, and CART, respectively. To avoid biases, we applied our trained classifiers to the independent test set. The fact that a small increase in accuracy and sensitivity was observed in the test set demonstrates the generalization performance of the classifiers. After regulating the probability threshold of the MLP, we obtained a best prediction result on the test set. Eventually, the MLP was selected as our classifier in this study (Supplemental Table S3, Fig. 2).
Identification of sequences of proteomic features differentially expressed between the HCC and LC patients by FT-ICR-MS
m/z
Peptide name
Peptide sequences
819.49 1076.14 1341.72 2551.44 3156.44 3812.88 4184.26 4465.92 4776.41
Not identified Fibrinogen a Not identified Fibrinogen a Transthyretin Not identified Not identified Antichymotrypsin α-1 antitrypsin
/ GDFLAEGGGVR / SSSYSKQFTSSTSYNRGDSTFES DSGPRRYTIAALLSPYSYSTTAVVTNPKE / / LVETRTIVRFNRPFLMIIVPTDTQNIFFMSKVTNPKQA LEAIPMSIPPEVKFNKPFVFLMIDQNTKSPLFMGKVVNPTQK
FT-ICR-MS, Fourier transform ion cyclotron resonance mass spectometry; HCC, hepatocellular carcinoma; LC, liver cirrhosis.
1548
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
N Wang et al.
Recently, with the development of artificial intelligence, artificial neural networks have been widely used in the area of medical data mining. MLP is a feedforward artificial neural network model that maps input data sets onto a set of appropriate output. This method can use a large number of simple units to process information in parallel and is capable of distinguishing data that are not linearly separable.21 Moreover, MLP utilizes a supervised learning technique named backpropagation for training the network,22,23 which gives a certain degree of flexibility to handling noise.24 Most importantly, MLP is a nonparametric dynamic model that is automatically self-training and can readjust the internal parameters by backpropagation when more samples enter the network.25 This technique demonstrates better performance than other machine learning methods, such as radial basis function, recurrent neural network, and self-organizing map.21 In this study, we provided a novel approach to distinguish HCC and LC patients based on bead extraction and MALDI-TOF/TOF MS analysis. This approach was simple, available, and not invasive compared with other diagnostic approaches. Biases from sample collection, MALDI-TOF/TOF MS performance, agents used, and the operator could influence the reproducibility.26,27 To minimize these influences, we standardized the collection and fractionation protocol, and optimized parameters of MALDI-TOF/TOF MS.28–30 Distribution of age and gender between HCC and LC group were taken into account as we collected the patient serum samples. Strikingly, the MALDI-TOF/TOF MS peptide profiling technology needed no knowledge of the amino acid sequence of the peaks or their biological functions for class prediction. Application of this approach by clinicians and technical staff appears promising in selecting the training cases, consistency in sample handling, proper operation of the precision instrument, and correct establishment of analysis algorithm. To further investigate HCC and LC, we have applied an optimal approach to identify the sequences of selected features. Sequence identification showed that one feature was a fragment of alpha-1antitrypsin (A1AT), and one feature was a fragment of ACT. Alpha-1-antitrypsin is a protease inhibitor belonging to the serpin superfamily. In our study, A1AT was increased in the sera of patients with HCC compared with LC, which was consistent with the results of Wang et al.15 and Naitoh et al.16 However, the exact mechanism and role of elevated serum level of A1AT in HCC remains unclear.31 It could be explained by a hypothesis that production of A1AT by tumor cells correlates with the regional proteolytic and inflammatory activity, which is probably involved in the protection of tumor cells.32 ACT is an acute-phase protein secreted by hepatocytes in response to cytokines such as oncostatin M. Decreased expression of ACT has been observed in human HCC tissues and cells,33 and it suggested that the attenuation of endogenous ACT expression during liver regeneration, as well as its reduced levels in HCC, could fulfill a physiopathologic role. In the present work, ACT was also down-regulated in sera of patients with HCC. HCC is a major cause of death in cirrhotic patients and is associated with LC in more than 90% of cases. Early diagnosis and treatment of HCC are expected to improve survival of patients. Our results indicate that the MLP model is a novel and useful serum peptide pattern for differentially diagnosing HCC and LC with a sensitivity of 90%. This model alone, or together with the serum AFP, was more effective than serum
Specific serum peptides of liver cancer
AFP detection for the early diagnosis of HCC in patients with HBV-related LC.
Acknowledgments This work was supported by grants from National Natural Science Foundation of China (No.81025010, 21007092), Shandong Provincial Natural Science Foundation, China (ZR2010HQ027), National Key Technology R&D Program (2009BAK61B04), Innovation of methodology (2010IM030300), and China National Basic Research Program (2010CB529404).
References 1 Bruix J, Boix L, Sala M, Llovet JM. Focus on hepatocellular carcinoma. Cancer Cell 2004; 5: 215–19. 2 Bruix J, Sherman M, Practice Guidelines Committee, American Association for the Study of Liver Diseases. Management of hepatocellular carcinoma. Hepatology 2005; 42: 1208–36. 3 Taylor-Robinson SD, Foster GR, Arora S, Hargreaves S, Thomas HC. Increase in primary liver cancer in the UK, 1979–94. Lancet 1997; 350: 1142–3. 4 Forner A, Vilana R, Ayuso C et al. Diagnosis of hepatic nodules 20 mm or smaller in cirrhosis: prospective validation of the noninvasive diagnostic criteria for hepatocellular carcinoma. Hepatology 2008; 47: 97–104. 5 Tiollais P, Pourcel C, Dejean A. The hepatitis B virus. Nature 1985; 317: 489–95. 6 Yuen MF, Yuan HJ, Wong DK et al. Prognostic determinants for chronic hepatitis B in Asians: therapeutic implications. Gut 2005; 54: 1610–14. 7 Wong VW, Chan SL, Mo F et al. Clinical scoring system to predict hepatocellular carcinoma in chronic hepatitis B carriers. J. Clin. Oncol. 2010; 28: 1660–5. 8 Bruix J, Llovet JM. Prognostic prediction and treatment strategy in hepatocellular carcinoma. Hepatology 2002; 35: 519–24. 9 Kim CK, Lim JH, Lee WJ. Detection of hepatocellular carcinomas and dysplastic nodules in cirrhotic liver: accuracy of ultrasonography in transplant patients. J. Ultrasound Med. 2001; 20: 99–104. 10 Liotta LA, Ferrari M, Petricoin E. Clinical proteomics: written in blood. Nature 2003; 425: 905. 11 Fiedler GM, Leichtle AB, Kase J et al. Serum peptidome profiling revealed platelet factor 4 as a potential discriminating Peptide associated with pancreatic cancer. Clin. Cancer Res. 2009; 15: 3812–19. 12 Patz EF Jr, Campa MJ, Gottlin EB, Kusmartseva I, Guan XR, Herndon JE 2nd. Panel of serum biomarkers for the diagnosis of lung cancer. J. Clin. Oncol. 2007; 25: 5578–83. 13 Liang T, Wang N, Li W et al. Identification of complement C3f-desArg and its derivative for acute leukemia diagnosis and minimal residual disease assessment. Proteomics 2010; 10: 90–8. 14 Visintin I, Feng Z, Longton G et al. Diagnostic markers for early detection of ovarian cancer. Clin. Cancer Res. 2008; 14: 1065–72. 15 Wang M, Long RE, Comunale MA et al. Novel fucosylated biomarkers for the early detection of hepatocellular carcinoma. Cancer Epidemiol. Biomarkers Prev. 2009; 18: 1914–21. 16 Naitoh A, Aoyagi Y, Asakura H. Highly enhanced fucosylation of serum glycoproteins in patients with hepatocellular carcinoma. J. Gastroenterol. Hepatol. 1999; 14: 436–45. 17 Mark Hall EF, Holmes G, Pfahringer B, Reutemann P, Ian H. The WEKA data mining software: an update. SIGKDD Explor. 2009; 11: 10–18.
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
1549
Specific serum peptides of liver cancer
N Wang et al.
18 Vergara IA, Norambuena T, Ferrada E, Slater AW, Melo F. StAR: a simple tool for the statistical comparison of ROC curves. BMC Bioinformatics 2008; 9: 265. 19 Vapnik V. Statistical Learning Theory. New York: John Wiley and Sons, 1998. 20 Attallah AM, Omran MM, Attallah AA et al. HCC-ART score, a simple, highly sensitive and specific test for early diagnosis of hepatocellular carcinoma: a large-scale, multicentre study. Br. J. Cancer 2013; 109: 1657–65. 21 Oztekin A, Delen D, Kong ZJ. Predicting the graft survival for heart-lungtransplantation patients: an integrated data mining methodology. Int. J. Med. Inform. 2009; 78: e84–96. 22 Rosenblatt F. Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms. Washington, DC: Spartan Books, 1961. 23 Rumelhart DE, McClelland JL, The PDP Research Group, eds. Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Volume 1: Foundations. Cambridge, MA: MIT Press, 1986. 24 Cross SS, Harrison RF, Kennedy RL. Introduction to neural networks. Lancet 1995; 346: 1075–9. 25 Zhang M, Yin F, Chen B et al. Pretransplant prediction of posttransplant survival for liver recipients with benign end-stage liver diseases: a nonlinear model. PLoS ONE 2012; 7: e31256. 26 Check E. Proteomics and cancer: running before we can walk? Nature 2004; 429: 496–7. 27 Coombes KR. Analysis of mass spectrometry profiles of the serum proteome. Clin. Chem. 2005; 51: 1–2. 28 Semmes OJ, Feng Z, Adam BL et al. Evaluation of serum protein profiling by surface-enhanced laser desorption/ionization
1550
29
30
31
32 33
time-of-flight mass spectrometry for the detection of prostate cancer: I. Assessment of platform reproducibility. Clin. Chem. 2005; 51: 102–12. Villanueva J, Shaffer DR, Philip J et al. Differential exoprotease activities confer tumor-specific serum peptidome patterns. J. Clin. Invest. 2006; 116: 271–84. Paradis V, Degos F, Dargere D et al. Identification of a new marker of hepatocellular carcinoma by serum protein profiling of patients with chronic liver diseases. Hepatology 2005; 41: 40–7. Topic A, Ljujic M, Radojkovic D. Alpha-1-antitrypsin in pathogenesis of hepatocellular carcinoma. Hepat. Mon. 2012; 12 (10 HCC): e7042. Sawaya R, Zuccarello M, Highsmith R. Alpha-1-antitrypsin in human brain tumors. J. Neurosurg. 1987; 67: 258–62. Santamaria M, Pardo-Saganta A, Alvarez-Asiain L et al. Nuclear alpha1-antichymotrypsin promotes chromatin condensation and inhibits proliferation of human hepatocellular carcinoma cells. Gastroenterology 2013; 144: 818–828. e814.
Supporting information Additional Supporting Information may be found in the online version of this article at the publisher’s web-site: Table S1 Clinical characteristics of study participants. Table S2 Reproducibility of mass spectra analyzed by MALDITOF/TOF MS. Table S3 Performance of the three classifiers on the test set.
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
doi:10.1111/jgh.12545
H E PAT O L O G Y
Serum peptide pattern that differentially diagnoses hepatitis B virus-related hepatocellular carcinoma from liver cirrhosis Na Wang,*,1 Yuan Cao,*,†,1 Wei Song,*,‡,1 Kun He,* Tao Li,* Jie Wang,* Bin Xu,* Hai-yan Si,† Cheng-Jin Hu† and Ai-Ling Li* *Institute of Basic Medical Sciences, National Center of Biomedical Analysis, Beijing, †Department of Laboratory Medicine, 90th General Hospital of Jinan, Jinan, Shandong, and ‡Department of Hematology and Oncology, The First Hospital of Jilin University, Changchun, Jilin, China
Key words differential diagnosis, hepatocellular carcinoma, liver cirrhosis, serum peptide. Accepted for publication 18 January 2014. Correspondence Ailing Li, Institute of Basic Medical Sciences, National Center of Biomedical Analysis, 27 Tai-Ping Road, Beijing 100850, China. Email: [email protected] Cheng-Jin Hu, Department of Laboratory Medicine, 90th General Hospital of Jinan, 25 Shifan Road, Jinan, Shandong 250031, China. Email: [email protected] 1
These authors contributed equally to this work. Conflict of interest: No.
Abstract Background: Although alpha-fetoprotein (AFP) is a useful serologic marker of hepatocellular carcinoma (HCC), it is not sufficiently sensitive to differentiate HCC and liver cirrhosis (LC) caused by hepatitis B virus (HBV) infection. Aims: The aim is to discover novel noninvasive specific serum biomarkers for the differential diagnosis of HBV-related HCC and LC. Methods: With a highly optimized peptide extraction and matrix-assisted laser desorption/ionization time of flight/time of flight mass spectrometric approach, we investigated serum peptide profiles of 80 HCC and 67 LC patients. Three supervised machine learning methods were employed to construct classifiers. Receiver operator curves were plotted to evaluate the performance of classifiers. Results: With a support vector machine-based strategy, we picked nine peaks with m/z ratios of 819.49, 1076.14, 1341.72, 2551.44, 3156.44, 3812.88, 4184.26, 4465.92, and 4776.41 to construct the classifier. We proposed a novel method for distinguishing HCC from cirrhosis, based on a multilayer perceptron (MLP) method. We obtained a sensitivity of 90.0%, specificity of 79.4%, and overall accuracy of 85.1% on an independent test set. The combination of the MLP model and serum AFP level outperformed serum AFP marker alone in distinguishing HCC patients from LC patients. In this experience, sensitivity increased from 62.5% to 87.5%, and specificity increased from 79.4% to 88.2%. Conclusions: Our results indicate that the MLP model is a novel and useful serum peptide pattern for distinguishing HCC and LC. The peptidome signature alone or together with serum AFP determination may be a more effective method for early diagnosis of HCC in patients with HBV-related LC.
Introduction Hepatocellular carcinoma (HCC) is one of the most common malignancies worldwide, accounting for approximately 600 000 deaths per year.1–4 The incidence of HCC is rising over the past two decades. Over 80% of HCC develop from liver cirrhosis (LC), mainly related to chronic hepatitis B virus infection. Early diagnosis of HCC can greatly improve the outcome, with better longterm survival and reduced recurrence risk for operative treatment. Because of the lack of effective means of early diagnosis, only 30–40% of patients with HCC are candidates for potentially curative treatments at the time of diagnosis.5–7 Therefore, the differential diagnosis between HBV-related HCC and LC is very important. At present, serum alpha-fetoprotein (AFP) is available for the diagnosis and monitoring of HCC. However, a considerable proportion of HCC patients do not have a significant elevation of serum AFP. Because of its low sensitivity, AFP is not an ideal 1544
tumor marker. In addition, imaging techniques (ultrasound, computed tomography, and magnetic resonance imaging) are widely used for HCC diagnosis. At earlier stages, when the tumor is less than 2 cm in size, the imaging techniques are seldom diagnostic for HCC.6,8,9 Accordingly, discovery of novel noninvasive, effective, and reliable serum biomarkers for the differential diagnosis of HBV-related HCC and LC would play a pivotal role in improving the prognosis of patients with HCC. The search for cancer biomarkers has recently turned, at least partially, from cancer-specific proteins, which are rare and usually disappointing in their sensitivity, to circulatory peptides. The development of mass spectrometry (MS) technology now permits the display of hundreds of small- to medium-sized peptides from several microliters of serum in a high-throughput manner.10 Using a precision instrument of matrix-assisted laser desorption/ ionization time of flight MS (MALDI-TOF MS), recent studies have identified cancer-specific peptide markers.11–14 However, the
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
N Wang et al.
serum peptidome study for distinguishing HCC and LC by MALDI-TOF/TOF MS analysis has not been performed. In this study, we used a MALDI-TOF/TOF MS strategy to investigate serum peptide profiles of 80 HCC and 67 LC patients. Three supervised machine learning methods were employed to construct classifiers. Using the multilayered perceptron (MLP) model, we obtained a sensitivity of 90.0%, specificity of 79.4%, and overall accuracy of 85.1% on an independent test set. This model alone, or together with the serum AFP, was more effective than serum AFP detection for the early diagnosis of HCC in patients with HBV-related LC.
Materials and methods Study overview. Study overview was summarized in Figure 1. Serum samples of 80 HCC and 67 LC patients were analyzed by MALDI-TOF-MS for peptide expression. Independent training and test sets were created, with similar representation of age and gender in each set. A support vector machine (SVM)based strategy was used to select peaks to differentially diagnose HCC from LC. With the selected peaks, three supervised machine learning methods were employed to construct classifiers based on the training set to distinguish HCC from LC. The test set was used to evaluate the three classifiers. We also evaluated the performance of the candidate classifier combined with serum AFP. Serum samples. Serum samples of 147 patients, including 80 HCC patients and 67 LC patients, were collected from the 90th hospital of Jinan in China. Tumor stage was defined according to the Barcelona Clinic Liver Cancer (BCLC) staging system;15 for the purpose of this study, we classified tumors with BCLC stage 0 + A as early-stage HCC.16 The clinical characteristics of HCC and LC were summarized in Supplemental Table S1. All patients were HBV infected, with a Child–Pugh class A to C. Blood biochemistry, AFP assay, computed tomography, and liver biopsy were performed on all patients. The study was approved by the
Specific serum peptides of liver cancer
local institutional review board and ethical committee. Informed consent was obtained from all patients. All samples were collected using a standard clinical protocol. The blood samples were collected from patients in the morning, allowed to clot for 2 h at room temperature, and centrifuged for 15 min at 1500–2000 rpm. Sera were then frozen at −80°C. For the reproducibility experiments, sera from 10 HCC patients were pooled. Six within-run assays were then performed on the MALDI-TOF/TOF MS instrument (Supplemental Table S2). The mean CV of within-run assays was 18.7% (range: 16.5–25.7%). Peptide extraction. Copper-chelated magnetic beads were used for extracting peptides from serum. For each sample, 5 μL serum was mixed with 5 μL beads and 20 μL of binding solution. Samples were washed three times using 100 μL of washing solution, then 20 μL of eluent buffer was added. A 1 μL of the eluent and 1 μL α-cyano-4-hydroxycinnamic acid matrix solution were mixed and spotted onto a 600-um-diameter spot size 384 MTP target plate (Bruker Daltonik, Billerica, MA, USA). The peptide calibration standard with the same matrix was applied to target spots for external calibration of the instrument. The processed samples were analyzed using MALDI-TOF/TOF MS (ultraflex III, Bruker Daltonics), equipped with a pulsed ion extraction ion source. MALDI-TOF/TOF-MS data preprocessing and peak selection. After curve smoothing in all sample spectra and subtracting the baseline, the peaks with m/z between 800 Da and 10 000 Da and with a signal-to-noise ratio over 5, were labeled using FlexAnalysis 2.4 software (Bruker Daltonik). A peak cluster was created including m/z value and signal intensities and then exported together into excel file format. An in-house MatchPeaks software was used to normalize and align all sample peak intensities. These peaks were fit in a spreadsheet for further statistical analysis. The SVM-based strategy was used to select peaks to identify their discriminatory power.
Figure 1 Study overview. This diagram shows the route to select distinguishing peptides and processed models for HCC and LC diagnosis. The numbers indicate total number of selected features at that stage of the study.
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
1545
Specific serum peptides of liver cancer
N Wang et al.
Classifier construction and evaluation. To distinguish LC and HCC, three supervised machine learning methods (MLP, sequential minimal optimization [SMO], and classification and regression tree [CART]) were employed to construct classifiers in Waikato Environment for Knowledge Analysis.17 In this study, independent training (n = 73) and test sets (n = 74) were created, with similar numbers of LC and HCC, and similar representation of age and gender. With these sets, we assessed the performance of the three classifiers by considering the number of correctly classified (true positives and true negatives) and incorrectly classified (false positives and false negatives) samples in the test set. Accuracy, sensitivity, and specificity were calculated. Receiver operator curves (ROC) were plotted to evaluate the performance of a range of classifiers. Area under curve (AUC) was given for each curve. StAR18 (Molecular Bioinformatics Laboratory, Pontificial Catholic University of Chile, Santiago, Chile) was used for statistical comparison of AUCs.
Results Dataset and feature selection. To screen serum peptides for HCC and LC, we used MALDI-TOF-MS to analyze serum samples from 80 HCC and 67 LC patients. About 70% of subjects serum spectra were used as training set (n = 73), and the rest were used as test set (n = 74; Fig. 1, Table 1). A total of 235 peaks with m/z between 800 Da and 10 000 Da were obtained from the 147 patients. After all peaks intensity were normalized using house MatchPeaks software (MS2PCA), feature variables were evaluTable 1
Construction and evaluation of the three classifiers using supervised machine learning methods. To discriminate HCC from LC, we constructed three classifiers using supervised machine learning methods MLP, SMO, and CART. To evaluate the performance of the three classifiers, we applied a preliminary test using 10-fold cross-validation on the training set. The accuracy, sensitivity, and specificity of MLP were 85.1%, 90%, and 79.4%, respectively. SMO gave comparable results with an accuracy of 75.7%, sensitivity of 82.5%, and specificity of 67.6%. CART gave the worst results (accuracy of 50%, sensitivity of 62.5%, and specificity of 35.3%). The ability of a classifier to discriminate data correctly in the test set was known as its generalization performance.19 We compared the generalization performance of a series of classifiers by plotting their performance on the test set in ROC space (Fig. 2, Supplemental Table S3). The MLP classifier exhibited the best results on the test set, with a sensitivity of 90.0% and specificity of 79.4% (overall accuracy of 85.1%). The SMO classifier showed similar sensitivity (82.5%) and poor specificity (67.6%) on the test set. CART gave the worst results (accuracy of 50%, sensitivity of 62.5%, and specificity of 35.3%). The AUC was 0.909, 0.751, and 0.582 for MLP, SMO, and CART, respectively. To evaluate the performance of MLP model, ROC of each peptide was plotted as follows (Fig. 3). AUC
Training set and test set
Number of HCC patients Number of LC patients Gender (male : female) Age (years) AFP (ng/mL) ALT (U/L) AST (U/L) AKP (U/L) Child–Pugh A B C Tumor number and size Single nodule ≤ 2 cm Single nodule 2–3 cm Three nodules ≤ 3 cm AJCC stage I II BCLC score 0 A
Training set (n = 73)
Test set (n = 74)
P value
40 33 48:25 54.2 ± 11.8 104.9 ± 164.3 75.4 ± 123.6 71.8 ± 68.8 146.7 ± 87.2
40 34 54:20 53.8 ± 11.4 90.6 ± 168.4 75.1 ± 103.5 94.8 ± 178.4 128.4 ± 90.5
0.928
51 16 6
50 18 6
23 15 2
25 15 0
22 18
24 16
18 22
19 21
0.441 0.557 0.630 0.607 0.120 0.461 0.860
0.617
0.651
0.823
AFP, alpha-fetoprotein; AJCC, American Joint Committee on Cancer; AKP, alkaline; ALT, alanine aminotransferase; BCLC, Barcelona Clinic Liver Cancer; HCC, hepatocellular carcinoma; LC, liver cirrhosis; NA, not available.
1546
ated using the SVM-based strategy and ranked by the square of the weight assigned by the SVM. Nine peptides were selected for classifier construction, which had m/z ratios of 819, 1076, 1341, 2551, 3156, 3812, 4184, 4465, and 4776.
Figure 2 Classifier performance in receiver operator curves (ROC) space. Red line shows ROC curve of multilayered perceptron (MLP). Blue line represents ROC curve of sequential minimal optimization (SMO). Green line represents ROC curve of and regression tree (CART). The area under curve is 0.900, 0.748 and 0.559 for MLP, SMO and , MLP; , SMO; , CART. CART, respectively.
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
N Wang et al.
Specific serum peptides of liver cancer
Figure 3 Performance of single feature peptide in receiver operator curves (ROC) space. The lines show ROC curve of single peptide of the nine peptides. The area under curve is 0.705, 0.629, 0.519, 0.674, 0.739, 0.696, 0.684, 0.675, and 0.661 for the peptide m/z 819, 1076, 1341, 2551, 3156, 3812, 4184, 4465, and 4776, respectively.
of each feature peptide was also calculated. None of AUCs of these proteins were better than that of MLP, which proven the performance of the MLP classifier.
The combination of MLP model analysis and AFP for the differential diagnosis of HCC and LC. Serum levels of AFP are currently applied for screening HCC. When the
level of AFP is over 20 ng/mL, the patient is considered to have HCC. We assayed the serum levels of AFP in all the HCC and LC patients from the training set and test set. When the cutoff value of serum AFP was 20 ng/mL, the outcome indicated the sensitivity of 62.5% and specificity of 79.4% on the test set. The MLP model gave a sensitivity of 90% and specificity of 79.4%. The sensitivity in MLP model was much higher than AFP detection. Importantly, combining MLP model and AFP detection, a sensitivity of 87.5%
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
1547
Specific serum peptides of liver cancer
N Wang et al.
was achieved, with a specificity of 88.6% (Table 2). To compare the accuracy among MLP model, AFP level, and combination of MLP model and AFP detection (MLP + AFP), we plotted ROC of the three methods. AUC was also calculated for each curve (see Fig. 4 and Table 2). Based on our test set, MLP + AFP gave the best result with AUC of 0.943, while MLP showed the comparable AUC of 0.909. However, there is no significant difference (P = 0.0561) on AUCs between the two algorithms by using StAR.18 The AUC of AFP is 0.751, which is similar to the report by Attallah.20 Obviously, the AUC of AFP is less than those of MLP or MLP + AFP (P < 0.05). Identification of feature peptides. Peptide sequencing was performed to identify the nine feature peptides in the sera of HCC patients and LC patients. Five peptides were positively identified by FT MS/MS, and the Fourier transform (FT) sequencing result is shown in Table 3. For reference, m/z 1076 and 2551 stand for peptides of Fibrinogen α, m/z 3156 for a peptide of Transthyretin, m/z 4465 for a peptide of Antichymotrypsin (ACT), and m/z 4776 for a peptide of α-1 antitrypsin.
Figure 4 Performances of multilayered perceptron (MLP), alphafetoprotein (AFP) and MLP + AFP in receiver operator curves (ROC) space. Blue line shows ROC curve of MLP. Red line represents ROC curve of MLP + AFP. Green line represents ROC curve of AFP. The area under curve is 0.943, 0.909, and 0.751 for MLP + AFP, MLP, and AFP, , MLP + AFP; , MLP; , AFP. respectively.
Table 2 Diagnostic performances of AFP, MLP and AFP + MLP on the test set Classifiers
TP
TN
FP
FN
ACC (%)
SE (%)
SP (%)
AUC
MLP MLP + AFP AFP
36 35 25
27 29 27
7 5 7
4 5 15
85.1 86.5 71.6
90.0 87.5 62.5
79.4 88.2 79.4
0.909 0.943 0.751
ACC, accuracy; AUC, area under curve; AFP, alpha-fetoprotein; FN, false negatives; FP, false positives; MLP, multilayered perceptron; SE, sensitivity; SP, specificity; TN, true negatives; TP, true positives.
Table 3
Discussion In this study, we established an MLP model to differentiate HCC and LC patients with a sensitivity of 90.0% and specificity of 79.4%. When the cutoff value of serum AFP was 20 ng/mL, the sensitivity was 62.5%, and the specificity was 79.4%. The combination of the MLP model and serum AFP level outperformed serum AFP alone in distinguishing HCC from LC. Using this combination, the sensitivity increased from 62.5% to 87.5%. To construct classifiers for distinguishing HCC from cirrhosis, three supervised machine learning methods were used. The classification accuracy was 85.1%, 75.7%, and 50% for MLP, SMO, and CART, respectively. To avoid biases, we applied our trained classifiers to the independent test set. The fact that a small increase in accuracy and sensitivity was observed in the test set demonstrates the generalization performance of the classifiers. After regulating the probability threshold of the MLP, we obtained a best prediction result on the test set. Eventually, the MLP was selected as our classifier in this study (Supplemental Table S3, Fig. 2).
Identification of sequences of proteomic features differentially expressed between the HCC and LC patients by FT-ICR-MS
m/z
Peptide name
Peptide sequences
819.49 1076.14 1341.72 2551.44 3156.44 3812.88 4184.26 4465.92 4776.41
Not identified Fibrinogen a Not identified Fibrinogen a Transthyretin Not identified Not identified Antichymotrypsin α-1 antitrypsin
/ GDFLAEGGGVR / SSSYSKQFTSSTSYNRGDSTFES DSGPRRYTIAALLSPYSYSTTAVVTNPKE / / LVETRTIVRFNRPFLMIIVPTDTQNIFFMSKVTNPKQA LEAIPMSIPPEVKFNKPFVFLMIDQNTKSPLFMGKVVNPTQK
FT-ICR-MS, Fourier transform ion cyclotron resonance mass spectometry; HCC, hepatocellular carcinoma; LC, liver cirrhosis.
1548
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
N Wang et al.
Recently, with the development of artificial intelligence, artificial neural networks have been widely used in the area of medical data mining. MLP is a feedforward artificial neural network model that maps input data sets onto a set of appropriate output. This method can use a large number of simple units to process information in parallel and is capable of distinguishing data that are not linearly separable.21 Moreover, MLP utilizes a supervised learning technique named backpropagation for training the network,22,23 which gives a certain degree of flexibility to handling noise.24 Most importantly, MLP is a nonparametric dynamic model that is automatically self-training and can readjust the internal parameters by backpropagation when more samples enter the network.25 This technique demonstrates better performance than other machine learning methods, such as radial basis function, recurrent neural network, and self-organizing map.21 In this study, we provided a novel approach to distinguish HCC and LC patients based on bead extraction and MALDI-TOF/TOF MS analysis. This approach was simple, available, and not invasive compared with other diagnostic approaches. Biases from sample collection, MALDI-TOF/TOF MS performance, agents used, and the operator could influence the reproducibility.26,27 To minimize these influences, we standardized the collection and fractionation protocol, and optimized parameters of MALDI-TOF/TOF MS.28–30 Distribution of age and gender between HCC and LC group were taken into account as we collected the patient serum samples. Strikingly, the MALDI-TOF/TOF MS peptide profiling technology needed no knowledge of the amino acid sequence of the peaks or their biological functions for class prediction. Application of this approach by clinicians and technical staff appears promising in selecting the training cases, consistency in sample handling, proper operation of the precision instrument, and correct establishment of analysis algorithm. To further investigate HCC and LC, we have applied an optimal approach to identify the sequences of selected features. Sequence identification showed that one feature was a fragment of alpha-1antitrypsin (A1AT), and one feature was a fragment of ACT. Alpha-1-antitrypsin is a protease inhibitor belonging to the serpin superfamily. In our study, A1AT was increased in the sera of patients with HCC compared with LC, which was consistent with the results of Wang et al.15 and Naitoh et al.16 However, the exact mechanism and role of elevated serum level of A1AT in HCC remains unclear.31 It could be explained by a hypothesis that production of A1AT by tumor cells correlates with the regional proteolytic and inflammatory activity, which is probably involved in the protection of tumor cells.32 ACT is an acute-phase protein secreted by hepatocytes in response to cytokines such as oncostatin M. Decreased expression of ACT has been observed in human HCC tissues and cells,33 and it suggested that the attenuation of endogenous ACT expression during liver regeneration, as well as its reduced levels in HCC, could fulfill a physiopathologic role. In the present work, ACT was also down-regulated in sera of patients with HCC. HCC is a major cause of death in cirrhotic patients and is associated with LC in more than 90% of cases. Early diagnosis and treatment of HCC are expected to improve survival of patients. Our results indicate that the MLP model is a novel and useful serum peptide pattern for differentially diagnosing HCC and LC with a sensitivity of 90%. This model alone, or together with the serum AFP, was more effective than serum
Specific serum peptides of liver cancer
AFP detection for the early diagnosis of HCC in patients with HBV-related LC.
Acknowledgments This work was supported by grants from National Natural Science Foundation of China (No.81025010, 21007092), Shandong Provincial Natural Science Foundation, China (ZR2010HQ027), National Key Technology R&D Program (2009BAK61B04), Innovation of methodology (2010IM030300), and China National Basic Research Program (2010CB529404).
References 1 Bruix J, Boix L, Sala M, Llovet JM. Focus on hepatocellular carcinoma. Cancer Cell 2004; 5: 215–19. 2 Bruix J, Sherman M, Practice Guidelines Committee, American Association for the Study of Liver Diseases. Management of hepatocellular carcinoma. Hepatology 2005; 42: 1208–36. 3 Taylor-Robinson SD, Foster GR, Arora S, Hargreaves S, Thomas HC. Increase in primary liver cancer in the UK, 1979–94. Lancet 1997; 350: 1142–3. 4 Forner A, Vilana R, Ayuso C et al. Diagnosis of hepatic nodules 20 mm or smaller in cirrhosis: prospective validation of the noninvasive diagnostic criteria for hepatocellular carcinoma. Hepatology 2008; 47: 97–104. 5 Tiollais P, Pourcel C, Dejean A. The hepatitis B virus. Nature 1985; 317: 489–95. 6 Yuen MF, Yuan HJ, Wong DK et al. Prognostic determinants for chronic hepatitis B in Asians: therapeutic implications. Gut 2005; 54: 1610–14. 7 Wong VW, Chan SL, Mo F et al. Clinical scoring system to predict hepatocellular carcinoma in chronic hepatitis B carriers. J. Clin. Oncol. 2010; 28: 1660–5. 8 Bruix J, Llovet JM. Prognostic prediction and treatment strategy in hepatocellular carcinoma. Hepatology 2002; 35: 519–24. 9 Kim CK, Lim JH, Lee WJ. Detection of hepatocellular carcinomas and dysplastic nodules in cirrhotic liver: accuracy of ultrasonography in transplant patients. J. Ultrasound Med. 2001; 20: 99–104. 10 Liotta LA, Ferrari M, Petricoin E. Clinical proteomics: written in blood. Nature 2003; 425: 905. 11 Fiedler GM, Leichtle AB, Kase J et al. Serum peptidome profiling revealed platelet factor 4 as a potential discriminating Peptide associated with pancreatic cancer. Clin. Cancer Res. 2009; 15: 3812–19. 12 Patz EF Jr, Campa MJ, Gottlin EB, Kusmartseva I, Guan XR, Herndon JE 2nd. Panel of serum biomarkers for the diagnosis of lung cancer. J. Clin. Oncol. 2007; 25: 5578–83. 13 Liang T, Wang N, Li W et al. Identification of complement C3f-desArg and its derivative for acute leukemia diagnosis and minimal residual disease assessment. Proteomics 2010; 10: 90–8. 14 Visintin I, Feng Z, Longton G et al. Diagnostic markers for early detection of ovarian cancer. Clin. Cancer Res. 2008; 14: 1065–72. 15 Wang M, Long RE, Comunale MA et al. Novel fucosylated biomarkers for the early detection of hepatocellular carcinoma. Cancer Epidemiol. Biomarkers Prev. 2009; 18: 1914–21. 16 Naitoh A, Aoyagi Y, Asakura H. Highly enhanced fucosylation of serum glycoproteins in patients with hepatocellular carcinoma. J. Gastroenterol. Hepatol. 1999; 14: 436–45. 17 Mark Hall EF, Holmes G, Pfahringer B, Reutemann P, Ian H. The WEKA data mining software: an update. SIGKDD Explor. 2009; 11: 10–18.
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
1549
Specific serum peptides of liver cancer
N Wang et al.
18 Vergara IA, Norambuena T, Ferrada E, Slater AW, Melo F. StAR: a simple tool for the statistical comparison of ROC curves. BMC Bioinformatics 2008; 9: 265. 19 Vapnik V. Statistical Learning Theory. New York: John Wiley and Sons, 1998. 20 Attallah AM, Omran MM, Attallah AA et al. HCC-ART score, a simple, highly sensitive and specific test for early diagnosis of hepatocellular carcinoma: a large-scale, multicentre study. Br. J. Cancer 2013; 109: 1657–65. 21 Oztekin A, Delen D, Kong ZJ. Predicting the graft survival for heart-lungtransplantation patients: an integrated data mining methodology. Int. J. Med. Inform. 2009; 78: e84–96. 22 Rosenblatt F. Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms. Washington, DC: Spartan Books, 1961. 23 Rumelhart DE, McClelland JL, The PDP Research Group, eds. Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Volume 1: Foundations. Cambridge, MA: MIT Press, 1986. 24 Cross SS, Harrison RF, Kennedy RL. Introduction to neural networks. Lancet 1995; 346: 1075–9. 25 Zhang M, Yin F, Chen B et al. Pretransplant prediction of posttransplant survival for liver recipients with benign end-stage liver diseases: a nonlinear model. PLoS ONE 2012; 7: e31256. 26 Check E. Proteomics and cancer: running before we can walk? Nature 2004; 429: 496–7. 27 Coombes KR. Analysis of mass spectrometry profiles of the serum proteome. Clin. Chem. 2005; 51: 1–2. 28 Semmes OJ, Feng Z, Adam BL et al. Evaluation of serum protein profiling by surface-enhanced laser desorption/ionization
1550
29
30
31
32 33
time-of-flight mass spectrometry for the detection of prostate cancer: I. Assessment of platform reproducibility. Clin. Chem. 2005; 51: 102–12. Villanueva J, Shaffer DR, Philip J et al. Differential exoprotease activities confer tumor-specific serum peptidome patterns. J. Clin. Invest. 2006; 116: 271–84. Paradis V, Degos F, Dargere D et al. Identification of a new marker of hepatocellular carcinoma by serum protein profiling of patients with chronic liver diseases. Hepatology 2005; 41: 40–7. Topic A, Ljujic M, Radojkovic D. Alpha-1-antitrypsin in pathogenesis of hepatocellular carcinoma. Hepat. Mon. 2012; 12 (10 HCC): e7042. Sawaya R, Zuccarello M, Highsmith R. Alpha-1-antitrypsin in human brain tumors. J. Neurosurg. 1987; 67: 258–62. Santamaria M, Pardo-Saganta A, Alvarez-Asiain L et al. Nuclear alpha1-antichymotrypsin promotes chromatin condensation and inhibits proliferation of human hepatocellular carcinoma cells. Gastroenterology 2013; 144: 818–828. e814.
Supporting information Additional Supporting Information may be found in the online version of this article at the publisher’s web-site: Table S1 Clinical characteristics of study participants. Table S2 Reproducibility of mass spectra analyzed by MALDITOF/TOF MS. Table S3 Performance of the three classifiers on the test set.
Journal of Gastroenterology and Hepatology 29 (2014) 1544–1550 © 2014 Journal of Gastroenterology and Hepatology Foundation and Wiley Publishing Asia Pty Ltd
