UNCORRECTED PROOF 1 2 3 A new NMR-based metabolomics approach for the diagnosis 4 of biliary tract cancer 5 He Wen1, , Sung Soo Yoo2, , Jinho Kang1 , Hee Goo Kim2 , Jin-Seok Park2 , Seok Jeong2 , Jung Il Lee2 , 6 Hyuk Nam Kwon1 , Sunmi Kang1 , Don-Haeng Lee2,*, Sunghyouk Park1,* 7 1 Department of Biochemistry, Inha University Hospital and Center for Advanced Medical Education by BK21 Project, College of Medicine, 8 Inha University, Shinheung-dong, Chung-gu, Incheon 400-712, Republic of Korea; 2 Division of Gastroenterology, Department of 9 Internal Medicine, Inha University Hospital and Center for Advanced Medical Education by BK21 Project, College of Medicine, 10 Inha University, Incheon, Republic of Korea 11 1213 Background & Aims: Biliary tract cancer is highly lethal at pre- 14 sentation, with increasing mortality worldwide. Current diagnos- 15 tic measures employing multiple criteria such as imaging, 16 cytology, and serum tumor markers are not satisfactory, and a 17 new diagnostic tool is needed. Because bile is a cognate metabo- 18 lite-rich bio-fluid in the biliary ductal system, we tested a new 19 metabolomic approach to develop an effective diagnostic tool. 20 Methods: Biles were collected prospectively from patients with 21 cancer (n = 17) or benign biliary tract diseases (n = 21) with per- 22 cutaneous or endoscopic methods. Nuclear magnetic resonance 23 spectra (NMR) of these biles were analyzed using orthogonal par- 24 tial least square discriminant analysis (OPLS-DA). 25 Results: The metabolomic 2-D score plot showed good separation 26 between cancer and benign groups. The contributing NMR signals 27 were analyzed using a statistical TOCSY approach with verifica- 28 tion. The diagnostic performance assessed by leave-one-out anal- 29 ysis exhibited 88% sensitivity and 81% specificity, better than the 30 conventional markers (CEA, CA19-9, and bile cytology). 31 Conclusion: The NMR-based metabolomics approach provides 32 good performance in discriminating cancer and benign biliary 33 duct diseases. The excellent predictability of the method suggests 34 that it can, at least, increase the currently available diagnostic 35 approaches. 36 Ó 2009 European Association for the Study of the Liver. Published 37 by Elsevier B.V. All rights reserved. 3839 40Introduction 41Biliary tract cancer arises from epithelial cells of the intrahepatic 42and extrahepatic bile ducts. Although this type of cancer is not 43very common, it is highly lethal, since most are locally advanced 44at presentation. Its incidence increases with age, and the mortal- 45ity is increasing worldwide [1­4]. Patients with biliary tract can- 46cer often present painless jaundice, pruritus, and/or anorexia. 47Hepatic resection and liver transplantation are the only curative 48options for this cancer, but the recurrence rate is high. 49The diagnosis of biliary tract cancer is usually done based on a 50combination of radiologic, histological, and tumor marker evi- 51dence, because each of these approaches alone has drawbacks. 52Tissue diagnosis, which could confirm the presence of cancer 53cells, cannot be generally performed due to tumor location, size, 54and desmoplastic characteristics [5­7]. For example, obtaining 55tissues through percutaneous fine needle aspiration is frequently 56not possible, since many of these tumors are located in the liver 57hilum amid large vascular structures [8,9]. 58Serum tumor markers, including carbohydrate antigen 19-9 59(CA19-9) and carcinoembryonic antigen (CEA), have been used 60to diagnose biliary tract cancer [10­12]. These proteins are onco- 61fetal antigens found at high levels in the fetal small intestine and 62gastrointestinal tumors. CA19-9 is mainly used in pancreatic and 63biliary tract cancer diagnosis, with sensitivities of about 80% and 6460%, respectively [10,11]. However, it can also be elevated in 65other malignancies such as pancreatic, colon, lung, and breast 66cancers, and other benign conditions such as pancreatitis, bile 67stasis, cholangitis, and inflammatory bowel disease. CEA is nor- 68mally found in embryonic entodermal tissues and fetal gastroin- 69testinal tissues, but also elevated in adult cancers, such as 70pancreatic, stomach, lung, and hepatobiliary cancers [13]. There- 71fore, these serum markers alone are not sufficient to diagnose bil- 72iary tract cancers, and other benign biliary duct complications 73can compromise their utility [14]. 74Bile cytology has been used widely for the diagnosis, because 75bile can be obtained relatively easily with Percutaneous Transhe- 76patic Cholangiography (PTC) and Endoscopic Retrograde Cholan- 77gioPancreatography (ERCP). However, ERCP cytology alone gives 78a low sensitivity of 35% [15], and additional brushing step was 79reported to improve the sensitivity [16]. This brush cytology is 80now the most common tissue sampling technique and it can be Journal of Hepatology 2009 vol. xxx j xxx­xxx Keywords: Bile; Biliary tract cancer; Metabolomics; Diagnosis. Received 14 April 2009; received in revised form 27 August 2009; accepted 1 September 2009 * Corresponding authors. Tel.: +82 32 8902548; fax: +82 32 8902549 (D.-H. Lee); Tel.: +82 32 8900935; fax: +82 32 8846726 (S. Park). E-mail addresses: ldh@inha.ac.kr (D.-H. Lee), spark@inha.ac.kr (S. Park). These authors contributed equally to this work. Abbreviations: NMR, nuclear magnetic resonance spectra; OPLS-DA, orthogonal partial least square discriminant analysis; STOCSY, statistical total correlation spectroscopy; CA19-9, carbohydrate antigen 19-9; CEA, carcinoembryonic antigen; PTC, percutaneous transhepatic cholangiography; ERCP, endoscopic retrograde cholangiopancreatography; PTBD, percutaneous transhepatic biliary drainage; ENBD, endoscopic nasobiliary drainage. Research Article JHEPAT 3124 No. of Pages 7 26 November 2009 ARTICLE IN PRESS Please cite this article in press as: Wen H et al. A new NMR-based metabolomics approach for the diagnosis of biliary tract cancer. J Hepatol (2009), doi:10.1016/j.jhep.2009.11.002 UNCORRECTED PROOF 81 performed for most biliary strictures detected by endoscopic chol- 82 angiography. Even with the brushing step, the reported sensitivity 83 is still low and variable, with its mean value around 60% [17­20]. 84 Moreover, the additional procedure could increase the risk of 85 infection [21]. Overall, diagnosis of biliary tract cancer, especially, 86 differentiating it from benign clinical conditions, is quite difficult, 87 and new diagnostic approaches are highly needed [22]. 88 Recently, a new ``-omics" approach, called metabolomics, has 89 emerged as a promising tool to differentiate individuals in dis- 90 ease or toxic conditions [23]. Compared with other omics 91 approaches, metabolomics deals with smaller molecular metabo- 92 lites in the body these change depending on the subject's envi- 93 ronmental states. It can be applied to any bio-fluid, such as 94 urine, serum, saliva, or bile, and is particularly useful for organs 95 that store or produce small molecular metabolites. Metabolomics 96 can be readily employed for new diagnostic approaches, as first 97 shown in a study with 36 coronary heart disease patients, where 98 it showed its utility as a rapid and non-invasive diagnostic tool 99 with high sensitivity and specificity [23,24]. Metabolomics has 100 subsequently shown promising results in diagnosing several can- 101 cers, such as those in breast, ovary, and prostate [25]. 102 Here, we have applied pattern recognition techniques and 103 expert data analysis to NMR spectra of biles taken from individ- 104 uals with biliary tract cancer or benign biliary tract diseases. The 105 objective of this study was to evaluate the performance of meta- 106 bolomic diagnosis of biliary tract cancer in comparison with the 107 conventional diagnostic tools including serum tumor markers 108 (CA19-9, CEA) and bile cytology. Our approach gave good distinc- 109 tion between the cancer and benign diseases and better sensitiv- 110 ity and specificity than the other approaches. This metabolomic 111 approach may become a reliable and convenient diagnostic tool 112 for biliary tract cancer. 113 Patients and methods 114 Patients 115 Informed consent was obtained from every patient enrolled in this study and the 116 study protocol conforms to the ethical guidelines of the 1975 Declaration of Hel- 117 sinki. The study was approved before initiation by the Institutional Review Board 118 at the Inha University Medical School and Hospital. 119 We prospectively obtained bile samples from patients with biliary tract can- 120 cer and benign biliary tract diseases at the Inha University Hospital (Incheon, 121 Korea) between January, 2006, and August, 2007. Patients with severe biliary sep- 122 sis were excluded from this study. This study included 17 patients with biliary 123 tract cancer and 21 patients with benign biliary tract disease (Table 1). The 124 patient groups were not matched on gender, age or disease stages to maximize 125 patient diversity. There were no exclusion criteria except for biliary sepsis, which 126 severely distorts metabolite profiles but can be easily diagnosed with other meth- 127 ods as reported previously [26]. 128 Assays and bile cytology 129 Serum CA19-9 and CEA were assayed with an immunoradiometric method and a 130 commercially available ELSA-CA19-9 and ELSA2-CEA (Cisbio International, Bed- 131 ford, MA). For patients with biliary tract cancer, routine diagnostic procedures 132 included abdominal CT scans or ultrasound. Upon identifying a duct stricture, 133 we performed percutaneous transhepatic biliary drainage (PTBD) or endoscopic 134 nasobiliary drainage (ENBD) as needed. 135 Sample collection 136 Bile samples were collected by PTBD, ENBD or during operation. The collected 137 biles were frozen at -80 °C immediately and freeze-dried in vacuo. Ten milligrams 138 of the dried samples were re-solubilized into 500 ll of a D2O + CD3OD mixture 139containing 10 mM sodium phosphate (pH 6.0). Insoluble material was removed 140by centrifugation, and 0.025% TSP was added for chemical shift referencing and 141normalization. 142NMR measurements 143All spectra were obtained by an NMR spectrometer (Bruker Biospin Avance 500) 144operating at a proton NMR frequency of 500.13 MHz. The acquisition parameters 145were essentially the same as previously reported [27,28]. The time domain data 146were Fourier transformed, phase corrected, and baseline corrected manually. This 147study made use of the NMR facility at Korea Basic Science Institute, which is sup- 148ported by Bio-MR Research Program of the Korean Ministry of Science and Tech- 149nology (E28080). 150Metabolomics data analysis 151To reduce the complexity of the NMR data for pattern recognition, the spectra 152were binned with 0.04 ppm width using an in-house Perl program. The signals 153were normalized against total integration values, and then, 0.025% TSP signal. 154The water and methanol regions were excluded. The numeric data were imported 155into statistical software. Matlab (MathWorks, Natick, MA), SIMCA-P version 11.0 156(Umetrics, Sweden), Chenomx (Spectral database; Edmonton, Alberta, Canada) 157and Excel (Microsoft, Seattle, WA) programs were used for data analysis. Orthog- 158onal projections to latent structure-discriminant analysis (OPLS-DA) were per- 159formed to distinguish cancer and benign patient groups. The statistical 160validation was performed using ``Y-scrambling" validation, where the class memTable 1. Clinical patient characteristics. Biliary tract cancer Benign biliary tract diseasef Clinical parameters (n = 17) (n = 21) Gender (M:F) 4.7:1 1.3:1 Age, yearsa 70.4 10.6 59.4 15.5 Methods of bile sampling PTBD 13 2 ENBD 3 14 Operation 1 5 Cancer stagesb I Ia: 5, Ib: 2 II IIa: 3, IIb: 2 III IIIa: 1, IIIb: 2, 11 Ic: 2 Diagnosis of cancerc Operationd 5 Operation and Bile cytologied 4 Bile cytologyd 3 Clinical and radiologicale 5 Out of 17 cancer patients, 9 patients were diagnosed by operation (Operation (5) + Operation and Bile cytology(4)). In addition, three un-operated patients showed positive in cancer cells in the drained bile. Therefore, total of 12 patients (71%) were diagnosed either by operation or bile cytology. Overall, the sensitivity of the bile cytology was 41%. For the rest of the cancer patients (five, 29%), histological examination (bile cytology, brushing cytology, guided fine needle aspiration) could not detect cholangiocarcinoma. However, radiological (cholangiography and abdominal CT), and clinical (obstructive jaundice, weight loss, abdominal pain or incidental abdominal mass detection) evidence justified the diagnosis of cholangiocarcinoma. Moreover, all of the patients died of cancer progression within one year of diagnosis, which gave additional support to our diagnosis. PTBD, Percutaneous transhepatic biliary drainage; ENBD, Endoscopic nasobiliary drainage. a Values expressed as the mean + SD (range). b According to American joint committee on cancer staging manual (2002, 6th Edition, Springer). c All of the patients died of cancer progression within one year of diagnosis. d These represent the gold standard of the biliary duct cancer diagnosis. e Radiological evidence includes cholangiography and abdominal CT. Clinical evidence includes obstructive jaundice, weight loss, abdominal pain or incidental abdominal mass detection. f One cancer patient had been treated with intrahepatic duct stone before the cholangiocarcinoma. Research Article 2 Journal of Hepatology 2009 vol. xxx j xxx­xxx JHEPAT 3124 No. of Pages 7 26 November 2009 ARTICLE IN PRESS Please cite this article in press as: Wen H et al. A new NMR-based metabolomics approach for the diagnosis of biliary tract cancer. J Hepatol (2009), doi:10.1016/j.jhep.2009.11.002 UNCORRECTED PROOF 161 bership was shuffled 200 times randomly, and the resulting Q2 and R2 values were 162 calculated. Prediction of the unknown samples was carried out by leave-one-out 163 analysis, as reported previously [29]. The conceptual explanation of these meth- 164 ods is given in Supplementary material S4. 165 Results 166 Patient characteristics 167 The biliary tract cancer group included 13 Klatskin tumors, two 168 CBD cancers, one gallbladder cancer, and one intrahepatic chol- 169 angiocarcinoma. There were 17 bile duct stones, two benign bil- 170 iary strictures, one choledochal cyst, and one other disease in the 171 benign biliary tract group. Patient characteristics of the two 172 groups were different because of the epidemiology of biliary tract 173 cancer (Table 1). Bile sampling was also different between the 174 two groups because treatment options for the two groups 175 differed. 176 NMR spectra and multivariate analysis 177 We obtained NMR spectra of bile samples from both patient 178 groups. The general spectral features were similar, with large 179 peaks in the aliphatic region (2.3­0.8 ppm) corresponding to 180 the bile acids, cholesterol, fatty acids, and other lipid compo- 181 nents, present abundantly in bile (Fig. 1). To analyze the NMR 182 data holistically and to establish the prediction model for biliary 183 tract cancer, we applied OPLS-DA multivariate analysis to the 184 NMR data. The results revealed that analysis with signals upfield 185 of 6.0 ppm gave better separation (data not shown), probably due 186 to the aliphatic nature of the bile components. Therefore, we per- 187 formed the subsequent analysis with the 0­6.0 ppm region sig- 188 nal. The OPLS-DA distinction model was obtained using one 189 predictive (Pp) and four orthogonal components (Po) (Fig. 2). 190 The majority of the normal and cancer samples appear clustered 191 in their respective regions with only a few overlaps between 192 them. The model featured an overall goodness of fit, R2 (Y), of 193 95% and an overall cross-validation coefficient, Q2 (Y), of 91%. 194 Out of the overall R2 (X) value of 0.87, 60% was structured but 195 uncorrelated to the response, and 27% was predictive. These 196 results show that there is considerable variation within each 197 group, but that our model can reliably differentiate between 198 them, even with large structured noise. 199 Statistical TOCSY analysis 200 With the efficient separation of the cancer and benign groups, we 201 further identified the variables responsible for the classification 202 rules. We utilized statistical total correlation spectroscopy (STO- 203 CSY), which can show the modeled correlation (P(corr)p) as NMR 204 lines, enabling straightforward interpretation of the variable con- 205 tributions [27,30,31]. The STOCSY plot (Fig. 3) shows that impor- 206 tant contributions for the separation come from signals at 1.50 207 (1), 1.06 (2), and 3.70 (3) ppm, which correspond to ­CH2­, ­ 208 CH3­, and ­CHn ­OR moieties that are common in bile acids. 209 Therefore, differences in the bile acid composition are impor- 210 tantly related to the class differentiation. However, the P(coor)p 211 values indicate that variations in these signals are not entirely 212 responsible for the class difference. This is not very surprising, 213 considering a previous report on coronary heart disease [23]. 214 There, only 20­30% of the variance of the most important vari- 215ables was related to the heart disease risk, but very high sensitiv- 216ity and specificity were still obtained. Therefore, the remaining 217variations in our case should result from subtle individual chem- 218ical differences in bile acids, such as the position of the double 219bonds and bile-metabolite conjugation. Benign 6 4 2 0 Cancer 6 4 2 0 H(PPM) Fig. 1. Representative 500 MHz 1 H-NMR spectra of bile samples from a benign biliary tract disease patient (top) and a biliary tract cancer patient (bottom). The spectra were taken for samples in 500 ll of D2O + CD3OD mixture containing 10 mM sodium phosphate (pH 6.0) and 0.025% TSP as a chemical shift reference. Po Pp - 200 - 100 0 100 - 70 - 50 - 30 - 10 0 10 30 50 Fig. 2. Orthogonal projections to latent structure-discriminant analysis (OPLS-DA) score plot of benign and cancer samples. Open triangle: Benign samples; Filled box: Cancer samples. The model was obtained using one predictive and four orthogonal component, with R2 (Y) of 95% and Q2 (Y) of 91%. Pp represents the predictive component and Po represents the orthogonal component. JOURNAL OF HEPATOLOGY Journal of Hepatology 2009 vol. xxx j xxx­xxx 3 JHEPAT 3124 No. of Pages 7 26 November 2009 ARTICLE IN PRESS Please cite this article in press as: Wen H et al. A new NMR-based metabolomics approach for the diagnosis of biliary tract cancer. J Hepatol (2009), doi:10.1016/j.jhep.2009.11.002 UNCORRECTED PROOF 220Statistical validation 221The separation result of the cancer and normal patients was sub- 222jected to ``Y-scrambling" statistical validation to test the possibility 223of chance correlation. We randomly permutated the Y -variable 224(cancer or benign group designation), re-built the statistical model, 225and observed the trends of the predictive power and goodness of fit 226at each step. Two hundred rounds of such reshuffling gave coher- 227ent decreases in both parameters and the extrapolated value of 228the Q2 of 0.3 (Fig. 4), which shows that the separation model is 229statistically sound, and that its high predictability is not due to 230over-fitting of the data. Although the current study may not cover 231all the possible variations in the patients, such as the bile duct 232obstruction time, we believe our validation through randomiza- 233tion of the Y-variable suggests that those variations should be 234orthogonal to, and thus not be a major factor for our differentiation 235between the cancer and normal groups. Unrelated variations were 236most likely partitioned into the orthogonal components of the pre- 237diction model and, thus, should not affect the predictability. 238Among the unrelated variations, gender and age could provide 239a large source of variation that may affect the differentiation. 240Therefore, we analyzed the patient data in subgroups that are H (ppm) Benign Cancer P(cov)p - 0.7 - 0.6 - 0.4 - 0.3 - 0.1 0 0.1 0.3 0.4 0.6 0.70.4 0.3 0.2 0.1 0.0 - 0.1 - 0.2 - 0.3 - 0.4 S - TOCSY 013 2456 P (corr) p 1 2 3 Fig. 3. Variable contributions from statistical total correlation spectroscopy (STOCSY). The color scale on the right indicates P(coor)p. The P(cov)p represents the modeled covariance and P(coor)p represents the modeled correlation. Peaks with labels are mentioned in the text (1: 1.50 ppm, 2: 1.06 ppm, 3: 3.70 ppm). 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Correlation coefficient to original values - 0.3 - 0.2 - 0.1 0. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Fig. 4. Statistical validation of the OPLS-DA analysis result by ``Y-scrambling". Two hundred permutations were performed, and the resulting R2 and Q2 values were plotted. Open triangle: R2 ; Filled square: Q2 . The solid line represents the regression line for R2 and the dashed line for Q2 . Research Article 4 Journal of Hepatology 2009 vol. xxx j xxx­xxx JHEPAT 3124 No. of Pages 7 26 November 2009 ARTICLE IN PRESS Please cite this article in press as: Wen H et al. A new NMR-based metabolomics approach for the diagnosis of biliary tract cancer. J Hepatol (2009), doi:10.1016/j.jhep.2009.11.002 UNCORRECTED PROOF241 not affected by these biases. First, we performed the differentiation 242 with only male patients, as the cancer group is primarily male and 243 the benign group has relatively even distribution. The actual result 244 (see Supplementary Fig. S1A) exhibited very similar differentiation 245 as our original model with all the patients (see Fig. 2), which con- 246 firms that our original differentiation is not based on the gender. If 247 the original model had been influenced by the gender, the male- 248 only analysis should have given much poorer, or even no, discrim- 249 ination between the cancer and benign groups. We also tested the 250 influence of age in our model. In separate analyses with younger 251 (see Supplementary Fig. S1B) and older groups (see Supplementary 252 Fig. S1C), the differentiation of cancerand benign groups were even 253 better than the one with all the patients (see Fig. 2). As stated 254 above, if our original model had been influenced mainly by age, 255 the differentiation should have been much worse in each sub- 256 group. These results again confirm the validity of our OPLS-DA 257 approach which can effectively exclude these possible confound- 258 ing factors in differentiating the groups based on the feature of 259 interest (cancer status, in our case). 260 Prediction and diagnostic performance test 261 To estimate the actual performance of our OPLS-DA model in 262 diagnosing biliary tract cancer, we performed a leave-one-out 263predictive test. For this, we left-out one patient sample at a time 264and constructed the OPLS-DA prediction model with the rest of 265the data (a training set). The prediction model was constructed 266with the same number of predictive and orthogonal components 267as the original OPLS-DA classification model. The class member- 268ship of the left-out sample was predicted using an a priori cut- 269off value of 0.5. This procedure was repeated until each and every 270sample had been tested once. Of the 21 benign disease samples, 27118 were predicted correctly as benign, and of the 17 cancer sam- 272ples, 15 were predicted correctly as cancer (Fig. 5). Therefore, our 273OPLS-DA metabolomics prediction model exhibited a sensitivity 274of 88% and a specificity of 81% for biliary tract cancer diagnosis, 275which is significantly better than conventional serum markers 276or cytology (Table 2). 277Discussion 278Biliary tract cancer is highly lethal and only surgical excision of 279the tumor can improve survival [32­35]. However, biliary tract 280cancer is often presented locally advanced, and the majority of 281the patients are elderly, with critical co-morbidity which 282increases the risk of operation. In this respect, it has been sug- 283gested that neither more advanced surgical techniques nor radi- N5 N6 N8 N7 N10 N11 N15 N9 N12 N13 N17 N19 N18 N20 N21 N22 N23 N24 N25 N26 N16 C13 C15 C17 C18 C16 C12 C11 C10 C9 C8 C7 C5 C6 C3 C2 C4 C1 - 0.1 0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 Obs ID (Primary) Predicted 0 10 20 30 N24 N25 N8 C10 C18 Fig. 5. Prediction of cancer and benign patients using leave-one-out analysis. One patient sample (unknown) was left-out at a time and an OPLS-DA prediction model was constructed with the rest of the data. The class membership of the left-out samples was predicted using an a priori cut-off value of 0.5 (dashed line) [23]. Cancer samples: black box; Benign samples: Red circle. The Y values of the filled symbols are from the analysis using the entire dataset. In the case of mis-classified samples, the predicted Y values from the leave-one-out analysis are also shown as open boxes (cancer patients) and open circles (normal patients). JOURNAL OF HEPATOLOGY Journal of Hepatology 2009 vol. xxx j xxx­xxx 5 JHEPAT 3124 No. of Pages 7 26 November 2009 ARTICLE IN PRESS Please cite this article in press as: Wen H et al. A new NMR-based metabolomics approach for the diagnosis of biliary tract cancer. J Hepatol (2009), doi:10.1016/j.jhep.2009.11.002 UNCORRECTED PROOF 284 ation therapy is likely to improve survival [36]. Therefore, cur- 285 rently, efforts are being directed to prevention and reliable detec- 286 tion. However, current diagnostic tools, such as serum tumor 287 markers and bile cytology, have limited utility for differentiating 288 cancer and benign diseases, and new diagnostic methods are 289 highly needed. 290 Here, we applied a metabolomics approach to biles obtained 291 directly from patients in order to assess its accuracy and reliabil- 292 ity in diagnosing biliary tract cancer. Our approach showed better 293 performance in terms of both specificity and sensitivity than con- 294 ventional data obtained from literature and our own patients 295 (Table 2). Although CEA showed perfect specificity for our 296 patients, its utility was significantly compromised due to its poor 297 sensitivity. In general, sensitivity is more important than specific- 298 ity in serious diseases such as cancer. Bile cytology, in theory, can 299 deliver perfect specificity, as it directly observes cancer cells in 300 the samples, but its reported sensitivity is rather poor to range 301 between 35% and 40% [15,37]. We also obtained 41% sensitivity 302 using bile cytology for our patients. Although brushing has been 303 shown to increase cytology sensitivity by about 20% [17­20], it 304 requires additional invasive steps such as ERCP or EST, which 305 could increase the risk of pancreatitis [21]. The final sensitivity 306 of brush cytology is still about 60%, significantly lower than our 307 metabolomics results. Our metabolomics approach gave high val- 308 ues for both sensitivity and specificity. Therefore, we believe that 309 the metabolomic diagnosis may be more clinically useful than 310 conventional techniques in biliary tract cancer diagnosis. Obvi- 311 ously, as with any other new diagnostic approaches, there are 312 limitations to our study. One such possibility is the effects of bil- 313 iary infection without clinical evidence of sepsis on the metabolic 314 profiles. This potential confounding factor, though, can be diag- 315 nosed by culture or PCR, and therefore, may be an interesting 316 subject for later studies. 317 To get deeper insights into the metabolic difference, we ana- 318 lyzed our data with targeted metabolic profiling for four 319 metabolites involved importantly in energy metabolic pathways: 320 choline, lactate, citrate and glucose. We used student's indepen- 321 dent t-test to see if the contents of these metabolites are 322 statistically different between cancer and benign groups (see 323 Supplementary Fig. S2). While choline (p > 0.85), lactate (p > 324 0.79), and glucose (p > 0.24) did not show any relevant differ- 325 ences, citrate level was statistically higher in cancer groups 326 (p < 0.05). The higher content of citrate is interesting, as it is 327 the starting molecule of the TCA cycle, the hallmark of the aerobic 328 energy metabolism. Citrate is formed through a condensation 329 reaction between oxaloacetate and acetyl CoA. The latter is also 330 the precursor of the cholesterol which is metabolized into bile 331 acids. The higher level of citrate in the cancer group might result 332 from the low dependence of the cancer cells on the aerobic 333energy metabolism consuming citrate in the TCA cycle, consistent 334with the Warburg effect in cancer cells. High level of citrate is 335expected to affect the concentration of acetyl CoA, its immediate 336precursor, which in turn can affect the bile acid formation. As cit- 337rate also has ­CH2­ group, this is consistent with our initial sug- 338gestion on contributing signals .Although confirmation of the 339above will require detailed flux analysis of all the involved path- 340ways, our metabolomic data provide an interesting initial evi- 341dence for the link between the energy metabolism and bile acid 342compositions in the cancer group. 343In addition to our main goal of differentiating cancer and 344benign patients, we also tested if our approach can differentiate 345the various stages of the biliary duct cancer. Although individual 346differentiation of stages I, II, and III were not satisfactory (data 347not shown), we obtained a good separation between stages I 348and II combined against stage III (see Supplementary Fig. S3). 349Although the number of patients is not large for each group, these 350data suggest that it may be possible to differentiate between rel- 351atively early (I and II) and later stage biliary duct cancers with our 352metabolomic approach. 353It should be noted that our metabolomics approach is ``non- 354invasive", as it uses bile that had been drained for therapeutic 355purposes and required no separate collection steps. In contrast, 356brush bile cytology requires additional steps, and serum markers 357require blood drawing. This fact also alleviates ethical problems, 358the inconvenience of additional visits, or pain for sample collec- 359tion for patients, providing a more convenient option. 360An efficient diagnostic method is best developed with tissues 361or bio-fluids that are cognate to the organs of interest. For exam- 362ple, urine metabolites have been used to predict kidney cancer or 363allograft rejection [29,38]. Here, we used bile for biliary duct can- 364cer diagnosis. Bile passes through the biliary duct before being 365secreted into the intestine, during which time it has direct con- 366tact with any surrounding cancer tissue. Especially in obstructive 367bile duct diseases, such as those targeted in this study, bile stays 368in the ducts for a long time, thus likely reflecting differences in 369the ductal epithelial cells. Conventional serum markers, such as 370CA-19-9 and CEA are detected from serum and, therefore, could 371reflect changes in other tissues, including colon or pancreatic 372cancers. Bile cytology, although using bile, may not always be 373able to retrieve cancer cells from the tissue, resulting in low 374sensitivity. Therefore, bile metabolomics seems more theoreti- 375cally relevant for the biliary tract cancer diagnosis than those 376approaches. 377Currently, biliary tract cancer is diagnosed by multiple criteria 378based on computerized tomography, magnetic resonance imag- 379ing, bile cytology, endoscopic ultrasonography, serum markers, 380and positron emission tomography. Our study shows that a met- 381abolomics approach, by itself, can differentiate biliary tract can- 382cer from benign diseases with high reliability. To the best of 383our knowledge, this study is the first report of a metabolomics 384diagnostic approach in the human hepatobiliary system outper- 385forming other conventional clinical criteria. A study with larger 386patient groups and standardized protocols could eventually lead 387to a dependable diagnostic tool for biliary tract cancer. 388Acknowledgements 389The authors who have taken part in this study declared that they 390do not have anything to declare regarding funding from industry 391or conflict of interest with respect to this manuscript. Table 2. Comparison of the diagnostic performance between conventional and metabolomicQ1 . Criteria CA19-9** CEA Bile cytology Metabolomics [Reference] [22] [9] [15,20] [current study] Sensitivity 81% (73%) 20% (68%) 41% (35­61 %) 88% Specificity 53% (63%) 100% (82%) N/A 81% * The numbers indicate the values obtained from the patients enrolled in the current study. The numbers in the parenthesis are from the literature. ** Cut-off value of >37 U/mL (both reference and our study). Cut-off value of >5.2 ng/mL in primary sclerosing cholangitis patients and cutoff value of 6.0 ng/mL in our study. 61% was obtained using brush cytology. Research Article 6 Journal of Hepatology 2009 vol. xxx j xxx­xxx JHEPAT 3124 No. of Pages 7 26 November 2009 ARTICLE IN PRESS Please cite this article in press as: Wen H et al. A new NMR-based metabolomics approach for the diagnosis of biliary tract cancer. J Hepatol (2009), doi:10.1016/j.jhep.2009.11.002 UNCORRECTED PROOF 392 Appendix A. Supplementary data 393 Supplementary data associated with this article can be found, in 394 the online version, at doi:10.1016/j.jhep.2009.11.002. 395 References 396 [1] Patel T. Increasing incidence and mortality of primary intrahepatic cholan- 397 giocarcinoma in the United States. Hepatology 2001;33:1353­1357. 398 [2] El Rassi ZE, Partensky C, Scoazec JY, Henry L, Lombard-Bohas C, Maddern G. 399 Peripheral cholangiocarcinoma: presentation, diagnosis, pathology and 400 management. Eur J Surg Oncol 1999;25:375­380. 401 [3] Taylor-Robinson SD, Toledano MB, Arora S, Keegan TJ, Hargreaves S, Beck A, 402 et al. Increase in mortality rates from intrahepatic cholangiocarcinoma in 403 England and Wales 1968­1998. Gut 2001;48:816­820. 404 [4] Patel T. Worldwide trends in mortality from biliary tract malignancies. BMC 405 Cancer 2002;2:10. 406 [5] Ohtsuka M, Ito H, Kimura F, Shimizu H, Togawa A, Yoshidome H, et al. Results 407 of surgical treatment for intrahepatic cholangiocarcinoma and clinicopatho- 408 logical factors influencing survival. Br J Surg 2002;89:1525­1531. 409 [6] Ueno N, Sano T, Kanamaru T, Tanaka K, Nishihara T, Idei Y, et al. 410 Adenosquamous cell carcinoma arising from the papilla major. Oncol Rep 411 2002;9:317­320. 412 [7] Saito M, Hige S, Takeda H, Tomaru U, Shibata M, Asaka M. Combined 413 hepatocellular carcinoma and cholangiocarcinoma growing into the com- 414 mon bile duct. J Gastroenterol 2001;36:842­847. 415 [8] Matsumoto A, Imamura M, Akagi Y, Kaibara A, Ohkita A, Mizobe T, et al. A 416 case report of disseminated recurrence of inferior bile duct carcinoma in 417 PTCD fistula. Kurume Med J 2002;49:71­75. 418 [9] Siqueira E, Schoen RE, Silverman W, Martin J, Rabinovitz M, Weissfeld JL, 419 et al. Detecting cholangiocarcinoma in patients with primary sclerosing 420 cholangitis. Gastrointest Endosc 2002;56:40­47. 421 [10] Steinberg W. The clinical utility of the CA 19-9 tumor-associated antigen. 422 Am J Gastroenterol 1990;85:350­355. 423 [11] Perkins GL, Slater ED, Sanders GK, Prichard JG. Serum tumor markers. Am 424 Fam Physician 2003;68:1075­1082. 425 [12] Qin XL, Wang ZR, Shi JS, Lu M, Wang L, He QR. Utility of serum CA19-9 in 426 diagnosis of cholangiocarcinoma: in comparison with CEA. World J Gastro- 427 enterol 2004;10:427­432. 428 [13] Carpelan-Holmstrom M, Louhimo J, Stenman UH, Alfthan H, Haglund C. CEA, 429 CA 19-9 and CA 72-4 improve the diagnostic accuracy in gastrointestinal 430 cancers. Anticancer Res 2002;22:2311­2316. 431 [14] Bjornsson E, Kilander A, Olsson R. CA 19-9 and CEA are unreliable markers 432 for cholangiocarcinoma in patients with primary sclerosing cholangitis. 433 Liver 1999;19:501­508. 434 [15] Davidson B, Varsamidakis N, Dooley J, Deery A, Dick R, Kurzawinski T, et al. 435 Value of exfoliative cytology for investigating bile duct strictures. Gut 436 1992;33:1408­1411. 437 [16] Osnes M, Serck-Hanssen A, Myren J. Endoscopic retrograde brush cytology 438 (ERBC) of the biliary and pancreatic ducts. Scand J Gastroenterol 439 1975;10:829­831. 440 [17] de Bellis M, Fogel EL, Sherman S, Watkins JL, Chappo J, Younger C, et al. 441 Influence of stricture dilation and repeat brushing on the cancer detection 442 rate of brush cytology in the evaluation of malignant biliary obstruction. 443 Gastrointest Endosc 2003;58:176­182. 444 [18] Kurzawinski T, Deery A, Davidson BR. Diagnostic value of cytology for biliary 445 stricture. Br J Surg 1993;80:414­421. 446 [19] Logrono R, Kurtycz DF, Molina CP, Trivedi VA, Wong JY, Block KP. Analysis of 447 false-negative diagnoses on endoscopic brush cytology of biliary and 448 pancreatic duct strictures: the experience at 2 university hospitals. Arch 449 Pathol Lab Med 2000;124:387­392. 450[20] Mahmoudi N, Enns R, Amar J, AlAli J, Lam E, Telford J. Biliary brush cytology: 451factors associated with positive yields on biliary brush cytology. World J 452Gastroenterol 2008;14:569­573. 453[21] Vandervoort J, Soetikno RM, Montes H, Lichtenstein DR, Van Dam J, 454Ruymann FW, et al. Accuracy and complication rate of brush cytology from 455bile duct versus pancreatic duct. Gastrointest Endosc 1999;49:322­327. 456[22] Kim HJ, Kim MH, Myung SJ, Lim BC, Park ET, Yoo KS, et al. A new strategy for 457the application of CA19-9 in the differentiation of pancreaticobiliary cancer: 458analysis using a receiver operating characteristic curve. Am J Gastroenterol 4591999;94:1941­1946. 460[23] Brindle JT, Antti H, Holmes E, Tranter G, Nicholson JK, Bethell HW, et al. Rapid 461and noninvasive diagnosis of the presence and severity of coronary heart 462disease using 1 H-NMR-based metabonomics. Nat Med 2002;8:1439­1444. 463[24] Goodacre R, Vaidyanathan S, Dunn WB, Harrigan GG, Kell DB. Metabolomics 464by numbers: acquiring and understanding global metabolite data. Trends 465Biotechnol 2004;22:245­252. 466[25] Claudino WM, Quattrone A, Biganzoli L, Pestrin M, Bertini I, Di Leo A. 467Metabolomics: available results, current research projects in breast cancer, 468and future applications. J Clin Oncol 2007;25:2840­2846. 469[26] Park SY, Park CH, Cho SB, Yoon KW, Lee WS, Kim HS, et al. The safety and 470effectiveness of endoscopic biliary decompression by plastic stent place- 471ment in acute suppurative cholangitis compared with nasobiliary drainage. 472Gastrointest Endosc 2008;68:1076­1080. 473[27] Kang J, Choi MY, Kang S, Kwon HN, Wen H, Lee CH, et al. Application of a 1 H 474nuclear magnetic resonance (NMR) metabolomics approach combined with 475orthogonal projections to latent structure-discriminant analysis as an 476efficient tool for discriminating between Korean and Chinese herbal 477medicines. J Agric Food Chem 2008;56:11589­11595. 478[28] Kang J, Lee S, Kang S, Kwon HN, Park JH, Kwon SW, et al. NMR-based 479metabolomics approach for the differentiation of ginseng (Panax ginseng) 480roots from different origins. Arch Pharm Res 2008;31:330­336. 481[29] Kim K, Aronov P, Zakharkin SO, Anderson D, Perroud B, Thompson IM, et al. 482Urine metabolomic analysis for kidney cancer detection and biomarker 483discovery. Mol Cell Proteomics 2009;8:558­570. 484[30] Wiklund S, Johansson E, Sjostrom L, Mellerowicz EJ, Edlund U, Shockcor JP, 485et al. Visualization of GC/TOF-MS-based metabolomics data for identifica- 486tion of biochemically interesting compounds using OPLS class models. Anal 487Chem 2008;80:115­122. 488[31] Cloarec O, Dumas ME, Craig A, Barton RH, Trygg J, Hudson J, et al. Statistical 489total correlation spectroscopy: an exploratory approach for latent biomarker 490identification from metabolic 1 H NMR data sets. Anal Chem 2005;77: 4911282­1289. 492[32] Nakeeb A, Pitt HA, Sohn TA, Coleman J, Abrams RA, Piantadosi S, et al. 493Cholangiocarcinoma: a spectrum of intrahepatic, perihilar, and distal 494tumors. Ann Surg 1996;224:463­473, discussion 473­465. 495[33] Henson DE, Albores-Saavedra J, Corle D. Carcinoma of the gallbladder: 496histologic types, stage of disease, grade, and survival rates. Cancer 4971992;70:1493­1497. 498[34] Henson DE, Albores-Saavedra J, Corle D. Carcinoma of the extrahepatic bile 499ducts: histologic types, stage of disease, grade, and survival rates. Cancer 5001992;70:1498­1501. 501[35] Farley DR, Weaver AL, Nagorney DM. ``Natural history" of unresected 502cholangiocarcinoma: patient outcome after noncurative intervention. Mayo 503Clin Proc 1995;70:425­429. 504[36] de Groen PC, Gores GJ, LaRusso NF, Gunderson LL, Nagorney DM. Biliary tract 505cancers. N Engl J Med 1999;341:1368­1378. 506[37] Park SW, Song SY, Chung JB, Jung BG, Moon YM, Kang JK, et al. Combined 507endoscopic transpapillary biopsy and exfoliative cytology for the diagnosis 508of bile duct cancer. Korean J Gastrointest Endosc 1999;19:588­596. 509[38] Wang JN, Zhou Y, Zhu TY, Wang X, Guo YL. Prediction of acute cellular renal 510allograft rejection by urinary metabolomics using MALDI-FTMS. J Proteome 511Res 2008;7:3597­3601. 512 JOURNAL OF HEPATOLOGY Journal of Hepatology 2009 vol. xxx j xxx­xxx 7 JHEPAT 3124 No. of Pages 7 26 November 2009 ARTICLE IN PRESS Please cite this article in press as: Wen H et al. A new NMR-based metabolomics approach for the diagnosis of biliary tract cancer. J Hepatol (2009), doi:10.1016/j.jhep.2009.11.002