MALDI-TOF MS analysis of permethylated N-glycans, which were enzymatically detached from serum glycoproteins, allowed relative quantification of 83 oligosaccharides. We analyzed a total of 202 serum samples using previously described methods with only minor modifications (30
). Comparison of average spectra in HCC cases (n
= 73) and controls (n
= 77) showed marked differences in glycan abundance (). Analysis of the 83 peak intensities by t
test showed significant differences (P
< 0.01) in the abundance of 57 glycans; we chose P
< 0.01 to adjust for the multiple comparisons. Details of the analysis are presented in Supplementary Table S1
together with description of the known glycan structures. Supplementary Fig. S1
provides a graphical overview of the spectra in the mass range of 1.5 to 5.5 kDa.
Average MALDI-TOF spectra in the mass range of 1.5 to 5.5 kDa in HCC cases (n = 73) and population controls (n = 77; inverted).
Structural composition of 46 of the 83 N-glycans was determined by a combination of enzymatic sequencing and tandem MS as described previously (29
). To select candidate markers for detection of HCC, we focused on the glycans with known structure to allow a robust validation of our selection in other laboratories. We selected 6 of the 48 N-glycans as candidate markers for classification of HCC by ACO-SVM computational methods, as described previously (30
). These six glycans were selected with >50% frequency in 100 repeats of the ACO-SVM algorithm carried out with 25 HCC and 24 population control spectra. Association of the glycans and covariates (age, gender, HCV and HBV infections, and smoking) with HCC (dependent variable) was analyzed by univariate logistic regression using 73 HCC and 77 control spectra (Supplementary Table S2
). Glycan intensities were dichotomized by the median value in population controls; the analysis of glycans as continuous variables did not substantially affect the outcome (data not shown). The logistic regression analysis showed that serologic markers of viral infections and five of the six selected glycans were strongly associated with HCC. The association of the sixth glycan with HCC bordered on significance and became significant after adjustment for HCV infection (see below).
To evaluate the association of these N-glycans with viral infections, we analyzed the association of each of the six N-glycans (independent variable) and viral infections (dependent variable) in the population controls (32% positive for HCV antibodies and 52% positive for HBV antibodies; see ). None of the selected N-glycans was associated with the presence of HBV antibodies; glycans 2 and 4 were associated with HCV antibodies (Supplementary Table S3
). Next, we used six multivariate regression models to evaluate the association of each glycan (independent variable) with HCC (dependent variable), following an adjustment for matching variables (age and gender) and for markers of HCV infection (). All six N-glycans were associated with HCC following the adjustment. We did not include HCV RNA in the regression models because it is correlated with anti-HCV (correlation coefficient = 0.823).
Association of glycans with HCC
Next, we examined the six glycans in comparison with the CLD control group, which is the most clinically relevant group in need of new markers for the detection of HCC. This group of participants had a biopsy-confirmed fibrosis (n
= 22) or cirrhosis (n
= 25). Glycans 1, 3, and 6 were less abundant in HCC cases, whereas the three remaining N-glycans increased in intensity from population controls to CLD to HCC. Multivariate logistic regression comparing HCC cases (n
= 73) with CLD controls (n
= 52) showed that glycans 1, 5, and 6 remained significantly associated with HCC after adjustment for age and gender (). We did not adjust for viral infection in this analysis because all participants in the CLD group were HBV negative and HCV positive and HBV was not associated with any of the glycans (see Supplementary Table S3
). Because all CLD and 80% of HCC cases carry HCV viral infection, the result strongly suggests that the observed change in N-glycans is associated with HCC. Descriptive statistics for the six glycans in population controls, CLD controls, and HCC are presented in Supplementary Table S4
Comparison of HCC with CLD controls
shows the ROC curves of the three individual glycans that are different compared with CLD controls, as well as their combination, in a blinded, independent validation set of HCC cases (n = 48) and CLD controls (n = 27). The area under the ROC curve for individual glycans ranged from 89% to 93%, whereas the combined classifier has a sensitivity of 90% and specificity of 89% in the blinded, independent validation set. Glycans 1 and 6 are triantennary and tetraantennary complex glycans that decrease in HCC. Glycan 5 is a bisecting glycan that increases in HCC patients. This is consistent with the general trends of changes observed in our study as discussed below.
Area under the ROC curve (AuROC) for three glycans comparing a blinded validation set of HCC cases (n = 47) and CLD controls (n = 27).
To further evaluate the potential of the three glycans for early detection of HCC, we analyzed the progressive glycan changes from fibrosis to cirrhosis and early cancer (stage I and II) disease (). It is interesting to note that 17 of the 18 early cancers and 23 of 25 cirrhotic controls were classified correctly by the three glycans. The prediction accuracy and suggest that the three selected glycans separate efficiently cirrhosis controls from early-stage HCC; we verified by regression analysis that all three glycans are significantly different at P < 0.05 between cirrhosis controls (n = 25) and early-stage HCC (n = 18). Because AFP was used as a selection criterion for the CLD group (<100 ng/mL included), we could not compare the prediction accuracy directly to AFP. However, 30% of the cases had AFP <200 ng/mL; of these 22 HCC cases, 18 were correctly classified by the three glycans.
Fig. 3 Box plots comparing three selected glycans in the following patient groups: fibrosis (n = 22; A), cirrhosis (n = 25; B), and early HCC (stage I and II; n = 18; C). Both the fibrosis and cirrhosis groups are significantly different from the early-stage (more ...)