Enhanced Detection of Early Hepatocellular Carcinoma by Serum SELDI-TOF Proteomic Signature Combined with Alpha-Fetoprotein Marker

A total of 240 patient subjects (120 HCC patients and 120 LC patients) were included in this study, and these patients were presented to the Queen Mary Hospital (Pokfulam, Hong Kong) between 2003 and 2005. The demographics and clinicopathological data are summarized in Table 1. Informed consent was obtained from these patients. This study was approved by the Joint University/Hospital Authority’s Institutional Review Board. All patients were age- and gender-matched Chinese having the same HBsAg profile. Preliminary diagnosis of HCC was based on preoperative investigations, which included blood biochemistry, AFP assay, chest x-ray, percutaneous ultrasonography, computed tomography, and hepatic angiography in selected patients. Liver function was assessed using Child-Pugh classification. Histopathological examination of the resected specimens was conducted by board-certified pathologists to confirm the diagnosis. Tumor stages were based on the AJCC staging system. Early HCC was defined as stage I (T1N0M0). Diagnosis of cirrhosis was based on liver histology as well as laboratory and imaging evidence of hepatic decompensation or portal hypertension. The median follow-up period of cirrhotic patients was 12 months with no sign of HCC development.

The training set consisted of 60 subjects with pathologically validated LC but without HCC and 60 HCC patients who were mainly at AJCC stage I and II (70%). Another 60 LC and 60 HCC patients were included in the blinded validation set. Pooled serum samples from healthy subjects were used as interassay quality control. Blood samples were taken prior to surgery and 3 months after surgery. Extracted serum was stored at −80°C. A commercial chemiluminescent immunoassay kit (ACS 180SE, Bayer, Leverkuesen, Germany) was used to determine the serum AFP level. In clinical practice, the upper normal range of AFP was 8–10 ng/mL and patients having >20 ng AFP/mL were considered to have liver abnormality. Therefore, this same AFP cutoff level was used in the current study to discriminate HCC in high-risk populations.23

Serum samples were thawed completely and centrifuged to remove debris at 9600 g for 10 min at 4°C. SELDI-TOF-MS profiling using the IMAC-3 chip (Ciphergen Biosystems, Fremont, CA) was done as previously described.18,22 In brief, 5 μL of the cleared serum was mixed with 5 μL phosphate-buffered saline (PBS) containing 8 M urea and 2% CHAPS and then diluted with 108 μL binding buffer (0.25 M sodium chloride and 1% CHAPS in PBS). After that, 100 μL of diluted serum samples were loaded on each array spot with a robotic bioprocessor and incubated for 1 h at room temperature. After washing off unbound molecules, each array was spotted twice with 0.5 μL of freshly prepared EAM matrix solution (sinapinic acid saturated with 0.5% trifluoroacetic acid and 50% acetonitrile).

Mass spectrometric analysis was performed by SELDI-TOF-MS in a PBS-IIc ProteinChip reader (Ciphergen Biosystems) according to an automated data collection protocol. Serum spectrum was generated by averaging the results of 65 laser shots at a laser intensity of 240, a detector sensitivity of 8, in which the qualified mass peaks (signal-to-noise ratio >5) with mass-to-charge ratios (m/z) in an optimized range of 3000–12,500 Da were detected. Spectral analysis was carried out by the Biomarker Wizard program (version 3.1; Ciphergen Biosystems). After the total ion current normalization, the spectra data of the training set was analyzed using the Biomarker Pattern software (version 4.1, Ciphergen Biosystems) for model building. A classification and regression tree (CART) analysis was performed, as previously described, based on the differentially expressed protein peaks in sera of LC and HCC patients.24

The construction of CART included 2 steps, namely the tree construction step and the tree pruning step. For the tree construction process, the BPS was used to search for the best peak with defined cutoff level so that the training dataset was split into two daughter nodes. The splitting decision was based on the peak intensity of a sample. Samples went to the left daughter node if their peak intensities were equal to or less than the cutoff intensity value; otherwise, the samples would go to the right daughter node. The software continued to repeat this splitting process on each daughter node in this manner until no further gain in the classification was achieved and terminal nodes were produced. Classification of the terminal nodes was decided by a group of samples (i.e., HCC and LC), which represented the majority of samples in that group. In the second step of CART, the classification tree was cut down to a desired size that yielded the least classification error. In this study, a decision tree was generated using the Gini method with nonlinear combinations. A 10-fold cross-validation test was used for the evaluation of the classification error. The performance of the classification algorithm was then challenged in a blind study in which the test set consisted of a new batch of 120 serum samples (60 samples from each group).

Pooled normal sera, which were spotted randomly on different chips, were used to test for the consistency of the SELDI-TOF-MS spectra among different protein chips. The coefficients of variance (CV) for the three highest peaks with relative intensities above 25 were <20%, and their m/z ratios were <0.1%, akin to that of other studies.25,26 This suggested that little variation in the protein spectra was present for the same sample among different chips, implicating the stability and reproducibility of the SELDI-TOF-MS spectra.

The serum spectra from 60 HCC and 60 LC subjects were used as the training set. After intensity normalization, 80 peaks were found to be statistically significant between HCC and LC patients (P < .05). Five protein peaks were regarded as classifiers to construct the CART. Figure 1a represents spectral views showing some of these protein peaks in the HCC sera compared with the LC sera. The spreadsheets of normalized intensities of these peaks were exported into the Biomarker Patterns software for CART analysis. Of the many classification trees generated by using different settings of Gini, advance, cost of BPS software, the most optimal classification tree with lowest error cost was eventually established. A CART with 6 terminal nodes was established, based on the 5 discriminative peaks with m/z of 3324, 3994, 4665, 4795, and 5152 that could separate HCC patients from LC patients (Table 2) (Fig. 1b). When the 5 proteomic peaks were used as classifiers, this CART algorithm would yield an accuracy rate of 96.7% (116 of 120), a sensitivity of 98.33%, and a specificity of 95% for predicting HCC in the training set. To validate the above finding, a blinded set consisting of a separate cohort of 60 HCC patients and 60 LC patients was used. Serum protein profiles of these samples were collected and treated the same as those in the training set. The normalized peak intensities in the blinded set were subjected to analysis. A sensitivity of 83% (50 of 60), a specificity of 92% (55 of 60), and an accuracy rate of 87.5% were yielded in the blind set. Also, 87% (40 of 46) of the HCC patients in stage I and 89% of patients in stage II were detected by the classification algorithm (Table 3).

To strengthen these findings, we performed additional experiments to determine the change of proteomic profiles in 10 pairs of serum samples collected from newly recruited HCC patients, before and 3 months after surgery. As shown in Fig. 2a, there is a decrease of the serum level of m/z 3324 peak after removal of the tumors. Although the difference is not significant (P = .131), there is a trend of declination in the serum from the postsurgery group. Interestingly, the serum level of m/z 4795 peak, which was initially lower in the HCC group than the LC group, was significantly increased after surgery (P = .01) (Fig. 2b).

The sensitivity and specificity of SELDI-TOF-MS pattern and AFP₂₀ to predict different stages of HCC were examined in 120 HCC patients from training and validation set. A sensitivity of 87% (40 of 46) for detecting stage I HCC patients, 89% (32 of 36) for detecting stage II and 97% (37 of 38) for detecting late-staged HCC patients (stage III) were obtained by CART analysis (Table 3). When AFP₂₀ was used to predict HCC stages, a sensitivity of 54% (25 of 46) for stage I HCC patients, 72% (26 of 36) for stage II HCC, and 84% (32 of 38) for late-staged HCC patients (stage III) were obtained, which was less favorable than CART, especially for detecting stage I HCC (P < .05). The diagnostic odds ratios (DOR) were also calculated in this study to compare the CART analysis and AFP₂₀ (Table 4) for predicting HCC in the blinded samples. DOR is a single indicator for diagnostic performance, such that higher DOR value means better discriminative power.27 The calculated DOR of CART algorithm was 92.72, implying that the odds for positivity among patients with HCC was 92.72 times higher than the odds for positivity among patients without HCC. The calculated DOR for AFP₂₀ was 9.11 only. Based on these values, CART algorithm performed 10 times better than AFP₂₀ in detecting HCC from LC patients (92.72 vs 9.11) (Table 4). Interestingly, a combined use of both CART and AFP₂₀ increased the sensitivity to 95%, with at least 1 of the 2 tests being positive, and the specificity to 98%, when both tests were negative. The estimated DOR was 931 for the combined tests, which was better than the individual test alone (931 vs 92.72 and 9.11) (Table 4). Despite that, no significant correlation was found between the intensity values of each discriminative peak and some of the clinicopathological data of HCC patients—age, sex, AFP level, tumor size, and tumor stage (data not shown).