signature=c8626889784bc3d331991756aab80078,Autoantibody signature in hepatocellular carcinoma using ...

Study design

This study included three phases (Fig. 1): discovery phase (I), test phase (II), and validation phase (III). In the discovery phase (I), serum samples from 50 HCC and 50 healthy were enrolled. These 100 samples were all obtained from Zhongshan Hospital and individually profiled on HuProt arrays for screening candidate proteins and fabricating the HCC-focused arrays. Then, 282 HCC, 130 cirrhotic, and 164 healthy were collected from Zhongshan Hospital and used for model construction in the test phase (II). Finally, 279 HCC, 119 cirrhotic, and 179 healthy collected from Eastern Hepatobiliary Surgery Hospital and Cancer Hospital of Guangxi Medical University were used for independent verification in the validation phase (III). The clinical data of patients are summarized in Additional file 5: Table S1. Clinical variables of each group in the test phase (II) and validation phase (III) were compared by Pearson’s chi-squared test, and there was no statistical significance.

Fig. 1

Study design using seromics. A large cohort of 1253 serum samples, including 611 HCC patients, 249 patients with liver cirrhosis (cirrhotic), and 393 healthy controls (healthy), were enrolled for discovery and evaluation of potential serum AAbs as HCC diagnostic biomarkers

AAb screening for construction of HCC-focused arrays

In the discovery phase (I), the HuProt arrays were employed to profile 100 serum samples collected from 50 HCC and 50 healthy (Additional file 1: Fig. S1). For selection of candidate proteins, three criteria should be satisfied after comparing HCC vs. Healthy, as described in the “Methods” section. Finally, 81 proteins that were more significantly bound by the autoantibodies of HCC group than by those of the healthy group were identified (Fig. 2a). A total of 100 proteins were printed to fabricate the HCC-focused arrays in combination with additional 19 proteins from cancer literature in general. Among these 19 AAbs, 17 were also present in the discovery HuProt array and satisfied 1–2 criteria. Another 2 AAbs, CENPF [16] and CDKN2A [5], were absent in the discovery HuProt array. Alternatively, more samples were enrolled in the test phase (II) and validation phase (III), which would help to accurately evaluate the distinguishing capacity of these AAbs. The biological function and expression level of these 100 proteins were also investigated based on HPA database and our previous multi-omics HCC data [27] (Additional files 2 & 3: Figs. S2 & S3). One serum sample (pooled from 10 randomly selected HCC individuals) was independently applied to a total of 47 different HCC-focused arrays to evaluate their potential variance. As shown in Fig. 2b, the variance was minimal with an average correlation coefficient of 0.95.

Fig. 2

Fabrication of HCC-focused arrays. a According to the screening results of HuProt arrays, 81 proteins (p ≤ 0.05, FC ≥ 1.2 and positive ratio ≥ 10%) were selected as potential candidates. A total of 100 proteins were printed to fabricate the HCC-focused arrays, including 19 proteins from previous reports. b Six representative HCC-focused arrays testing the same sample exhibited high reproducibility. The diagonal indicates the SNR distribution of the sample, the lower left indicates the bivariate scatter plot with a fitted line, and the upper right indicates the correlation coefficient and the significance (***p < 0.001). c HCC-focused arrays were incubated with samples from one HCC patient, one patient with liver cirrhosis, and one healthy control, respectively. Three-dimension renderings of the signal intensities were shown, indicating that the array worked well

Identification of AAb biomarkers for HCC detection

Next, HCC-focused arrays were tested using serum samples from a large cohort of HCC individually. In the test phase (II), the signals of each protein between HCC and healthy or cirrhotic were compared, respectively. Examples of array image for HCC, cirrhotic, and healthy were provided in Fig. 2c. We identified a total of 55 potential biomarkers using the following criteria: p < 0.05, FC ≥ 1.2, and sensitivity > 15% with at least 90% specificity. Among them, 24 AAbs were able to classify HCC patients versus healthy, 17 AAbs were able to classify HCC versus cirrhotic, and the remaining AAbs were able to classify both HCC patients versus healthy and HCC versus cirrhotic (Additional file 6: Table S2).

To select predictors for model development, we performed 10-fold cross validation for the 55 potential biomarkers (Fig. 3a). The differential AAbs in each fold were used as input to a logistic regression that classified HCC patients versus controls. Within each fold, stepwise variable selection identified the most discriminative subset of the biomarker candidates [28]. Biomarker candidates selected in ten folds were characterized as predictors in a consensus logistic regression model, and 7 predictors were identified including CIAPIN1, EGFR, MAS1, SLC44A3, ASAH1, UBL7, and ZNF428 (Additional file 4: Fig. S4). The performance of the combinatorial 7 AAbs was then evaluated for HCC detection.

Fig. 3

Identification of combinatorial biomarker panel and development of ANN model. a Predictors were selected using 10-fold cross validation. The subjects were systematically rotated between ten folds. Within each fold, differential AAbs were determined comparing HCC patients to controls. The predictors for further model development were generated using the potential biomarkers, which worked in ten folds in the cross validation. b The correlations between any two proteins from the 7 predictors were calculated using all samples (HCC, cirrhotic, and healthy) in the test phase (II). The diagonal indicates the SNR distribution of the sample, the lower left indicates the bivariate scatter plot with a fitted line, and the upper right indicates the correlation coefficient and the significance (*p < 0.05, **p < 0.01, ***p < 0.001). c Schematic representation of the ANN model to predict HCC. Fully connected feedforward neural-networks including 7 input nodes (7 predictors), 5 neurons in the hidden layer, and 2 output nodes were chosen. Back propagation of error algorithm was used as the learning rule, and the average committee vote was used to classify the patient samples

Performance of the 7-AAb panel in test/validation phase

The correlations between any two proteins from the 7 predictors were calculated using all samples (HCC, cirrhotic, and healthy) in the test phase (II). The results showed that the closest connection existed between MAS1 and ASAH1 with a coefficient of 0.77 (p < 0.001; Fig. 3b). It has been reported that neural network analysis was potentially more powerful than traditional statistical techniques when the interaction among variables was complex. Thus, ANN model based on these 7 predictors was further explored in the test phase (II). We built a three-layer neural network with 7 input nodes, 5 hidden neurons, and 2 output neurons (Fig. 3c). The committee vote was performed by averaging all outputs and then to classify the samples (Fig. 4). As shown in Table 1, the ANN model for this 7-AAb panel could identify HCC with a sensitivity of 68.6% and a specificity of 92.1% (AUC = 0.894, HCC vs. controls [healthy + cirrhotic]), which was superior to AFP (cutoff = 400 ng/mL, sensitivity = 28.4%, specificity = 98.7%, AUC = 0.808).

Fig. 4

Workflow for the ANN-model. For the test phase (II), 576 samples were randomly split into 10 equally sized groups. One ANN model was built using 90% of cases as training set and the remaining 10% as verification set. This procedure was performed 10 times to generate 10 ANN models. Five hundred ANN models were obtained after a total running of 50 times. Each ANN model provided the following outputs: 0 indicates healthy control and 1 indicates HCC. The committee vote was performed by averaging all outputs and then to classify the samples. The samples in the validation phase (III) used 500 ANN models for the blind test

Table 1 Performance of the 7-AAb panel and AFP in HCC detection

Second, 577 serum samples from an independent cohort were used (phase III) to validate the performance of this 7-AAb panel. Based on the ANN-model, this panel had a sensitivity of 73.4% and a specificity of 90.1% for HCC detection (HCC vs. controls [healthy + cirrhotic], AUC = 0.902) (Table 1), as well as a sensitivity of 80.6% and a specificity of 90.1% for AFP− HCC detection (AFP− HCC vs. Controls [healthy + cirrhotic], AUC = 0.926) (Table 2). Importantly, this panel detected HCC with high sensitivity (62.2–77.5%), outperforming AFP (30.7%) (Table 1).

Table 2 Evaluation of the 7-AAb panel in AFP− HCC detection

When combining the test phase (II) and validation phase (III), we found that this model also performed well. It reached to a sensitivity of 71.6% and a specificity of 90.0% in detecting HCC (HCC vs. controls [healthy + cirrhotic], AUC = 0.898) (Table 1), and a sensitivity of 76.1% and a specificity of 89.1% in detecting AFP− HCC (AFP− HCC vs. controls [healthy + cirrhotic], AUC = 0.912) (Table 2).

The 7-AAb panel’s performance for HBsAg− and HBsAg+-HCC

Chronic HBV infection is the leading cause of HCC in Eastern Asian countries and most African countries [3]. Hepatitis B surface antigen (HBsAg) is used to determine whether a patient has a recent or long-standing infection of HBV. In our cohort, approximately 70% (439/611) of HCC patients were HBsAg positive (HBsAg+-HCC). For HBsAg+-HCC detection, our model provided sensitivity of 59.6–79.1% and specificity of 85.2–95.5%, while AFP (cutoff 400 ng/mL) provided sensitivity of 31.4–34.7% and specificity of 96.7–100% (Additional file 7: Table S3). We also explored the feasibility of the model in HCC patients with negative HBsAg (HBsAg−-HCC). Test phase (II) and validation phase (III) contained 46 and 30 HBsAg−-HCC patients, respectively. We found that the ANN model of this panel was able to efficiently detect HBsAg−-HCC patients from controls (AUC 0.822–0.932), superior to AFP at a cutoff of 400 ng/mL with an AUC of 0.567–0.647 (Additional file 7: Table S3).

The 7-AAb panel’s performance for different HCC stages

Patients with early-stage HCC can benefit from curative treatments like tumor resection, liver transplantation, or ablation [18]. The performance of our model for HCC patients at different stages were also considered in our study. The evaluation for different stages of BCLC is provided in Table 3 and the others including TNM and Chinese HCC stages in Additional files 8 & 9: Tables S4 & S5. For early stage HCC (BCLC: 0, A; TNM: IA, IB; Chinese: Ia, Ib) detection, our model demonstrated significantly improved performance with 5–20% increases of AUC compared with AFP (cutoff 400 ng/mL). For HCC patients at intermediate or late stages (BCLC: B, C; TNM: II, III, IV; Chinese: II, III) detection, our model in combination with AFP (cutoff 400 ng/mL) achieved sensitivity of 72.2–88.6% and specificity of 89.3–96.6% (AUC 0.887–0.967) for distinguishing HCC patients from controls (healthy and cirrhotic). This combination achieved sensitivity of 79.2–94.3% and specificity of 90.1–97.6% (AUC 0.918–0.985) for distinguishing HCC patients from healthy, and sensitivity of 58.3–88.6% and specificity of 83.7–98.8% (AUC 0.829–0.943) for distinguishing HCC patients from cirrhotic. Thus, the 7-AAb panel based on ANN-model could be effectively applied for early-stage HCC detection.

Table 3 Performance of the 7-AAb panel and AFP to detect HCC with different BCLC stages