A novel immunodiagnosis panel for hepatocellular carcinoma based on bioinformatics and the autoantibody-antigen system.

Hepatocellular carcinoma (HCC) is a malignancy with a dismal survival rate. The novel autoantibodies panel may provide new insights for the diagnosis of HCC. Biomarkers screened by two methods (bioinformatics and the antigen-antibody system) were taken as candidate tumor-associated antigens (TAAs). Enzyme-linked immunosorbent assay was used to detect the corresponding autoantibodies in 888 samples of verification and validation cohorts. The verification cohort was used to verify the autoantibodies. Samples in the validation cohort were randomly divided into a train set and a test set with the ratio of 6:4. A diagnostic model was established by support vector machines within the train set. The test set further verified the model. Eleven TAAs were selected (AAGAB, C17orf75, CDC37L1, DUSP6, EID3, PDIA2, RGS20, PCNA, TAF7L, TBC1D13, and ZIC2). The titer of six autoantibodies (PCNA, AAGAB, CDC37L1, TAF7L, DUSP6, and ZIC2) had a significant difference in any of the pairwise comparisons among the HCC, liver cirrhosis, and normal control groups. The titer of these autoantibodies had an increasing tendency. Finally, an optimum diagnostic model was constructed with the six autoantibodies. The AUCs were 0.826 in the train set and 0.773 in the test set. The area under the curve (AUC) of this panel for diagnosing early HCC was 0.889. The diagnostic ability of the panel reduced with the progress of HCC. The positive rate of the panel in diagnosing alpha-fetoprotein (AFP)-negative patients was 75.6%. For early HCC, the sensitivity of the combination of AFP with the panel was 90.9% and superior to 53.2% of AFP alone. The novel immunodiagnosis panel combining AFP may be a new approach for the diagnosis of HCC, especially for early-HCC cases.

alpha‐fetoprotein

area under the curve

enzyme‐linked immunosorbent assay

hepatocellular carcinoma

liver cirrhosis

normal controls

signal‐to‐noise ratio

support vector machines

autoantibodies against tumor‐associated antigen

tumor‐associated antigens

weighted gene coexpression network analysis

INTRODUCTION

Liver cancer, especially primary liver cancer, is one of the leading causes of cancer‐related death worldwide, of which hepatocellular carcinoma (HCC) accounts for 80%‐90%. The morbidity and mortality of liver cancer ranked seventh and second in all malignant tumors, respectively. Clinically, only 10%‐20% of patients with HCC could be treated by surgical operation due to the difficulty of early diagnosis of HCC. , Without a specific treatment plan, the median survival time of advanced liver cancer is only 1‐2 months. , Alpha‐fetoprotein (AFP), as the only serum biomarker for HCC diagnosis in clinical practice, has been widely used. However, its sensitivity is only about 60%. , Hence, it is an urgent need to identify effective and noninvasive biomarkers for the early diagnosis of HCC.

A great deal of evidence showed that autoantibodies against tumor‐associated antigen (TAAbs) arose in the blood at an early stage of tumorigenesis. , , Compared with other serological markers, , , TAAbs persist in the blood for long periods. , TAAbs can serve as tumor‐diagnostic biomarkers also owing to their easy measurement in serum and immune amplification effect. , , , Researchers also found that the diagnostic value of a single TAAb was not particularly ideal in regard to sensitivity and specificity. , , Given the problem, scholars mainly focused on combining multiple TAAbs to diagnose tumors. , Different panels are needed for different types of cancers to enhance diagnostic power.

In recent years, the rapid development of bioinformatics provided many research methods for the screening of tumor markers. Among them, weighted gene coexpression network analysis (WGCNA) is an excellent means of extracting gene modules and correlating them with clinical traits. , WGCNA has the unique advantage of converting gene expression data into coexpression modules and providing new insights into genes that may be responsible for phenotypic characteristics. , For these reasons, WGCNA was widely used in several cancers to identify pivot biomarkers for cancer diagnosis or prognosis. , , ,

Autoantibody‐antigen system is used as a common method for disease detection such as chronic hepatitis B and plasmodium. In the same way, the autoantibody‐antigen system also can be used to screen disease markers. Among these, the utilization of a human proteomics chip to screen tumor‐related markers is a significant method. , The chip used in this study contains all human recombinant proteins that can be purified at present to comprehensively screen tumor‐associated antigens (TAAs) related to HCC.

In this study, bioinformatics and the autoantibody‐antigen system were used to screen TAAs related to HCC. After verification and validation, 753 samples in the validation phase were divided into a train set and a test set to evaluate the ability of the anti‐TAAs panel as noninvasive markers to diagnose HCC.

MATERIALS AND METHODS

Serum samples and research design

A total of 888 serum samples were taken into consideration, including 296 HCC cases, 296 liver cirrhosis (LC) cases, and 296 normal controls (NCs). The subjects among the three groups were age (±2 years) and sex matched according to 1:1 ratio. These samples were acquired from the Tumor Epidemiology Laboratory of Zhengzhou University. All subjects signed informed consent forms, and the study was approved by the Ethics Review Committee of Zhengzhou University (ZZURIB 2019‐001). The verification phase including HCC (n = 45), LC (n = 45), and NC (n = 45) patients was used to validate the selected TAA. The remaining samples included in the evaluation phase were randomly divided into a train set and a test set according to the ratio of 6:4. The train set was used to build a diagnostic model, and the test set further verified the model (Figure 1).

FIGURE 1

Work flow chart of the research. HCC, hepatocellular carcinoma; LC, liver cirrhosis; NC, normal control; TAAs, tumor‐associated antigens; TAAbs, autoantibodies against tumor‐associated antigens; SVM, support vector machines

Data download and preprocessing

Seven microarray datasets, including GSE84005, GSE76297, GSE64041, GSE54236, GSE57957, GSE39791, GSE77314, were downloaded from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/) database (Table S1). The expressions of the probes in the chip were normalized by the “normalizeBetweenArrays” function in the “limma” package. Each dataset was annotated with its corresponding annotation platform. When a gene had multiple probes, the maximum value was used as the expression value of the gene. GSM1979485, GSM1979420, and GSM1979407 were removed because they did not match the adjacent tissues in GSE76297. GSM1310594 in GSE54236 was deleted for the same reason. The “Combat” function in the “sva” package was used to perform batch calibration among microarray datasets. The corrected dataset was analyzed.

Screening of candidate TAAs

One of the methods used in this study to screen HCC‐related markers was WGCNA. The basic theory of the WGCNA algorithm has been presented in a previous study. The “WGCNA” package was used to calculate the gene coexpression network. The “step‐by‐step” method was used to build the WGCNA network. Cluster analysis was used to remove outliers by the average linkage method. The scale‐free topology fit index served as a function of the soft‐thresholding power to choose an optimum threshold. The threshold was used to construct adjacency matrix and coexpression similarity. Based on dissimilarity topological overlap, genes were hierarchically clustered to construct geneTree. Dynamic hybrid branch cutting method was used to identified gene modules. The cutoff values of minimum size (minimum cluster size) and sensitivity to cluster splitting were 30 and 2, respectively. The threshold was set at 0.25 to merge modules with similarities. WGCNA has an ability to calculate the correlation between gene modules and clinical phenotypes. The correlation was tested by the Pearson test. Finally, the most relevant module was selected.

Another method for screening candidate TAAs was the human proteomics chip. The human proteomics chip used in the study is currently the highest throughput protein chip. It contains more than 21 000 recombinant proteins (https://cdi.bio/huprot/). Hence, the human proteomics chip is an effective method for comprehensive screening of TAAs. It was applied to detect the TAA concentration in the sera of 10 HCC cases and 10 NCs. The experiment was completed under the company standard procedure (BCBIO Biotechnology). The detected data were normalized based on signal‐to‐noise ratio (SNR) value. When the SNR of a protein was greater than 4, it was defined as positive protein. On this basis, we set an appropriate cutoff threshold for each protein based on the SNR of the protein in the control group and calculated the positive rate of the protein in the cancer group and the healthy control group. For each protein, we assumed the two sets of samples were from two identical populations. Then, a one‐tailed test was performed by Mann‐Whitney’s U, represented by P‐value. The following criteria were used to screen meaningful TAAs in the human proteomics chip: fold change (FC) > 1 and P < .05; positive rate in HCC cases ≥ 50%; and negative rate in the NC group ≥ 90%.

The intersection of the biomarkers screened by WGCNA and the human proteomics chip was treated as candidate TAAs.

Indirect enzyme‐linked immunosorbent assay (ELISA)

Indirect ELISA was used to detect the concentration of TAAbs in serum. The specific experimental methods and procedures have been introduced in detail in previous studies. , The 11 recombinant proteins (AAGAB, C17orf75, CDC37L1, DUSP6, EID3, PDIA2, RGS20, PCNA, TAF7L, TBC1D13, and ZIC2) were all diluted in the coating buffer at a concentration of 0.125 μg/mL. HCC, LC, and NC samples were evenly distributed on each plate. Every plate had five replicate sera and three blanks. Different plates were normalized by repeated sera.

Diagnostic model construction based on SVM

Support vector machines (SVM) are a set of machine learning methods. Compared with other machine learning methods, SVM is very powerful in discerning nuances in complex data sets, which has been used as a classifier for cancer classification. To build the most suitable SVM learning diagnosis model, the kernel function was selected and the "radial" and gamma was set to 0.1. The model was used to distinguish HCC from NC and applied to the diagnosis of HCC subgroups such as early stage (Barcelona Clinic Liver Cancer [BCLC], stage 0‐B) and late stage (BCLC, stage C‐D). Electrochemiluminescence immunoassay was used to detect the concentration of AFP. The positive rates of three groups (TAAb panel, AFP, and combination of TAAb panel and AFP) in recognizing HCC were compared. And the positive rates of these three groups were also compared in early and late HCC. According to clinical standards, the threshold of AFP was set to 7 ng/mL. HCC patients were divided into an AFP‐positive group (AFP[+]) and an AFP‐negative group (AFP[−]). The panel was used to diagnose AFP(+) patients and AFP(−) patients. The model's ability to recognize LC was also evaluated. The diagnostic ability of the models was assessed by receiver operating characteristic curves (ROC) analysis and the area under the ROC curve (AUC).

Statistical analysis

The processing of microarray datasets and the realization of WGCNA were carried out in the R 4.0.4 language. Sample matching and analysis of differences among the three groups were implemented by the IBM SPSS 26.0 software. Before analysis, data were assessed for normal distribution by the Kolmogrov‐Simirnov normality test. If the data conformed to the normal distribution, they were analyzed by the ANOVA; otherwise, they were tested by the Kruskal‐Wallis test. The statistical difference between the two AUCs were evaluated by the MedCalc 11.0 software. The P‐values calculated were two tailed, and significant differences were set at 95% level (P < .05).

RESULTS

Demographic characteristics

This experiment was mainly divided into three phases (Figure 1). In the discovery phase, WGCNA and human proteomic chips were used to screen candidate TAAs. In the verification and validation phase, ELISA was used to detect the titers of 11 corresponding TAAbs in the serum samples, including 296 HCC cases, 296 LC cases, and 296 NCs. The basic information of the participants is shown in detail in Table 1. Statistical analysis showed that there was no difference in the distribution of age and gender among the three groups (P > .05).

TABLE 1

Characteristics of participants

Characteristics	Verification dataset			Validation dataset
	HCC (n = 45)	LC (n = 45)	NC (n = 45)	HCC (n = 251)	LC (n = 251)	NC (n = 251)
Male n (%)	33 (73.3)	33 (73.3)	33 (73.3)	204 (80.3)	204 (80.3)	204 (80.3)
Age
Mean ± SD	53.8 ± 9.9	53.5 ± 9.6	53.4 ± 9.8	52.6 ± 10.5	52.4 ± 1.05	52.1 ± 10.3
BCLC, n (%)
0	0 (0)	IA	IA	4 (1.6)	IA	IA
A	4 (8.8)	IA	IA	40 (15.9)	IA	IA
B	8 (17.8)	IA	IA	35 (13.9)	IA	IA
C	9 (20.0)	IA	IA	55 (21.9)	IA	IA
D	0 (0)	IA	IA	8 (3.2)	IA	IA
NA	24 (53.4)	IA	IA	109 (43.7)	IA	IA
AFP, n (%)
>7 ng/mL	22	NA	NA	131 (52.2)	NA	NA
<7 ng/mL	18	NA	NA	86 (34.3)	NA	NA
NA	5	NA	NA	34 (13.5)	NA	NA

Abbreviations: AFP, alpha‐fetoprotein; BCLC, Barcelona Clinic Liver Cancer; HCC, hepatocellular carcinoma; IA, inapplicable; LC, liver cirrhosis; NA, not available; NC, normal control; SD, standard deviation.

Weighted coexpression network construction and key module identification

A total of 794 samples were included to construct a gene coexpression network. Sample clustering demonstrated that three outliers (GSM979132, GSM979123, and GSM1398656) need to be unlocked (Figure 2A). Then, the remaining 791 samples were used for WGCNA. The thresholding of β = 4 was selected to determine the scale‐free scale of the network (Figure 2B). Through the thresholding power, genes were divided into 17 modules in total (Figure 2C). The turquoise module was obtained because it had the highest correlation with the sample category (r = .73, P < .0001, Figure 2D). It also showed a high correlation with clinical information (cor = 0.9, P < .001 Figure 2E). Therefore, the genes in the turquoise module were chosen for subsequent analyses.

FIGURE 2

Identification of modules related to the clinical features of hepatocellular carcinoma. A, Identification and deletion of abnormal samples. B, Recognition of soft threshold in weighted gene coexpression network analysis (WGCNA). C, Gene clustering dendrogram based on dissimilarity measure (1‐TOM). D, Heat map of correlation between the gene clustering module and clinical characteristics of hepatocellular carcinoma. E, Module membership vs gene significance in the turquoise module

Tumor‐associated antigen selected in the discovery phase

The human proteomics chip data were analyzed according to the standards established in the method. Finally, 11 TAAs were obtained by the two methods, including AAGAB, C17orf75, CDC37L1, DUSP6, EID3, PDIA2, RGS20, PCNA, TAF7L, TBC1D13, and ZIC2. The basic characteristics of the 11 TAAs are shown in Table S2.

Detection of autoantibodies in the verification and validation phase

To determine whether the 11 TAAbs had a potential diagnostic value, their titers were measured in two phases. In the verification phase, ELISA was used to detect the titer of the 11 TAAbs in 135 samples. The scatter plot of the optical density (OD) values of the 11 TAAbs is shown in Figure 3A. The titer of autoantibodies against CDC37L1, ZIC2, DUSP6, PCNA, EID3, TAF7L, and AAGAB had differences among the three groups. Then, the seven TAAbs were examined in the validation cohort. Results showed that the average level of the OD values of six autoantibodies (PCNA, AAGAB, CDC37L1, TAF7L, DUSP6, and ZIC2) had a significant difference in any of the pairwise comparisons between the three groups (Figure 3B). Furthermore, the OD value of the six autoantibodies was the highest in the HCC group and the lowest in the NCs group. Finally, six autoantibodies (PCNA, AAGAB, CDC37L1, TAF7L, DUSP6, and ZIC2) were chosen for further analysis.

FIGURE 3

Scatter plot of the titers of 11 autoantibodies against tumor‐associated antigen in verification phase (A) and validation phase (B). The superscript letter “a” indicates that the OD value of this autoantibody was different among the three groups. OD, optical density

Establishment of a diagnostic model for distinguishing HCC from NC

Receiver operating characteristic curves analysis was carried out to appraise the diagnostic value of each anti‐TAA autoantibody in the validation cohort. The AUCs of the six TAAbs ranged from 0.671 to 0.741 (Figure 4A‐F).

FIGURE 4

The diagnostic capabilities of six autoantibodies in the validation phase. CI, confidence interval; AUC, area under the curve

The six TAAbs were used as a panel to construct a diagnostic model by the train set. The AUC of the model was 0.826 (95% confidence interval [CI]: 0.779‐0.873). Sensitivity, specificity, and accuracy rates were 77.5%, 76.2%, and 77.0%, respectively (Figure 5A, Table 2). The test set was used to verify the model, and a slightly lower AUC (Figure 5B, Table 2) was acquired. There was no statistical difference in AUC between the train set and the test set (z = 1.299, P = .1940).

FIGURE 5

Receiver operating characteristic (ROC) curves for the panel by (A) tian, (B) test, (C) early HCC and normal control, (D) late HCC and normal control, (E) AFP(−) HCC and normal control, (F) AFP(+) HCC and normal control, (G) HCC and liver cirrhosis, (H) early HCC and liver cirrhosis, (I) liver cirrhosis and normal control. AFP, alpha‐fetoprotein; AFP(−), AFP‐negative group; AFP(+), AFP‐positive group; AUC, area under the curve; BCLC, Barcelona Clinic Liver Cancer; CI, confidence interval; early HCC, BCLC, stage 0‐B; HCC, hepatocellular carcinoma; late HCC, BCLC stage C‐D; LC, liver cirrhosis; NC, normal control

TABLE 2

Evaluation of the diagnostic value of the SVM model in different datasets

	AUC (95% CI)	Se (%)	Sp (%)	Youden's index	PLR	NLR	PPV (%)	NPV (%)	Accuracy (%)
Train	0.826 (0.779‐0.873)	78.0	76.0	0.536	3.250	0.289	76.5	77.6	77.0
Test	0.773 (0.708‐0.837)	73.0	75.0	0.480	2.920	0.360	74.5	73.5	74.0
Early HCC vs NC	0.889 (0.836‐0.943)	81.0	87.0	0.684	6.231	0.218	86.2	82.1	84.0
Later HCC vs NC	0.811 (0.736‐0.886)	78.0	73.0	0.508	2.889	0.301	74.3	76.8	75.5
AFP(+) vs NC	0.822 (0.772‐0.872)	65.6	88.5	0.542	5.739	0.384	85.2	72.2	77.3
AFP(−) vs NC	0.781 (0.710‐0.851)	75.6	76.7	0.523	3.304	0.312	76.8	76.2	76.5
HCC vs LC	0.652 (0.604‐0.699)	58.2	64.5	0.227	1.639	0.648	62.1	60.7	61.4
early HCC vs LC	0.710 (0.628‐0.788)	56.0	79.0	0.342	2.667	0.557	72.7	64.2	67.5
LC vs NC	0.656 (0.608‐0.705)	55.8	74.9	0.307	2.223	0.590	69.0	62.9	65.4

Abbreviations: AFP(−), AFP negative group; AFP(+), AFP positive group; AUC, area under the receiver operating characteristic curve; CI, confidence interval; early HCC, Barcelona Clinic Liver Cancer 0‐B; HCC, hepatocellular carcinoma; later HCC: Barcelona Clinic Liver Cancer C‐D; NC, normal control; NLR, negative likelihood ratio; NPV, negative predictive value; PLR, positive likelihood ratio; PPV, positive predictive value; Se, sensitivity; Sp, specificity; SVM, support vector machines.

The model produced by the train set was also used to diagnose the early stage (BCLC, stage 0‐B) and late stage (BCLC, stage C‐D) of HCC. The AUC of this panel to diagnose early HCC was 0.889 (Figure 5C and Table 2), and that to diagnose late stage was 0.811 (Figure 5D and Table 2). The difference between the two AUCs was not statistically significant (z = 1.673, P = .094).

Combination of AFP and autoantibodies

There were 217 HCC patients with AFP information. The positive rate of AFP was only 59.9% (130/217) (Table 3), while the positive rate of the panel was 77.9% (169/217). When the AFP and the panel were connected in parallel, the positive rate reached 90.3% (196/217). There were statistical differences in positive rates among them (χ 2 = 55.69, P < .001, Table 3).

TABLE 3

Comparison of single and parallel detection of AFP and model

	Panel	AFP	Panel + AFP	χ2/P
Number of positive	169a	130b	196c	55.69/<.001
Number of negative	48	87	21

Each subscript letter denotes a subset of work categories whose column proportions do not differ significantly from each other at the 0.05 level.

Abbreviation: AFP, alpha‐fetoprotein.

According to the clinical cutoff value (7 ng/mL), HCC patients were divided into an AFP(+) group and an AFP(−) group. The model was used to diagnose AFP(+) patients and AFP(−) patients (Figure 5E,F and Table 2). In the AFP(+) group, the AUC value was 0.822, which was slightly higher than that of the AFP(−) group. However, there was no statistical difference between the two groups (z = 0.923, P = .356).

In early‐HCC patients, the sensitivity of AFP was only 53.2% (41/77). Interestingly, 80.5% (29/36) of AFP(−) HCC sera were anti‐TAAs positive. If the panel and AFP were combined to diagnose HCC, 90.9% (70/77) of early‐HCC patients were correctly diagnosed (Table 4). The change in sensitivity was statistically significant (χ 2 = 27.03, P < .001).

TABLE 4

Sensitivity of the combination of AFP and the panel to detect early and late HCC

Anti‐TAAs panel	Serum AFP levels				Total
	Early HCC		Late HCC
	>7 ng/mL	<7 ng/mL	>7 ng/mL	<7 ng/mL
Positive	35	29	40	7	111
Negative	6	7	10	0	23
Total	41	36	50	7	134

Sensitivity (%) of early HCC: (35 + 29 + 6)/(35 + 29 + 6 + 7) = 90.9%; Sensitivity (%) of late HCC: (40 + 7 + 10)/(40 + 7 + 10 + 0) = 100%.

Abbreviations: AFP, alpha‐fetoprotein; HCC, hepatocellular carcinoma; TAAs, tumor‐associated antigens.

In late‐HCC patients, the AFP‐positive rate was 87.7% (50/57). All AFP‐negative patients were diagnosed as HCC cases by the panel. When AFP was used in combination with the panel, all late patients were correctly identified (Table 4). The change in positive rate was statistically significant (χ 2 = 5.14, P = .023).

Identification of LC using the optimal model

To explore the progression pattern of HCC, the diagnostic ability of the model in the course of the disease was analyzed. First, the model's ability to differentiate HCC from LC was examined. The results showed that the AUC was 0.652 (95% CI: 0.604‐0.699). Sensitivity, specificity, and accuracy rates were 58.2%, 64.7%, and 61.4%, respectively (Figure 5G, Table 2). The model's ability to discriminate early HCC from LC was also examined, and a slightly higher AUC was acquired (Figure 5H, Table 3). The AUC of differentiating LC from NC was 0.656 (95% CI: 0.608‐0.705). Sensitivity, specificity, and accuracy rates were 55.8%, 74.9%, and 65.4%, respectively (Figure 5G, Table 3).

DISCUSSION

At present, HCC is still one of the lethal cancers threatening human health in the world. Diagnosis of HCC has always been a research focus. This article mainly used bioinformatics and the autoantibody‐antigen system to screen TAAs related to HCC. Then, ELISA was used to detect the titer of the corresponding autoantibodies in the sera, and SVM was used to build a diagnostic model. In the end, we successfully constructed a panel with a good diagnostic performance for detecting HCC.

In this study, bioinformatics and the autoantibody‐antigen system were combined to screen the candidate TAAs. Among them, WGCNA is an excellent algorithm in screening the biological markers, and it has been used in many studies. The human proteomics chip contains 21 000 human recombinant proteins in this study, which can comprehensively screen TAAs. , , , In summary, bioinformatics can screen for genes that play important roles in the occurrence and development of HCC. Hence, those two methods were used together to screen valuable TAAs as much as possible.

Next, the concentration of autoantibodies against selected TAAs was validated using ELISA. Through the verification and the validation phase, six TAAbs (PCNA, AAGAB, CDC37L1, TAF7L, DUSP6, and ZIC2) had a significant difference in any of the pairwise comparisons between the three groups. Interestingly, we also found the titer of the six autoantibodies had an increasing tendency with the disease stage, which may provide early warning of the onset of HCC.

In recent years, many studies have shown that exploration of a panel is helpful for the diagnosis of cancer. Furthermore, machine learning can provide promising tools; it has been applied in pancreatic cancer and prostate cancer. According to the research of Zhang et al, the diagnostic ability of a single antigen‐antibody system was limited, but combining multiple TAAbs can improve the diagnostic performance for cancer detection. Hence, the six TAAbs were used to construct a diagnostic model using SVM. The panel showed good diagnostic performance for detecting HCC. In addition, an interesting phenomenon that caught our attention was that the panel was more valuable in diagnosing early HCC than late HCC. There are similar results in other cancers. , This decrease may be the result of the loss of antigens to help the tumor escape immune surveillance. In other words, the panel constructed by six autoantibodies may be used for the diagnosis of early HCC. Compared with previous studies, , , our research focuses on systematically and comprehensively screening HCC‐related autoantibodies and verifying them in phases. Moreover, we constructed a robust HCC diagnostic panel with large samples, which had not been achieved before either. Review of previous studies showed that our research was consistent with the findings of related studies. For example, Wang et al developed a 22‐autoantibody detector to diagnose prostate cancer, with 81.6% sensitivity and 88.2% specificity. These strongly demonstrated that a customized autoantibodies panel is a promising and powerful immunodiagnostic tool for certain types of cancer, such as HCC and prostate cancer.

This study also analyzed the association between AFP and TAAbs in HCC diagnosis. The result showed that there was no statistical difference between the AUCs of the two groups (AFP[+] group and AFP[−] group). It can be concluded that TAAbs had no relationship with AFP. Namely, TAAbs are independent diagnostic markers that can be used to diagnose AFP(−) HCC patients. Some research studies revealed the AFP value was in the normal range in approximately 40% of diagnosed HCC cases. , The positive rate of AFP in HCC patients is consistent with several studies (Table 3). The positive rate of our model was 78%, which is significantly higher than that of AFP (Table 3). It is gratifying that the positive rate reached 90.3% when the two methods were connected in parallel. In order to further explore the supplementary ability of this panel for AFP to diagnose HCC, we divided patients with AFP information into an early‐HCC and a late‐HCC group. To our surprise, the positive rate of "panel+AFP" for diagnosis of early HCC was significantly improved, with a sensitivity of 90.9%, while AFP was only 53%. Those results were consistent with the conclusions of previous researches. , The above research results indicated the immunodiagnostic panel can be used as a marker for the diagnosis of AFP(−) HCC patients, and the panel combined with AFP may improve the sensitivity of diagnosing HCC, especially for early‐HCC cases.

The current research has several advantages. First, the method was used to screen HCC‐related TAAs should be mentioned. The method of combining bioinformatics and autoantibody‐antigen system could be more accurate for screening out TAAs that may play an important role in the development of HCC. Second, multistage verification in large samples was applied, which makes the results more credible. Third, compared with AFP in parallel, the sensitivity of the panel in diagnosing HCC was higher, and the results indicated that the novel immunodiagnosis panel combined with AFP may be a new approach for the diagnosis of HCC, especially for early‐HCC cases. However, there are some limitations to our research. First, the human proteomics chip does not include proteins with post‐translational modifications and structural changes. Therefore, those related TAAbs cannot be detected. However, the purpose of combining with bioinformatics was to screen out important indicators as accurately as possible. Second, the function of the screened indicators in the occurrence and development of HCC needs further verification.

In summary, our research data showed that the customized autoantibodies panel provides a supplementary method for the diagnosis of HCC, especially for early‐HCC and AFP(−) patients. The study also proves that autoantibodies are a promising direction for cancer biomarker research. The ability of this method to diagnose cancer depends on the careful choice of the autoantibodies panel. Our findings may encourage future clinical studies to explore the function of related proteins in the occurrence and progression of HCC.

DISCLOSURE

The authors declare no conflict of interest.

Tab S1‐S2

Click here for additional data file.