Proteome Multimarker Panel for the Early Detection of Hepatocellular Carcinoma: Multicenter Derivation, Validation, and Comparison.

Conventional methods for the surveillance of hepatocellular carcinoma (HCC) by imaging, with and without serum tumor markers, are suboptimal with regard to accuracy. We aimed to develop and validate a reliable serum biomarker panel for the early detection of HCC using a proteomic technique. This multicenter case-control study comprised 727 patients with HCC and patients with risk factors but no HCC. We developed a multiple reaction monitoring-mass spectrometry (MRM-MS) multimarker panel using 17 proteins from the sera of 398 patients. Area under the receiver operating characteristics curve (AUROC) values of this MRM-MS panel with and without α-fetoprotein (AFP) and protein induced by vitamin K absence or antagonist-II (PIVKA-II) were compared. The combination and standalone MRM-MS panels had higher AUROC values than AFP in the training (0.940 and 0.929 vs 0.775, both P < 0.05), test (0.894 and 0.893 vs 0.593, both P < 0.05), and confirmation sets (0.961 and 0.937 vs 0.806, both P < 0.05) in detecting small single HCC. The combination and standalone MRM-MS panels had significantly higher AUROC values than the GALAD score (0.945 and 0.931 vs 0.829, both P < 0.05). Our proteome 17-protein multimarker panel distinguished HCC patients from high-risk controls and had high accuracy in the early detection of HCC.

Introduction

Hepatocellular carcinoma (HCC) accounts for 75–85% of all primary liver cancer cases, and its mortality rate is nearly equivalent to its global incidence.[1] The burden of HCC is highest in East Asia, where cirrhosis, hepatitis B and C virus, and excessive alcohol consumption are the major risk factors.[2]

Although controversies exist over whether HCC surveillance programs provide a survival benefit for at-risk patients, a randomized controlled trial and several retrospective studies have concluded that surveillance through imaging studies, such as ultrasonography and serum α-fetoprotein (AFP) testing, results in early detection of HCC and reduced mortality.[3−7] However, in a meta-analysis of patients with cirrhosis, abdominal ultrasonography, a universally recommended imaging modality for the surveillance of HCC, had a low sensitivity of 63% in detecting early-stage HCC with concomitant use of AFP.[8] Considering the operator-dependent nature of ultrasonography and the high costs and potential physical harms of other imaging modalities (e.g., radiation hazards with computed tomography), there is a need for the development of a highly sensitive yet economical and safe measure for the early detection of HCC.[9]

The molecular heterogeneity of HCC has prompted attempts to integrate various serum biomarkers to detect the disease at an early stage.[10] The GALAD score, which comprises age, sex, AFP, lectin-bound AFP, and protein induced by vitamin K absence or antagonist-II (PIVKA-II), is one of the most widely examined biomarker panels.[11] Similarly, multiple reaction monitoring–mass spectrometry (MRM-MS), a targeted proteomic approach, quickly and accurately analyzes hundreds of proteins as potential biomarkers in various types of cancer.[12−16] Our group has proposed a multimarker panel for the detection of early-stage HCC, consisting of 28 proteins that have been identified by MRM-MS.[17]

This multicenter study aimed to develop and evaluate a new multimarker panel for the detection of HCC with target proteins using a previous proteomic dataset, examine whether the incorporation of AFP and PIVKA-II into the panel improves its accuracy, validate the panel in an independent cohort, and compare its performance with that of the GALAD score.

Methods

Study Design and Participants

This case–control study was based on two independently established datasets of HCC patients and high-risk controls who selected per the same criteria from three referral centers [Seoul National University Hospital (SNUH), Asan Medical Center (AMC), and Samsung Medical Center (SMC)] in Seoul, Korea. The derivation set included 568 patients from all three participating centers—284 HCC patients and 284 non-HCC controls—whose serum samples were drawn between January 3, 2011, and September 3, 2013. They were randomly divided into training set (199 HCC patients and controls each) and test set (85 HCC patients and controls each). The panel was developed and validated using these training and test sets, respectively. Serum samples were also drawn from 159 patients—109 HCC patients from SNUH and AMC and 50 non-HCC controls from AMC—between September 7, 2013, and August 3, 2020, to establish another independent confirmation set (shown in Figure ).

Figure 1

Overview of the development of the MRM-MS panel. The study was performed with two independent cohorts: derivation and confirmation. A total of 383 target candidates were trained and internally validated with the derivation cohort, and the final 17-marker panel was validated with the confirmation cohort. Subgroups—tumor size under 2 cm, LC, and CHB—were also tested with regard to the performance of the model.

Patients with HCC were eligible if they were aged 20–80 years and diagnosed with single HCC that was smaller than 5 cm or 2–3 HCCs, each smaller than 3 cm, within 3 months prior to collection of their sera. The diagnosis of HCC was made according to radiographic or histological findings following the updated 2011 American Association for the Study of Liver Disease (AASLD) guidelines.[18] Patients with radiological vascular invasion or extrahepatic metastasis of HCC by computed tomography (CT) or magnetic resonance (MRI) imaging, which detects advanced HCC, were excluded. The high-risk controls comprised patients aged 20–80 years with cirrhosis or chronic hepatitis B (CHB) or C (CHC) who were confirmed not to have had HCC by abdominal ultrasonography, computed tomography, or magnetic resonance imaging within 3 months prior to collection of their samples. Common exclusion criteria were (i) impaired hepatic function, defined as Child-Pugh class C; (ii) poor performance status, considered an Eastern Cooperative Oncology Group performance status score of 3 or above; and (iii) any other malignancy within 3 years prior to the collection of sera.

Cirrhosis was diagnosed clinically per the following criteria: (i) histological findings, (ii) thrombocytopenia (<100,000/mm3) and a blunted, nodular liver edge with splenomegaly (>12 cm), or (iii) the presence of ascites, varices, or hepatic encephalopathy. CHB and CHC were defined as a positive serum hepatitis B surface antigen test and a positive serum anti-hepatitis C virus antibody test, respectively, on two separate occasions that were separated by at least 6 months.

Biomarker Candidates

In a previous study, 383 peptides, corresponding to 176 proteins, had been identified as being suitable for quantitative MRM-MS analysis. Starting with 2189 proteins that were identified in five biobank resources, semiquantitative MRM-MS assay was performed on pooled serum samples of HCC patients and controls with cirrhosis or CHB, in which 23,184 peptides, representative of 1693 proteins, were filtered by a proteomic approach using prediction servers and a database. Among 1583 reproducibly detectable peptides, 542 were differentially expressed in individual samples of HCC patients compared with controls, 421 of which were verified as being acceptable for subsequent screening according to preexisting mass spectrometry spectral data. A subsequent quantitative MRM-MS analysis targeted 385 peptides that had undergone an interference screen using a stable isotope-labeled standard peptide mixture.[17]

In this study, we performed MRM-MS quantification for 383 target peptides, correlating to 176 proteins, except for two peptides for which compatible stable isotope-labeled standard peptides were not obtained due to unavailability. Their exclusion did not affect the performance of the developed panel because the two peptides were part of proteins that were represented by other peptides in the MRM-MS quantification.

Quantitative MRM-MS Analysis

Sample block randomization was performed for each set. The six most abundant proteins were removed from the samples by a Multiple Affinity Removal System Human-6 (MARS Hu-6 × 100 mm; Agilent Technologies, Santa Clara, CA) column that was loaded onto a high-performance liquid chromatography system (Shimadzu Co., Kyoto, Japan), wherein the column was exchanged for every 200 samples that were depleted. A total of 100 μg of proteins from each sample was hydrolyzed with sequencing-grade modified trypsin (Promega, Madison, WI). The MRM-MS assays on the training and test sets were conducted on an Agilent 6490 triple quadrupole mass spectrometer (Agilent Technologies) that was equipped with a Jet Stream Electrospray source that was coupled to a 1260 Infinity HPLC system (Agilent Technologies), and the assay on the confirmation set was performed on an Agilent 6495 triple quadrupole mass spectrometer (Agilent Technologies) that was coupled to the same HPLC system.

Detailed information on the MRM-MS procedure is provided in Supplementary Methods.

Data Preprocessing

Log transformation was applied selectively to the raw data from the 383-multiplex MRM assay such that the average of the skewness values of the training set was at a minimum, to secure the normality of the distribution (Figure S1).[19] We performed Student’s t test and calculated area under the receiver operating characteristics (AUROC) value for the transformed data of each peptide, leaving 107 peptides (79 proteins) with P < 0.05 by Student’s t test and AUROC > 0.6. Based on the P values and AUROC values, a single peptide was selected for each protein that was associated with multiple peptides. Consequently, 79 candidate proteins were included in the development of the panel.

Panel Development and Assessment

Protein expression levels were compared between HCC patients and controls in the training set for the 79 biomarker candidates by the Mann–Whitney U test. We used logistic regression analysis to build the MRM-MS multimarker panel, with recursive feature elimination with cross-validation to determine the markers that were to remain in the panel. AFP and PIVKA-II were later combined with the MRM-MS panel as continuous variables. The discriminatory abilities of AFP and the MRM-MS panel with and without AFP and/or PIVKA-II were evaluated by AUROC analysis. By DeLong test, we compared, in pairs, the AUROC values of the standalone MRM-MS panel, AFP, and the combination MRM-MS panels in the training, test, and confirmation sets, with that of the GALAD score in a subset (GALAD cohort) of the confirmation cohort.[20] Youden index was used as the optimal cutoff from the training set and applied to the test and confirmation set at which the sensitivity and specificity were calculated for the standalone MRM-MS and combination MRM-MS panels.[21] The sensitivities and specificities were compared in each set with those of AFP at a cutoff of 20 ng/mL, a common value for screening HCC.[22] All reported P values were two-sided, and P values less than 0.05 were deemed to be significant.

All statistical analyses were conducted in R (version 4.0.4; R development Core Team, Vienna, Austria; http://www.R-project.org) and SPSS (version 25.0; SPSS, Chicago, IL).

Results and Discussion

Characteristics of the Study Population

Most participants had cirrhosis (62.3, 65.9, and 82.6% for HCC patients and 78.9, 75.3, and 100.0% for controls in the training, test, and confirmation sets, respectively) and Child-Pugh class A liver function (92.4, 89.4, and 96.3% and 87.9, 89.4, and 70.0%, respectively). All HCC patients with noncirrhotic liver, except 1 in the confirmation set, had CHB or CHC. No participant had a coinfection with hepatitis B and C viruses. Most HCC patients had a single tumor (90.5, 89.4, and 82.6% in the training, test, and confirmation sets, respectively) (shown in Table ).

Table 1

Characteristics of the Study Populationa,b

	training set			test set			confirmation set
	HCC (n = 199)	control (n = 199)	P value	HCC (n = 85)	control (n = 85)	P value	HCC (n = 109)	control (n = 50)	P value
age, years	58.0 (52.0–64.0)	56.0 (51.5–61.0)	0.010	58.0 (50.0–63.0)	56.0 (52.0–64.0)	0.332	61.0 (56.0–67.0)	57.0 (51.0–61.0)	0.001
sex, N (%)			<0.001			0.007			<0.001
female	34 (17.1%)	80 (40.2%)		23 (27.1%)	41 (48.2%)		17 (15.6%)	26 (52.0%)
male	165 (82.9%)	119 (59.8%)		62 (72.9%)	44 (51.8%)		92 (84.4%)	24 (48.0%)
body mass index, kg/m2	24.7 (22.5–26.7)	24.4 (22.3–26.5)	0.504	24.8 (22.6–26.3)	24.0 (22.4–25.8)	0.161	24.5 (22.5–26.6)	24.4 (23.0–27.9)	0.765
platelet, ×103/μL	142.0 (112.5–175.0)	87.0 (65.0–127.0)	<0.001	140.0 (110.0–180.0)	92.0 (64.0–147.0)	<0.001	140.0 (102.0–186.0)	74.0 (56.0–102.0)	<0.001
creatinine, mg/dL	0.9 (0.8–1.0)	0.8 (0.6–0.9)	<0.001	0.8 (0.8–1.0)	0.8 (0.6–0.9)	0.002	0.9 (0.7–1.0)	0.7 (0.6–0.9)	<0.001
albumin, g/dL	3.8 (3.5–4.0)	4.2 (3.8–4.5)	<0.001	3.8 (3.4–4.0)	4.4 (4.0–4.6)	<0.001	3.8 (3.6–4.1)	3.9 (3.7–4.3)	0.829
ALT, IU/L	41.0 (26.0–114.5)	23.0 (17.0–33.0)	<0.001	51.0 (28.0–121.0)	26.0 (18.0–35.0)	<0.001	26.0 (19.0–36.0)	23.0 (19.0–30.0)	0.236
total bilirubin, mg/dL	0.9 (0.5–1.2)	1.1 (0.7–1.6)	<0.001	0.9 (0.6–1.3)	1.0 (0.8–1.4)	0.115	0.7 (0.5–0.9)	1.6 (1.0–2.3)	<0.001
prothrombin time, INR	1.1 (1.0–1.2)	1.1 (1.0–1.2)	<0.001	1.2 (1.1–1.3)	1.1 (1.0–1.2)	<0.001	1.1 (1.0–1.1)	1.1 (1.1–1.3)	<0.001
risk factors			<0.001			0.346			0.004
cirrhosis, N (%)	124 (62.3%)	157 (78.9%)		56 (65.9%)	64 (75.3%)		90 (82.6%)	50 (100.0%)
noncirrhotic CHB, N (%)	65 (32.7%)	34 (17.1%)		20 (23.5%)	16 (18.8%)		15 (13.8%)	0
noncirrhotic CHC, N (%)	10 (5.0%)	8 (4.0%)		9 (10.6%)	5 (5.9%)		3 (2.8%)	0
others, N (%)	0			0			1 (0.9%)
Child-Pugh class, N (%)			0.183			1.000			<0.001
A	183 (92.4%)	175 (87.9%)		76 (89.4%)	76 (89.4%)		105 (96.3%)	35 (70.0%)
B	15 (7.6%)	24 (12.1%)		9 (10.6%)	9 (10.6%)		4 (3.7%)	15 (30.0%)
alcoholic, N (%)	20 (10.1%)	12 (6.0%)	0.197	6 (7.1%)	8 (9.4%)	0.780	14 (12.8%)	6 (12.0%)	1.000
diabetes mellitus, N (%)	45 (22.6%)	19 (9.5%)	0.001	11 (12.9%)	5 (5.9%)	0.189	30 (27.5%)	5 (10.0%)	0.023
hypertension, N (%)	60 (30.2%)	13 (6.5%)	<0.001	17 (20.0%)	7 (8.2%)	0.047	34 (31.2%)	13 (26.0%)	0.632
AFP ng/mL, median (IQR)	6.0 (2.9–14.1)	2.5 (1.7–3.8)	<0.001	5.7 (2.7–19.3)	2.8 (2.0–3.9)	<0.001	10.9 (3.5–43.9)	3.0 (2.3–5.6)	<0.001
PIVKA-II mAU/mL, median (IQR)	32.0 (23.0–83.5)	16.0 (13.0–20.0)	<0.001	37.0 (23.0–78.0)	16.0 (13.0–20.0)	<0.001	42.0 (26.0–216.0)	17.0 (14.0–21.0)	<0.001
tumor number, N (%)
1	180 (90.5%)			76 (89.4%)			90 (82.6%)
2	12 (6.0%)			6 (7.1%)			15 (13.8%)
3	7 (3.5%)			3 (3.5%)			4 (3.7%)
tumor size, cm	2.4 (1.6–3.0)			2.4 (1.5–3.7)			2.5 (1.5–3.2)
AJCC stage, n (%)
IA	77 (38.7%)			31 (36.5%)			42 (38.5%)
IB	105 (52.8%)			45 (52.9%)			48 (44.0%)
II	17 (8.5%)			9 (10.6%)			19 (17.4%)

Data are provided in N (%) or median (interquartile range).

HCC, hepatocellular carcinoma; ALT, alanine aminotransferase; INR, international normalized ratio; CHB, chronic hepatitis B; CHC, chronic hepatitis C; AFP, α-fetoprotein; PIVKA-II, protein induced by vitamin K absence or antagonist-II; AJCC, American Joint Committee on Cancer; IQR, interquartile range.

Development of an MRM-MS Panel

Of the 79 candidate proteins that were significantly associated with the presence of HCC (P < 0.05), 17 proteins [PPBP (platelet basic protein), SERPINC1 (antithrombin-III), MCAM (cell surface glycoprotein), CFL1 (cofilin-1), C4A (complement C4-A), AMBP (protein AMBP), THBS1 (thrombospondin-1), UCHL3 (ubiquitin carboxyl-terminal hydrolase isozyme L3), LCAT (phosphatidylcholine-sterol acyltransferase), C1QC (complement C1q subcomponent subunit C), CNDP1 (β-Ala-His dipeptidase), C2 (cytochrome c oxidase subunit 2), CA2 (carbonic anhydrase 2), C6 (complement component C6), SAA4 (serum amyloid A-4 protein), SERPINA10 (protein Z-dependent protease inhibitor), and APOH (β-2-glycoprotein 1)] were selected as optimal constituents of the MRM-MS panel by stepwise selection. The predicted probability of HCC cases according to the panel was calculated in the training set (shown in Table S1).

Performance of the MRM-MS Panel Excluding General Proteins

Although 17 protein markers were suggested by machine learning, several were not specifically related to HCC. The performance of a panel that excluded general proteins, such as C4A, C6, SAA4, and APOH, was confirmed prior to the analysis of the 17-protein panel (Figure S2).

In the training set, the standalone MRM-MS panel had a significantly higher AUROC value (0.912 vs 0.759; P < 0.001) than AFP alone. The addition of AFP to the MRM-MS panel (designated the MA panel) improved the AUROC value (0.939; 95% CI, 0.910–0.956; P = 0.025), as did that of PIVKA-II to it (termed the MP panel) (0.926; 95% CI, 0.901–0.951; P = 0.175). The combination MRM-MS panel with AFP and PIVKA-II (the MAP panel) had a significantly higher AUROC value (0.938; 95% CI, 0.915–0.960; P = 0.005) than AFP alone.

In the test set, the MRM-MS panel yielded a significantly higher AUROC value (0.919 vs 0.691; P < 0.001) than AFP alone. The AUROC values of the MA (0.919; 95% CI, 0.877–0.962; P = 0.841), MP (0.928; 95% CI, 0.889–0.967; P = 0.191), and MAP panels (0.930; 95% CI, 0.891–0.970; P = 0.288) were comparable with that of the standalone MRM-MS panel.

In the confirmation set, the MRM-MS panel had a significantly higher AUROC value (0.947 vs 0.779, P < 0.001) than AFP alone. The AUROC values of the MA (0.9577; 95% CI, 0.930–0.985; P < 0.001), MP (0.966; 95% CI, 0.942–0.990; P < 0.001), and MAP panels (0.973; 95% CI, 0.952–0.994; P < 0.001) were also higher than that of the standalone MRM-MS panel.

In a machine learning-based biomarker study, the classifier processes the best set of combination solely using the given quantitative value of the data. From our result, the 17-marker combination, including the four general markers, performed best in distinguishing patients and healthy controls alone, without AFP or PIVKA values, which are not quantifiable by MRM-MS.

In addition, considering that blood samples that are procured from the median cubital vein do not represent a specific organ, certain markers that are irrelevant to HCC might have been selected as candidates. Because the presence of HCC might impact not merely the immune and endocrine systems but also the entire body, removing markers that are related to inflammation, which would show a definitive difference between patients and controls, might decrease the accuracy of a multimarker panel that is developed for surveillance.

Performance of the MRM-MS Panel with and without AFP and PIVKA-II in Detecting HCC

In the training set, the standalone MRM-MS panel had a significantly higher AUROC value (0.937 vs 0.759; P < 0.001) than AFP alone. The MRM-MS panel had greater sensitivity (80.4 vs 17.1%) and lower specificity (93.5 vs 97.5%) than AFP. The addition of AFP to the MRM-MS panel improved the AUROC value (0.949; 95% CI, 0.929–0.969; P = 0.006), whereas that of PIVKA-II to it did not (0.941; 95% CI, 0.920–0.963; P = 0.073). The combination MRM-MS panel with AFP and PIVKA-II had a significantly higher AUROC value (0.953; 95% CI, 0.934–0.971; P = 0.002), similar sensitivity (80.4%), and greater specificity (96.5%) compared with the standalone MRM-MS panel (shown in Figures and S3).

Figure 2

Performance of AFP and the MRM-MS panel with and without AFP and PIVKA-II in detecting HCC. ROC curves of AFP, standalone MRM-MS panel, MRM-MS + AFP (MA) panel, MRM-MS + PIVKA-II (MP) panel, and MRM-MS + AFP + PIVKA-II (MAP) panel in the (A) training set and (B) confirmation set.

In the test set, the MRM-MS panel yielded a significantly higher AUROC value (0.891 vs 0.691; P < 0.001), greater sensitivity (78.8 vs 24.7%), and lower specificity (81.2 vs 97.7%) than AFP. The AUROC values of the MA (0.891; 95% CI, 0.844–0.938; P = 0.598), MP (0.901; 95% CI, 0.856–0.945; P = 0.086), and MAP panels (0.907; 95% CI, 0.863–0.952; P = 0.280) were comparable with that of the standalone MRM-MS panel. However, the sensitivity and specificity were higher in the MAP panel (81.2 and 82.3%, respectively) versus the standalone MRM-MS panel.

In the confirmation set, the MRM-MS panel had a significantly higher AUROC value (0.940 vs 0.779, P < 0.001), higher sensitivity (88.1 vs 35.8%), and lower specificity (82.0 vs 96.0%) than AFP. The AUROC values of the MA (0.958; 95% CI, 0.929–0.986; P = 0.004), MP (0.953; 95% CI, 0.923–0.983; P = 0.001), and MAP panels (0.970; 95% CI, 0.947–0.994; P < 0.001) were also higher than that of the standalone MRM-MS panel. The sensitivity and specificity of the MAP panel were 89.9 and 98.0%, respectively, both higher compared with the standalone MRM-MS panel (shown in Table ).

Table 2

Performance of AFP and the MRM-MS Panel with and without AFP and PIVKA-II in Detecting HCCa

	AUROC 95% CI	P value	sensitivity (%)	specificity (%)
Training Set
AFP	0.759 (0.712–0.806)		17.1	97.5
standalone MRM-MS panel	0.937 (0.915–0.959)	<0.001b	80.4	93.5
MA panel	0.949 (0.929–0.969)	0.006c	88.4	87.5
MP panel	0.941 (0.920–0.963)	0.073c	81.4	94.5
MAP panel	0.953 (0.934–0.971)	0.002c	80.4	96.5
Test Set
AFP	0.691 (0.610–0.772)		24.7	97.7
standalone MRM-MS panel	0.891 (0.844–0.938)	<0.001b	78.8	81.2
MA panel	0.898 (0.852–0.945)	0.598c	89.4	74.1
MP panel	0.901 (0.856–0.945)	0.086c	81.2	78.8
MAP panel	0.907 (0.863–0.952)	0.280c	81.2	82.3
Confirmation Set
AFP	0.779 (0.707–0.852)		35.8	96.0
standalone MRM-MS panel	0.940 (0.906–0.975)	<0.001b	88.1	82.0
MA panel	0.958 (0.929–0.986)	0.004c	92.7	70.0
MP panel	0.953 (0.923–0.983)	0.001c	88.1	98.0
MAP panel	0.970 (0.947–0.994)	<0.001c	89.9	98.0

AUROC—area under the receiver operating characteristics curve; AFP—α-fetoprotein; MA—multiple reaction monitoring–mass spectrometer + AFP; MP—multiple reaction monitoring–mass spectrometer + protein induced by vitamin K absence or antagonist-II; MAP—multiple reaction monitoring–mass spectrometer + AFP + protein induced by vitamin K absence or antagonist-II.

AUROC of the standalone MRM-MS panel versus AUROC of AFP.

AUROC of MA, MP, or MAP panel versus AUROC of the standalone MRM-MS panel.

Performance of the MRM-MS Panel with and without AFP and PIVKA-II in Detecting Small (<2 cm) Single HCC

Patients with small (i.e., <2 cm) single HCC and high-risk controls in each set were analyzed. In the training set, the standalone MRM-MS panel had a significantly higher AUROC value (0.929 vs 0.775, P < 0.001), greater sensitivity (76.6 vs 22.1%), and lower specificity (93.5 vs 97.5%) compared with AFP. The AUROC values of the MA (0.938; 95% CI, 0.909–0.968; P = 0.072), MP (0.931; 95% CI, 0.898–0.963; P = 0.498), and MAP panels (0.940; 95% CI, 0.911–0.969; P = 0.074) were similar to that of the standalone MRM-MS panel. The sensitivity and specificity of the MAP panel were 72.7 and 96.5%, respectively (shown in Figures and S3).

Figure 3

Performance of AFP and the MRM-MS panel with and without AFP and PIVKA-II in detecting small (<2 cm) single HCC. ROC curves of AFP, standalone MRM-MS panel, MRM-MS + AFP (MA) panel, MRM-MS + PIVKA-II (MP) panel, and MRM-MS + AFP + PIVKA-II (MAP) panel in the (A) training set and (B) confirmation set.

In the test set, the MRM-MS panel yielded a significantly higher AUROC value (0.893 vs 0.593; P < 0.001), better sensitivity (74.2 vs 22.6%), and lower specificity (81.2 vs 97.5%) than AFP. The AUROC values of the MA (0.895; 95% CI, 0.934–0.957; P = 0.882), MP (0.893; 95% CI, 0.832–0.953; P = 0.938), and MAP panels (0.894; 95% CI, 0.832–0.957; P = 0.943) were comparable with that of the standalone MRM-MS panel. The sensitivity and specificity of the MAP panel were 77.4 and 82.4%, respectively.

In the confirmation set, the standalone MRM-MS panel had a significantly higher AUROC value (0.937 vs 0.806; P = 0.019), greater sensitivity (88.1 vs 31.0%), and lower specificity (82.0 vs 96.0%) than AFP. The AUROC values of the MP (0.946; 95% CI, 0.892–1.000; P = 0.04) and MAP panels (0.961; 95% CI, 0.912–1.000; P = 0.026) were significantly higher versus the standalone MRM-MS panel, unlike the MA panel (0.955; 95% CI 0.901–1.000; P = 0.062). The sensitivity and specificity of the MAP panel were 90.5 and 98.0%, respectively (shown in Table ).

Table 3

Performance of AFP and the MRM-MS Panel with and without AFP and PIVKA-II in Detecting Small (<2 cm) Single HCCa

	AUROC 95% CI	P value	sensitivity (%)	specificity (%)
Training Set
AFP	0.775 (0.710–0.813)		22.1	97.5
standalone MRM-MS panel	0.929 (0.896–0.962)	<0.001b	76.6	93.5
MA panel	0.938 (0.909–0.968)	0.072c	83.1	87.4
MP panel	0.931 (0.898–0.963)	0.498c	76.6	94.5
MAP panel	0.940 (0.911–0.969)	0.074c	72.7	96.5
Test Set
AFP	0.593 (0.464–0.721)		22.6	97.5
standalone MRM-MS panel	0.893 (0.834–0.952)	<0.001b	74.2	81.2
MA panel	0.895 (0.834–0.957)	0.882c	90.3	74.1
MP panel	0.893 (0.832–0.953)	0.938c	74.2	78.8
MAP panel	0.894 (0.832–0.957)	0.943c	77.4	82.4
Confirmation Set
AFP	0.806 (0.717–0.896)		31.0	96.0
standalone MRM-MS panel	0.937 (0.878–0.996)	0.019b	88.1	82.0
MA panel	0.955 (0.901–1.000)	0.062c	95.2	70.0
MP panel	0.946 (0.892–1.000)	0.040c	88.1	98.0
MAP panel	0.961 (0.912–1.000)	0.026c	90.5	98.0

AUROC—area under the receiver operating characteristics curve; AFP—α-fetoprotein; MA—multiple reaction monitoring–mass spectrometer + AFP; MP—multiple reaction monitoring–mass spectrometer + protein induced by vitamin K absence or antagonist-II; MAP—multiple reaction monitoring–mass spectrometer + AFP + protein induced by vitamin K absence or antagonist-II.

AUROC of the standalone MRM-MS panel versus AUROC of AFP.

AUROC of MA, MP, or MAP panel versus AUROC of the standalone MRM-MS panel.

Performance of the MRM-MS Panel with and without AFP and PIVKA-II vs GALAD in Detecting HCC

The GALAD cohort was a subset of the confirmation set and was composed of 53 HCC patients and 50 controls from a single participating center (AMC) (Table S2). The MRM-MS panel had a significantly higher AUROC value (0.931 vs 0.718; P <0.001), greater sensitivity (88.7 vs 52.8%), and lower specificity (82.0 vs 92.0%) than the GALAD score. The AUROC value of the MAP panel was significantly higher (0.958; 95% CI, 0.916–1.000; P = 0.009) than the standalone MRM-MS panel (Figure S4). The addition of AFP and PIVKA-II to the MRM-MS panel improved its sensitivity (90.6%) and specificity (98.0%) (shown in Table S3).

Subgroup Analysis

A separate subgroup analysis was performed for patients with cirrhosis in each set and those with noncirrhotic CHB in the training and test sets. An analysis of participants with noncirrhotic CHB in the confirmation set was not feasible because all controls in the set had cirrhosis.

For patients with cirrhosis, the standalone MRM-MS model yielded a significantly higher AUROC value (0.927 vs 0.790; P < 0.001), higher sensitivity (76.6 vs 20.2%), and lower specificity (94.9 vs 97.5%) than AFP in the training set. The AUROC value of MAP panel (0.945; 95% CI, 0.919–0.970; P = 0.021) was significantly higher versus the standalone MRM-MS panel. In the test and confirmation sets, the MRM-MS model also had a significantly higher AUROC value (0.875 and 0.944 vs 0.712 and 0.798; P = 0.005 and P < 0.001, respectively), greater sensitivity (76.8 and 88.9% vs 28.6 and 36.7%, respectively), and lower specificity (78.1% and 82.0 vs 96.9 and 96.0%, respectively) than AFP alone. The AUROC value of the MRM-MS panel improved with the addition of AFP and PIVKA-II (i.e., MAP panel) in the confirmation set (0.973; 95% CI, 0.949–0.997; P = 0.001), but this increase was not significant in the test set (0.897; 95% CI, 0.839–0.954; P = 0.303) (shown in Table S4A and Figure S5).

For patients with noncirrhotic CHB in the training set, the MRM-MS panel had a significantly higher AUROC value (0.951 vs 0.734; P < 0.001), greater sensitivity (86.2 vs 13.8%), and lower specificity (85.3 vs 97.1%) than AFP, whereas the MAP panel had a statistically similar AUROC value (0.962; 95% CI, 0.929–0.994; P = 0.111) as the standalone MRM-MS panel. In the test set, the MRM-MS panel also had a significantly higher AUROC value (0.909 vs 0.702; P < 0.001) compared with AFP alone, and the addition of AFP and PIVKA-II (i.e., the MAP panel) did not significantly improve the AUROC value (0.931; 95% CI, 0.854–1.000; P = 0.298) (shown in Table S4B and Figure S5).

We have developed a multimarker panel of 17 proteins by MRM-MS analysis for detecting HCC. This MRM-MS panel performed significantly better than AFP in differentiating HCC patients from controls in every dataset, especially with regard to sensitivity. The incorporation of AFP and PIVKA-II into the MRM-MS panel improved its performance in every set, although statistical significance was not achieved with the test set. The diagnostic superiority of the MRM-MS panel over AFP was maintained in patients with small single HCC, but the significant improvement that was affected by the addition of AFP and PIVKA-II was observed only in the confirmation set. The MRM-MS panel also outperformed the GALAD score and demonstrated its efficacy in identifying HCC patients among cirrhotic patients and those with noncirrhotic CHB.

The low concentrations of potential biomarkers and the wide range of protein levels are major setbacks in the proteomic analysis of serum.[23] MRM-MS, which can detect attomole levels of peptides[24−27] and quantify hundreds of peptides in an automated manner,[28,29] is an effective modality in examining and validating candidate biomarkers for early detection of HCC. The data that are generated by MRM-MS assays are highly reproducible.[30]

HCC surveillance programs that are based on abdominal ultrasonography with and without serum AFP assay have low sensitivities in cirrhotic patients,[8,31] whose alterations in liver parenchyma are associated with poor imaging quality[32] and a high risk of HCC.[33] Although recent studies have examined the value of cross-sectional imaging modalities (e.g., computed tomography and magnetic resonance imaging) as potential alternatives,[34−36] their limited accessibility and high cost are major barriers to their widespread use.[37] HCC surveillance using an MRM-MS panel has the advantages of being readily available and cost-effective because it requires only blood samples for mass spectrometry assays.

No study that has aimed to identify new HCC serum markers through mass spectrometry[38−42] has validated the discriminatory performance of the identified markers in a chronologically separately collected, multicenter cohort. The current study is also distinct from a previous report by our group,[17] in that data preprocessing was adopted before the selection of the markers and that fewer markers were involved in the developed panel. The smaller number of markers renders the MRM-MS panel economically competitive with regard to the amount of internal standard peptides and reagents that are needed for mass spectrometry assays.[43]

Our MRM-MS panel showed excellent accuracy in discriminating HCC patients, with an AUROC value of 0.891 being the lowest across all analyses, considering that a value over 0.8 is excellent.[44] However, its specificity was lower than that of AFP, as was the case with our group’s earlier panel. AFP, despite its low sensitivity of 39–65% in detecting HCC,[45] improved the sensitivity when added to the MRM-MS panel, although its specificity declined. This drawback was overcome by introducing PIVKA-II, which is more specific than AFP in detecting HCC.[46] A similar trend was observed when the analysis was confined to patients with small single HCC. These results imply that the serum levels of proteins in the MRM-MS panel are affected by the presence of HCC through mechanisms that are independent of those that affect the levels of AFP and PIVKA-II. Notably, the MRM-MS panel was consistently reliable when classifying patients with cirrhosis and those with noncirrhotic CHB, in contrast to AFP, the levels of which are frequently elevated in patients with cirrhosis or CHB.[47]

Our study has several limitations. The depletion of the six most abundant proteins before the MRM-MS assays might have eliminated other proteins with significant discriminatory value, as inferred from the finding that AFP was not included in the list of the 17 markers. This issue has been discussed in other proteomic studies[48,49] and was addressed in our study through the subsequent addition of AFP, which is the only commonly used serum biomarker for screening HCC.[45]

Further, it is unknown whether the MRM-MS panel is useful for patients with other risk factors for HCC, such as alcoholic liver cirrhosis, or nonalcoholic fatty liver disease. Because the main etiological cause of HCC is shifting from CHB or CHC to NASH or ALD due to vaccination and antivirals, further validation in the ALD and NASH populations is required.[50,51] Also, the effectiveness of the MRM-MS panel in patients of various ethnicities is unverified—all of our participants were Korean. Finally, no comparison with abdominal ultrasonography was performed in this study.

Conclusions

In conclusion, this MRM-MS multimarker panel, comprising 17 proteins, showed excellent performance in distinguishing HCC patients from high-risk controls with cirrhosis or CHB or CHC. Its combination with AFP and PIVKA-II enhanced its performance, although statistical significance was not consistently reached. Prospective studies are warranted to determine whether our MRM-MS panel is a viable alternative to abdominal ultrasonography in HCC surveillance.