Serum Proteomic Profile Based on the TGF‐β Pathway Stratifies Risk of Hepatocellular Carcinoma

ABSTRACT

Background

Hepatocellular carcinoma (HCC) is the third leading cause of cancer‐related deaths, primarily due to late‐stage diagnosis. In this multicenter study, our goal is to identify functional biomarkers that stratify the risk of HCC in patients with cirrhosis (CP) for early diagnosis.

Methods

Five thousand and eight serum proteins (Somascan) were analysed in Cohort A (477 CP, including 125 HCC). Clustering analysis of the TGF‐β pathway‐associated protein signature was performed in a longitudinal, prospective Cohort B (312 CP, in which 18 cases developed HCC over a 5‐year follow‐up period). Next, a multivariable prediction model was built using logistic regression analysis of cross‐sectional data from a matched subgroup (n = 328, Cohort C). Model performance was 10‐fold cross‐validated across the entire Cohort A (n = 477).

Results

Longitudinal follow‐up analysis revealed that patients with elevated TGF‐β‐related protein signature displayed a five‐fold increased risk of developing HCC (9.68% vs. 1.91%). Compared to cirrhosis, serum MSTN, TGFBR2, and AFP levels raised in HCC were validated by ELISA (n = 200, odds ratio = 1.4–2.9, p < 0.05). In Cohort C, 88 proteins were significantly altered in HCC compared to cirrhosis (p < 0.05). The six‐protein panel (TGFBR2, MSTN, AFP, COL18A1, GLUL, TP63) displayed a strong performance in the matched cohort C (AUC 0.87, sensitivity 0.88, specificity 0.72), alongside four clinical factors (Age, Sex, BMI, Bilirubin). A 10‐fold cross‐validation demonstrated a mean AUC of 0.86 in cohort A, with strong predictive power in obese/MASLD/ALD‐related patients (AUCs: 0.862–0.921).

Conclusions

The mechanism‐based panel effectively stratifies HCC risk in cirrhotic patients, underscoring the need for Phase II/III validation.

Summary.

• Serum TGF‐β‐related proteins (TGFBR2 and MSTN) are significantly elevated in cirrhotic patients with HCC compared to those with cirrhosis alone.

• Elevated serum TGF‐β‐related protein signature is associated with a five‐fold increased risk of HCC development in cirrhotic patients.

• A TGF‐β pathway‐based protein panel demonstrates superior diagnostic performance for HCC in clinically relevant subgroups stratified by liver disease aetiology, obesity, and AFP status.

Abbreviations

AFP

α‐fetoprotein

ALP

alkaline phosphatase

ALT

alanine aminotransferase

AST

aspartate aminotransferase

BMI

body mass index

GGT

gamma glutamyl transferase

HBV

hepatitis B virus

HCC

hepatocellular carcinoma

HCV

hepatitis C virus

LI‐RADS

Liver Reporting and Data System

MAPK

mitogen‐activated protein kinase

MASH

nonalcoholic steatohepatitis

NSAIDs

nonsteroid anti‐inflammatory drugs

PI3K

phosphoinositide 3‐kinase

RFU

relative fluorescence units

TGF‐β

transforming growth factor β

1. Introduction

The importance of effective screening strategies for HCC remains urgent due to the poor prognosis associated with late‐stage diagnosis at presentation. Early‐stage HCC is amenable to effective therapies such as hepatic resection or liver transplantation, with a 5‐year survival ranging from 70% to 80% [1, 2]. However, the median overall survival of advanced‐stage HCC is only 2 to 6 months [3] despite notable recent advances in systemic treatments such as immunotherapy. Current screening strategies use abdominal ultrasound with or without alpha‐fetoprotein (AFP); however, recent large‐scale studies have highlighted the limitations of abdominal ultrasound in the detection of early‐stage HCC, particularly in patients with obesity and non‐viral etiologies of cirrhosis [4, 5, 6]. The combination of ultrasound and AFP has a higher sensitivity than ultrasound alone, but still misses approximately one‐third of early‐stage HCC.

Multiple prediction models have been developed to combine the predictive power of key clinical variables with blood‐based protein markers, addressing risk stratification and early detection of HCC in patients with cirrhosis [7]. Of emerging biomarker strategies, the GALAD panel, including gender, age, AFP, AFP‐L3, and Des‐gamma‐carboxy prothrombin (DCP) is the best validated [8]. Other promising approaches include methylated DNA markers, cell‐free DNA, and extracellular vesicle‐based protein assays [9, 10, 11]. Blood candidate markers for HCC include: (i) Osteopontin (OPN), Golgi Protein 73 (GP73), and Glypican 3 (GPC3), which perform well in AFP‐negative patients with HCC [12, 13, 14]; and (ii) Midkine (MDK) and Dickkopf‐related protein 1 (DKK1), which outperform AFP in diagnosing early‐stage HCC [15, 16]. Clinical variables that are potential indicators of risk for HCC include: (i) demographics (age, sex, race); (ii) laboratory test values albumin, platelets, bilirubin, fibroblast growth factor 21, vitamin D, serum liver enzymes [alkaline phosphatase (ALP), aspartate aminotransferase (AST), ALT, gamma‐glutamyl transferase (GGT), AFP]; (iii) liver stiffness; (iv) other comorbidities [high body mass index (BMI), diabetes mellitus, hepatitis B virus (HBV), hepatitis C virus (HCV)]; (v) behaviour (smoking history, alcohol intake, diet, exercise); and (vi) medication history [statins, metformin, aspirin, other nonsteroid anti‐inflammatory drugs (NSAIDs)]. However, an integrated and reproducible set of markers and models reflecting proteins mechanistically involved in the pathobiology of cirrhosis progression to cancer, which also incorporates previously described markers, has yet to be identified.

Identifying biomarkers for HCC risk stratification and prediction is also challenging due to the molecular heterogeneity, multiple etiologies, and divergent pathophysiology of the tumour [17, 18]. A comprehensive understanding of the genetic and epigenetic alterations driving HCC progression requires a multi‐faceted approach, combining large‐scale human genomic studies with mechanistic investigational research using animal models. Factors associated with the initiation and progression of HCC include viral hepatitis, exposure to toxins, metabolic syndromes, and gene mutations [19, 20, 21]. The Human Cancer Genome Atlas (TCGA) related studies have provided valuable insights into the genomic landscape of human HCC [22, 23, 24, 25, 26, 27], revealing frequent mutations in key pathways such as WNT (44%), p53 (31%), and TERT (44%), as well as epigenetic alterations like CDKN2A silencing (53%) and TGF‐β signalling dysregulation (43%). PI3K, MYC, and MET signalling pathway alterations also contribute to HCC pathogenesis [28, 29, 30].

Among these pathways, TGF‐β signalling is significant because it plays a major role in the progression of disease from fibrosis to cirrhosis, through the modulation of stromal, immune, and hepatocyte functions, ultimately leading to hepatocarcinogenesis [31, 32]. In addition, genes encoding components and regulators of the TGF‐β signalling pathway are altered in ~40% of human liver and gastrointestinal cancers [24]. Furthermore, multiple genetically engineered mouse models with impaired TGF‐β signalling are susceptible to HCC [31].

In this study, we address the challenge of identifying new markers by hypothesizing that functional markers reflecting ongoing pathology will yield important biomarkers that stratify risk for HCC. Additionally, we sought to incorporate and validate well‐described human markers to identify a serum protein‐based panel that differentiates between HCC and non‐HCC in human patients with cirrhosis. Earlier, we identified a significant potential value of the TGF‐β pathway‐associated protein signature in HCC risk prediction for cirrhotic patients in a retrospective cohort comprising 216 serum samples collected from five Institutions [33]. In this prospective cohort (Cohort B), we employed both targeted and untargeted approaches, focusing on a subset of potential protein biomarkers associated with the TGF‐β pathways, which queried approximately 5000 serum proteins. Using cross‐sectional data from the matched cohort C, we identified a six‐protein panel of TGF‐β Pathway‐related biomarkers with functional relevance and predictive value in detecting HCC and early‐stage HCC in patients with cirrhosis. Through cross‐validation, the performance of the six‐protein panel was subsequently validated in the entire cohort A and separate subgroups of cirrhotic patients, demonstrating the model's strength across various settings of cirrhosis, including different aetiologies, the presence or absence of obesity, and variable AFP status.

2. Patients and Methods

2.1. Study Population

In this multicenter study, aimed at elucidating HCC risk prediction using serum biomarkers in cirrhotic patients, we recruited adult patients from five medical centers participating in the Translational Research Consortium for the Early Detection of Liver Cancer, which is sponsored by the National Institutes of Health. These centers included George Washington University (GW), the University of Maryland (UMD), the University of California at Los Angeles (UCLA), the University of Hawaii (UH), and Northwell Health (NW). Eligible patients were categorised as (i) Cirrhosis with treatment‐naïve HCC (designated HCC+) or (ii) cirrhosis without HCC (designated HCC−), both at enrollment and on imaging done within 6 months after enrollment to exclude prevalent HCC. HCC diagnosis required either histological confirmation or characteristic findings on contrast‐enhanced imaging (CT or MRI) that satisfied the Liver Imaging Reporting and Data System (LI‐RADS) 5 lesion(s). Early‐stage HCC was defined as Barcelona Clinic Liver Cancer Stage A. Cirrhosis was diagnosed using serum biomarkers (FibroSure/FibroTest > 0.74, APRI > 2, or FIB‐4 > 3.25), histology, imaging, elastography, or clinical evidence of portal hypertension. The research was conducted in accordance with both the Declaration of Helsinki and the Istanbul Declaration. Written informed consent was obtained from each patient. The Institutional Review Board approved the study at all participating institutions. Clinical and laboratory data were collected at each center at the time of enrollment, and the Child‐Pugh score was calculated based on enrollment data.

The study included 477 patients with cirrhosis, enrolled from 2011 to 2022 (Cohort A): 125 with hepatocellular carcinoma (HCC+) and 352 cirrhotic‐only patients (HCC−). All clinical characteristics of these patients were provided in Table S1. For the prospective cohort (Cohort B), 312 patients with cirrhosis were followed longitudinally over a median period of 5 years (Figure S1). We utilised Propensity Score Matching (PSM) based on age, sex, and aetiology status to match the HCC+ patients with HCC− patients across the entire cohort A with the R package “MatchIt” (version 4.5.5). Patients matched via PSM were assigned to the matched cohort (n = 328, Cohort C, Figure S1). The patients included in the matched cohort C were matched according to a mixed ratio (1:1, 1:2, or 1:3), aiming to maximise sample size and mimic the imbalance between HCC− and HCC+ in real clinical settings.

2.2. Serum Proteomics Profiling

Four hundred seventy‐seven serum samples from Cohort A collected at enrollment were subjected to proteomic analysis by the SOMAscan Assay (Somalogic, Boulder, CO) [34]. The version of the SOMAscan Assay in use at the time of sample submission was the ~5000‐plex in the first batch (n = 146) and the ~7000‐plex in the second (n = 70) and third batches (n = 261). All samples passed the QC criteria. A linear scale provided by SOMAscan was used for data preprocessing to adjust for batch effects between different versions. All Somascan results of the assay were provided as relative fluorescence units (RFUs, Table S2). Six TGF‐β pathway‐related potential markers for HCC (MSTN, TGFBR2, SPTBN1, HMGB1, IL‐6, COL1A1) together with AFP, were measured by enzyme‐linked immunosorbent assay (ELISA) following the manufacturer's protocols (AFP: Cobas e602; MSTN, TGFBR2, IL‐6, COL1A1: Boster Bio EK1265, EK1307, EK0410, EK1692; HMGB1: Biomatik EKF57253; SPTBN1: #MBS9329768) in patients with cirrhosis (n = 200, 57 with HCC, Figure S1).

2.3. Statistical Analysis

For the SomaScan data from the matched cohort C (n = 328), we calculated mean values for HCC+ and HCC−, fold change (HCC+ vs. HCC−), and odds ratios (ORs) from conditional logistic regression models and a multivariable model adjusted by potential covariates (age, sex, aetiology, BMI, Bilirubin, and AFP). Multiple testing was accounted for by calculating the false discovery rate (FDR)‐adjusted p‐value. Median imputation was conducted for the missing clinical lab results. For all dysregulated proteins with an absolute fold change of more than 1.2 and a p‐value less than 0.05 from conditional logistic regression, we performed pathway and biology enrichment analysis using QIAGEN Ingenuity Pathway Analysis [35] and protein–protein interaction network analysis using the STRING database version 12.0 [36]. Clustering analyses were performed with the R package “Complexheatmap” (version 2.12.1). We compared the areas under the receiver operating characteristic (ROC) curve (AUCs) of three models: (1) using only the traditional risk factors (basic model: age, sex, BMI, bilirubin); (2) using the selected proteins only (protein model); and (3) combining the basic and protein models. Sequential backward and forward selections were performed to optimise the final panel. The flow chart provides a summary of the protein marker selection process, as shown in Figure S2. We used multivariable logistic regression for model building and validation. We evaluated the performance of our panel utilising 10‐fold cross‐validation [37]. First, we randomly divided the entire cohort A (n = 477) into 10 unique training sets (including 90% of the whole dataset) and validation sets (the remaining 10%). The prediction model was then developed within each training dataset using the selected panel of variables, and the validation datasets were used to calculate the AUC. We calculated the average AUC and its standard deviation (95% confidence interval [CI]). Proteins were log‐transformed to reduce skewness and remove outliers. The SMOTE algorithm was employed to address class imbalance while evaluating our panel in the subgroups of the combined cirrhotic cohort [38]. SAS (version 9.4, Cary, NC) and R (version 4.2.1) were used for data analysis, and unless otherwise specified, p < 0.05 indicates a significant effect.

3. Results

3.1. Patient Characteristics

Among the prospective cohort B (n = 312, cirrhotic patients), 18 patients (5.8%) developed HCC based on imaging criteria during a median follow‐up period of 5 years. Based on the baseline characteristics (Table 1), the cirrhotic patients who developed HCC during follow‐up (n = 18) displayed lower baseline serum bilirubin levels (p < 0.05) and higher serum albumin levels (p < 0.01) compared to those without HCC progression. Baseline characteristics of the matched cohort C (n = 328, 109 pairs of HCC+ and matched HCC−) are shown in Table 2. HCC+ patients exhibited higher serum levels of AFP (p = 0.001) and ALT (p = 0.007), and a lower serum level of bilirubin (p = 0.043) compared to the HCC− group.

TABLE 1.

Demographic and clinical data for the prospective cirrhotic participants (n = 312, Cohort B) using SOMAscan assay.

Demographic or clinical parameter	Entire cohort B (n = 312)	Cirrhosis only (n = 294)	HCC developed (n = 18)	p
Age mean (SD)	58 (12)	58 (12)	59 (12)	0.718 a
Males, No. (%)	190 (61%)	180 (61%)	10 (56%)	0.632 b
Race or ethnicity				0.150 b
Asian	19 (6.1%)	18 (6.1%)	1 (5.6%)
African American	39 (12.5%)	36 (12.2%)	3 (16.7%)
White	215 (68.9%)	204 (69.4%)	11 (61.1%)
Pacific islander or Other	29 (9.0%)	26 (8.8%)	3 (16.7%)
Hispanic	10 (3.2%)	10 (3.4%)	0 (0%)	0.384 b
BMI mean (SD), (kg/m2)	29.8 (6.5)	29.9 (6.6)	27.8 (5.2)	0.122 a
Aetiology, No. (%)
HBV	12 (3.8%)	11 (3.7%)	1 (5.6%)	0.698 b
HCV	95 (30.4%)	90 (30.6%)	5 (27.8%)	0.800 b
MASLD/MASH	94 (30.1%)	89 (30.3%)	5 (27.8%)	0.823 b
ALD	112 (35.9%)	103 (35.0%)	9 (50%)	0.199 b
Child Pugh Class, No. (%)				< 0.05 b
A	116 (37.2%)	105 (35.8%)	11 (61.1%)
B	118 (37.8%)	113 (38.4%)	5 (27.8%)
C	62 (19.9%)	62 (21.1%)	0 (0%)
Unknown	16 (5.1%)	14 (4.8%)	2 (11.1%)
Lab values, median [IQR]
ALT, U/L	33 [23–52]	33 [24–53]	25.5 [21–43.8]	0.073 c
AST, U/L	47.5 [33.2–76]	48.5 [34–77.8]	38 [33–61.8]	0.256 c
ALP, U/L	121 [85–171]	121.5 [85–171]	102 [77–220]	0.778 c
Bilirubin, mg/dL	1.4 [0.7–2.8]	1.5 [0.7–2.8]	1.0 [0.7–1.2]	< 0.05 c
Creatinine, mg/dL	0.9 [0.7–1.2]	0.9 [0.7–1.2]	0.9 [0.7–1.1]	0.935 c
Albumin, g/dL	3.7 [3–4.3]	3.6 [3–4.3]	4.2 [3.8–4.6]	< 0.01 c
AFP, ng/mL	3.7 [2.4–6.1]	3.7 [2.4–6.3]	3.7 [2.5–4.3]	0.710 c

^a t‐test.

^b Chi‐square test.

^c Mann Whitney U Test.

TABLE 2.

Baseline characteristics of cirrhotic patients with HCC (HCC+) and cirrhotic patients without HCC (HCC−) from matched (n = 328, Cohort C) and unmatched patients (n = 149).

Characteristics	Matched (n = 328, cohort C)		Unmatched (n = 149)
	HCC+ (n = 109)	HCC− (matched, n = 219)	HCC+ (n = 16)	HCC− (n = 133)
Age, mean (SD), y	64.79 (7.77)	63.02 (10.45)	68.44 (9.91)	52.23 (10.83)
Male, No. (%)	85 (78.0%)	160 (73.1%)	16 (100.0%)	55 (41.4%)
Race or Ethnicity, No. (%)
White	53 (48.6%)	154 (70.3%)	8 (50.0%)	93 (69.9%)
African American	22 (20.2%)	22 (10.0%)	4 (25.0%)	18 (13.5%)
Asian	15 (13.8%)	21 (9.6%)	2 (12.5%)	2 (1.5%)
Pacific islander or Other	19 (17.4%)	22 (10.0%)	2 (12.5%)	20 (15%)
Hispanic, No. (%)	4 (3.7%)	3 (1.4%)	0 (0.0%)	7 (5.3%)
BMI, mean (SD), kg/m2	29.94 (5.82)	30.31 (6.29)	27.17 (3.83)	28.57 (6.84)
Aetiology, No. (%)
HBV	10 (9.2%)	13 (5.9%)	7 (43.8%)	0 (0%)
HCV	53 (48.6%)	88 (40.2%)	10 (62.5%)	12 (9%)
MASLD/MASH	29 (26.6%)	86 (39.3%)	2 (12.5%)	30 (22.6%)
ALD	35 (32.1%)	53 (24.2%)	10 (62.5%)	73 (54.9%)
Child Pugh Class, No. (%)
A	66 (60.6%)	92 (42.0%)	7 (43.8%)	42 (31.6%)
B	37 (33.9%)	84 (38.4%)	7 (43.8%)	47 (35.3%)
C	6 (5.5%)	30 (13.7%)	2 (12.5%)	39 (29.3%)
Unknown	0 (0.0%)	13 (5.9%)	0 (0.0%)	5 (3.8%)
Lab values, median [IQR]
ALT, U/L	39 [24–65]	31 [22–46]	35.5 [18–54.5]	36 [25–51]
AST, U/L	48 [31–84]	41 [29–65]	35.5 [32–49]	57 [38–88]
ALP, U/L	106 [85–144]	105 [80–145]	90 [61.8–105.8]	145 [101.8–209.5]
Bilirubin, mg/dL	0.8 [0.6–1.4]	1.1 [0.7–1.9]	0.8 [0.6–1.0]	1.8 [0.9–4.6]
Creatinine, mg/dL	0.9 [0.7–1.1]	1 [0.7–1.2]	0.9 [0.8–1.1]	0.8 [0.6–1.0]
Albumin, g/dL	3.9 [3.3–4.3]	3.9 [3.1–4.5]	4.3 [3.5–4.5]	3.5 [2.9–4.1]
AFP, ng/mL	9 [4.1–50.6]	3.4 [2.2–5.5]	4.6 [3.0–23.9]	3.7 [2.4–5.9]

3.2. Biomarkers Associated With the TGF‐β Pathway Potentially Stratify HCC Risk in Cirrhotic Patients

Based on previous genomic and transcriptomic analyses of the TCGA‐LIHC dataset and animal model data, we have reported the critical role of the TGF‐β pathway in the progression of chronic liver disease to HCC [23, 24, 39]. A retrospective analysis of serum proteins from 216 cirrhosis patients revealed a significant TGF‐β pathway signature that potentially differentiated HCC from non‐HCC in patients with cirrhosis [33]. Therefore, we hypothesised that the TGF‐β pathway‐related serum protein signature could serve as a potential biomarker for risk stratification of HCC in cirrhosis. To test our hypothesis, we initially conducted an unsupervised clustering analysis of the 29 TGF‐β‐associated proteins using Somascan data from the prospective cirrhotic cohort B (n = 312) (Figure 1, Table S3). After normalisation, the patients were stratified into two main clusters based on their protein signature. Overall, among cirrhotic patients that developed HCC during follow‐up (n = 18), we observed that 72.2% of those (13 out of 18, z‐score > 0) had higher serum levels of Myostatin (MSTN) and 66.7% of them (12 out of 18, z‐score > 0) showed higher serum levels of Mothers against decapentaplegic homologue 3 (SMAD3) and Transforming Growth Factor Beta Receptor 2 (TGFBR2). Furthermore, cirrhotic patients with relatively higher serum levels of SMAD3, TGFBR2, and MSTN exhibited a five‐fold increased risk of developing HCC compared to those who had lower levels of these three proteins (9.68% [15 cases out of 155] vs. 1.91% [3 cases out of 157]).

FIGURE 1.

Longitudinal follow‐up of a TGF‐β pathway‐related biomarker panel that stratifies HCC risk in cirrhotic patients. Upper Panel: Unsupervised cluster analysis of the TGF‐β pathway‐associated proteins was performed in the prospective cirrhosis cohort (n = 312, Cohort B); a heatmap pattern of the TGF‐β pathway‐associated proteins abundance and associated clinical variables is shown. Lower Panel: Risk ratio of HCC development in cirrhotic patients with relatively low or high serum SMAD3/TGFBR2 signature in the prospective cohort (n = 312, Cohort B).

Consistent with our prior reports [33, 40], this clustering analysis suggests that TGFBR2 and MSTN could serve as potential diagnostic markers for HCC. Experimentally, ELISA assays validated increased expression of MSTN (FC: 1.5; p < 0.05), TGFBR2 (FC: 1.4; p < 0.05) and AFP (FC: 2.9, p < 0.05) in HCC+ patients (n = 57) compared to HCC− patients (n = 143) (Figure 2, Table S4). We also examined the expression of another four TGF‐β pathway‐related molecules (HMGB1, IL‐6, COL1A1, and SPTBN1), which are significantly elevated in liver tissue of patients with HCC and associated with poor prognosis [41, 42, 43, 44, 45, 46]. AFP, MSTN, and SPTBN1 were found to be elevated in HCCs associated with MASLD/ALD (Figure 3, p < 0.05). At the same time, serum HMGB1, IL‐6, COL1A1, and SPTBN1 were not significantly altered between overall HCC+ and HCC− (Figure 2). These findings underscore the potential clinical utility of TGF‐β pathway‐related molecules (MSTN, TGFBR2, and SPTBN1) as biomarkers for risk assessment and early detection of HCC, especially in those with MASLD/ALD.

FIGURE 2.

ELISA validation of TGF‐β pathway associated serum proteins in cirrhotic patients with any aetiology. Serum protein concentrations of MSTN, TGFBR2, AFP, HMGB1, IL‐6, COL1A1, and SPTBN1 were detected via ELISA. The boxes represent the interquartile range (IQR). The solid line in the box is the median, and the “+” sign is the mean. The bottom and top edges of the box are the lower and upper quartiles, which is known as the IQR. The whiskers range from the minimum to the lower quartile and from the upper quartile to the maximum. The extreme values (within 1.5 times the IQR from the upper or lower quartile) are the ends of the lines extending from the IQR. Points at a greater distance from the median than 1.5 times the IQR are plotted individually as circles. HCC−: Cirrhosis only (n = 143); HCC+: HCC with cirrhosis (n = 57). T‐test was used for p‐value calculation between HCC− and HCC+ groups. *p < 0.05; **p < 0.01.

FIGURE 3.

ELISA validation of TGF‐β pathway associated serum proteins in cirrhotic patients with MASLD/ALD (n = 137, 28 with HCC). Serum protein concentrations of MSTN, TGFBR2, AFP, HMGB1, IL‐6, COL1A1, and SPTBN1 were detected via ELISA. The boxplots were shown as in Figure 2. HCC−: Cirrhosis only (n = 109); HCC+: HCC with cirrhosis (n = 28). T‐test was used for p‐value calculation between HCC− and HCC+ groups. *p < 0.05; **p < 0.01.

3.3. Identification of Serum Proteins Differentially Associated With HCC in the Context of Cirrhosis

Using Somascan data of the matched cohort C (n = 328), after adjustment by the potential covariates, we identified 88 of 4728 proteins that had statistically different expression levels in HCC+ when compared to HCC− patients (with absolute fold change > 1.2 or < 0.8, and p < 0.05) via conditional logistic regression analysis: 46 increased and 42 decreased in HCC+ (Figure 4A, Table S5). The degree of difference in Somascan relative fluorescence unit (RFU) levels of 8 representative biomarkers is shown in Figure 4B. Compared to the HCC− group, HCC+ patients exhibited significantly higher protein levels of Glypican 3 (GPC3), Glutamine Synthetase (GLUL), Isocitrate Dehydrogenase 1 (IDH1), Alcohol Dehydrogenase 4 (ADH4), Integrin‐Linked Kinase (ILK), Enolase 2 (ENO2), Uridine Diphosphate Glucose Pyrophosphorylase 2 (UGP2), and Tumour Protein 63 (TP63). In Table 3, the fold change and multivariable odd ratios (95% confidential intervals [CIs]) per standard deviation for 5 of these proteins were 1.92 and 1.34 (95% CI: 0.70–2.58) for GPC3; 1.68 and 2.09 (95% CI: 0.86–5.09) for GLUL; 1.77 and 1.40 (95% CI: 0.91–2.18) for TP63; 1.31 and 1.26 (95% CI: 0.75–2.10) for ENO2; and 1.35 and 1.73 (95% CI: 0.71–4.20) for ILK.

FIGURE 4.

Differentially expressed proteins in HCC+ patients and HCC− patients in the matched cohort C (sample size: 109 pairs). (A) Volcano plot of the Somascan serum protein differential analysis between HCC+ (n = 109) and HCC− (n = 219) in the matched cohort C (Grey: Not significant; Green: Fold Change > 1.2 or < 0.8; Blue: P < 0.05; Red: Fold change > 1.2 or < 0.8, and p < 0.05). (B) Expression pattern of 8 representative proteins among the top dysregulated proteins (p < 0.05; absolute fold change > 1.2), using relative fluorescence units (RFU) derived from SomaScan. The boxplots were shown as in Figure 2. T‐test was used for p‐value calculation.

TABLE 3.

Top elevated SomaScan serum proteins comparing HCC+ and matched cirrhosis only patients (n = 328, Cohort C), according to fold change (absolute > 1.2) and conditional logistic regression (p < 0.05).

Entrez gene symbol	HCC+ mean, RFU	HCC− mean, RFU	Pooled SD, RFU	Fold change	p	Adjust p	OR (95% CI)	Multivariable OR (95% CI)
GPC3	2007	1046	1316	1.92	0.000	0.059	2.40 (1.62–3.55)	1.34 (0.70–2.58)
TP63	1353	764	1518	1.77	0.024	0.691	1.43 (1.05–1.95)	1.40 (0.91–2.18)
C3	7293	4249	8353	1.72	0.004	0.555	1.56 (1.16–2.10)	0.52 (0.23–1.16)
AFP	702	412	692	1.7	0.021	0.691	1.85 (1.10–3.12)	1.20 (0.70–2.06)
GLUL	8506	5049	6864	1.68	0.003	0.555	2.17 (1.30–3.65)	2.09 (0.86–5.09)
ADI1	2486	1556	3239	1.6	0.008	0.602	8.00 (1.71–37.49)	2.97 (0.30–29.27)
UGP2	1221	770	1116	1.59	0.016	0.666	2.13 (1.15–3.94)	1.56 (0.89–2.74)
DYNLL1	1057	670	1665	1.58	0.013	0.633	13.39 (1.73–103.89)	2.08 (0.23–18.65)
RNF128	2289	1495	2978	1.53	0.039	0.704	7.88 (1.11–55.96)	7.74 (0.57–105.40)
LIN7A	1482	986	2252	1.5	0.046	0.711	3.12 (1.02–9.57)	0.48 (0.05–4.90)
FCN2	19 289	12 895	21 471	1.5	0.009	0.602	1.46 (1.10–1.94)	1.22 (0.74–2.01)
ADH4	7206	4962	6149	1.45	0.009	0.602	1.43 (1.09–1.87)	1.36 (0.63–2.93)
C4A|C4B	25 567	17 767	19 634	1.44	0.002	0.555	1.51 (1.16–1.97)	1.34 (0.77–2.32)
HGS	8430	5900	6888	1.43	0.022	0.691	1.37 (1.05–1.79)	1.85 (0.57–6.04)
MX1	2084	1454	1779	1.43	0.004	0.555	1.48 (1.13–1.95)	1.28 (0.85–1.93)
SSMEM1	4008	2806	3323	1.43	0.045	0.711	1.51 (1.01–2.27)	0.81 (0.41–1.60)
PRPS2	1421	1038	879	1.37	0.008	0.602	1.65 (1.14–2.38)	1.30 (0.63–2.71)
ALPG	276	203	228	1.36	0.028	0.704	1.34 (1.03–1.74)	1.05 (0.60–1.82)
C3	12 223	9080	7292	1.35	0.006	0.555	1.44 (1.11–1.87)	1.73 (0.90–3.30)
ILK	3448	2554	2062	1.35	0.002	0.499	1.51 (1.16–1.95)	1.73 (0.71–4.20)
MPIG6B	28 750	21 710	20 800	1.32	0.019	0.679	1.34 (1.05–1.70)	1.18 (0.33–4.17)
COX6C	2679	2023	1668	1.32	0.005	0.555	1.92 (1.21–3.05)	1.06 (0.56–1.99)
ENO2	5135	3932	3137	1.31	0.005	0.555	1.47 (1.12–1.94)	1.26 (0.75–2.10)
UBL4A	3236	2498	2126	1.3	0.029	0.704	1.33 (1.03–1.71)	0.82 (0.24–2.84)
CKMT1A	31 553	24 371	28 729	1.29	0.023	0.691	1.34 (1.04–1.72)	2.28 (1.22–4.26)
ACAA1	8580	6632	7628	1.29	0.045	0.711	1.26 (1.00–1.58)	0.58 (0.28–1.18)
OTC	8611	6762	7422	1.27	0.041	0.709	1.27 (1.01–1.60)	0.86 (0.35–2.14)
PNMT	3732	2940	2299	1.27	0.013	0.633	1.46 (1.08–1.96)	1.07 (0.63–1.84)
IRAK4	521	411	370	1.27	0.046	0.711	1.42 (1.01–2.02)	1.33 (0.87–2.05)
CKB	1482	1166	1141	1.27	0.039	0.704	1.51 (1.02–2.23)	1.03 (0.68–1.56)
APLN	1035	823	570	1.26	0.015	0.658	1.75 (1.11–2.76)	0.79 (0.27–2.26)
VAMP3	1138	905	888	1.26	0.050	0.713	3.03 (1.00–9.18)	3.34 (0.72–15.52)
OSTN	485	389	330	1.25	0.047	0.711	1.33 (1.00–1.75)	1.27 (0.81–2.00)
ARPC1B	7982	6390	4720	1.25	0.046	0.711	1.29 (1.00–1.66)	0.49 (0.22–1.09)
ACY1	9953	8071	7073	1.23	0.020	0.688	1.33 (1.05–1.69)	0.70 (0.31–1.56)
RAB27B	7623	6232	4714	1.22	0.031	0.704	1.31 (1.03–1.67)	0.45 (0.14–1.48)
DNAJC4	2554	2090	1728	1.22	0.038	0.704	1.80 (1.03–3.14)	2.05 (0.40–10.45)
FERMT3	18 617	15 258	9701	1.22	0.017	0.672	1.36 (1.06–1.75)	0.90 (0.28–2.94)
SHMT2	9609	7852	6726	1.22	0.032	0.704	1.29 (1.02–1.62)	1.58 (0.78–3.18)
IDH1	12 792	10 588	5930	1.21	0.003	0.555	1.47 (1.14–1.90)	0.98 (0.47–2.04)
NT5C3A	654	538	240	1.21	0.004	0.555	1.66 (1.18–2.34)	1.10 (0.64–1.88)
TIMP1	38 181	31 683	36 719	1.21	0.032	0.704	1.37 (1.03–1.82)	0.97 (0.50–1.85)
SEMA3A	1363	1127	481	1.21	0.006	0.555	1.67 (1.16–2.41)	1.60 (0.94–2.70)
XPNPEP1	6480	5345	4437	1.21	0.049	0.713	1.34 (1.00–1.78)	1.11 (0.55–2.25)
ALPP	321	265	188	1.21	0.031	0.704	1.36 (1.03–1.81)	1.04 (0.58–1.86)
HERC4	4936	4081	2229	1.21	0.014	0.635	1.51 (1.09–2.09)	1.72 (0.53–5.58)

Note: p value (Wald test), odds ratios, and 95% confidence intervals were calculated from conditional logistic regression matched by age, sex, aetiology, BMI, Bilirubin, and AFP at blood collection. Multivariable model was adjusted by age, sex, aetiology, BMI, Bilirubin, and AFP.

Abbreviations: CI, confidence interval; OR, odds ratio.

3.4. Systems Biology Analysis of Dysregulated Serum Proteins Distinguishing HCC in Cirrhotic Patients

To further explore if altered serum protein profiles could reflect the biological mechanisms that mediate disease progression from cirrhosis to HCC, we performed an ingenuity pathway analysis (IPA) [35] using the 46 elevated proteins in HCC+ patients from cohort C as input (conditional logistic regression analysis, fold change > 1.2, p < 0.05). The dysregulated serum proteins were enriched in these top canonical pathways, among which post‐translational protein phosphorylation, regulation of IGF transport and uptake, noradrenaline and adrenaline degradation, creatine‐phosphate biosynthesis, and serine and glycine biosynthesis were most significantly affected (Figure S3A). IPA analysis predicted that Transforming Growth Factor Beta 1 (TGFB1), Hypoxia‐Inducible Factor 2 Alpha (HIF‐2α), Tumour Protein p53 (TP53), Forkhead Box O3 (FOXO3), and Kirsten Rat Sarcoma Viral Oncogene Homologue (KRAS) were the most statistically involved upstream regulators (Figure S3B). The top diseases and molecular functions associated with these elevated proteins were cancer and cell‐to‐cell signalling and interactions (Figure S3C,D). Previous studies have reported that mutations in potential driver genes such as Telomerase Reverse Transcriptase (TERT), TP53, and Beta 1 Catenin (CTNNB1) predispose individuals with cirrhosis to the development of HCC [47]. To investigate the potential functional relationship between proteins altered in our matched cohort C and these driver molecules in HCC, we conducted a network analysis using the top 46 increased proteins, two TGF‐β pathway‐related proteins (MSTN and TGFBR2), and three driver factors (TERT, TP53, and CTNNB1). This analysis identified seven proteins (TGFBR2, GPC3, GLUL, TP63, ENO2, ILK, IDH1) associated with HCC drivers and AFP (Figure S3E, Table S6), suggesting their potential value as functional biomarkers for HCC risk stratification or early diagnosis of HCC in patients with cirrhosis.

3.5. A Clinical and Serum Biomarker Panel With Improved Performance for HCC Detection in Cirrhotic Patients

To identify the optimal panel for HCC diagnosis, we constructed predictive models using five proteins from the above group (TP63, GPC3, GLUL, ENO2, ILK, fold change > 1.2, p < 0.05) together with COL18A1 (fold change < 0.8, p < 0.05, with the smallest q value: 0.064, Table S5), TGFBR2, MSTN and AFP (ELISA validated) in the matched cohort C (n = 328) via multivariable logistic regression. These models were further tested via 10‐fold cross‐validation across the entire cohort A.

In the matched cohort C, the multivariable logistic regression model using six proteins (TGFBR2, MSTN, AFP, COL18A1, GLUL, TP63) yielded an AUC of 0.84, compared with an AUC of 0.59 in the clinical basic model (Figure 5A). Adding these six proteins to four clinical variables (age, sex, BMI, and bilirubin) further enhanced the AUC to 0.87 (95% 0.83–0.91). By maximising the Youden J index, the model sensitivity was 0.88, and the specificity was 0.72 at the optimal predicted cutoff. The model's performance was then further validated in the entire cohort A (n = 477, including matched samples) using 10‐fold cross‐validation. The AUC statistic of the combined panel was 0.88 (95% CI, 0.84–0.91) across the entire cohort A (Figure 5B, Table S7), indicating significant discrimination.

FIGURE 5.

Receiver operating characteristic curves comparing the logistic regression models to distinguish HCC+ from cirrhosis only patients. (A, B) For the matched cohort C and the entire cohort A, basic model included age at blood collection, sex, body mass index ≥ 30 kg/m² (yes or no), and bilirubin. (C–F) Different subgroups from the entire cirrhotic cohort A (n = 477, including the matched cohort), filtering either by early‐stage HCC at T1 (C), or elevated BMI (≥ 30) (D), or with/without HBV/HCV infection (E, F). Protein model included six proteins (TGFBR2, MSTN, AFP, COL18A1, GLUL, TP63).

We further evaluated the performance of the six‐protein panel in different subgroups from the entire cohort A (n = 477), filtering by early‐stage HCC, low AFP (< 20 ng/mL), elevated BMI (≥ 30), or presence or absence of HBV/HCV infection. The panel performed equally well in these critical subgroups (AUCs: 0.862–0.921) (Figures 5C–F, S4 and S5, Table S8), particularly in obese (BMI > 30) cirrhotic cohorts (AUC of 0.90) and MASLD/ALD‐related cirrhotic cohorts (AUC of 0.92).

4. Discussion

In this multicenter study of serum SomaScan proteomics in cirrhotic patients, we identified 46 candidate HCC biomarkers associated with biologically relevant and plausible pathways, including those related to post‐translational protein phosphorylation, regulation of IGF transport and uptake, noradrenaline and adrenaline degradation, creatine phosphate biosynthesis, and serine and glycine biosynthesis. An unsupervised clustering analysis revealed that an elevated baseline signature of TGF‐β‐associated proteins (MSTN, SMAD3, and TGFBR2) correlated with a higher incidence of HCC in patients with cirrhosis. Experimental validation by ELISA confirmed that baseline levels of TGFBR2, MSTN, and AFP were significantly higher in HCC+ patients compared to HCC− patients. Incorporating six key proteins (i.e., TGFBR2, MSTN, AFP, COL18A1, GLUL, and TP63) into the prediction model improved the AUC from 0.59 (clinical basic model) to 0.87 (combined clinical basic and protein model). The performance of this panel was validated across the entire cohort A via 10‐fold cross‐validation, achieving a mean AUC of 0.86. Our findings may lead to significant improvements in HCC risk stratification for patients with cirrhosis in clinical practice.

The dysregulated proteins and pathways we identified could provide valuable markers through insights into the pathophysiology of HCC development from cirrhosis. For example, TGFBR2 is a transmembrane protein essential for TGF‐β signalling regulation, which is linked to the progression of liver cirrhosis and HCC. Studies have shown reduced TGFBR2 levels in liver tissue from HCC patients compared to those with cirrhosis, correlating significantly with aggressive HCC features [40, 48]. Myostatin (MSTN), a member of the TGF‐β superfamily, is primarily known for its role in regulating muscle growth, and has been observed to be significantly elevated in HCC compared to cirrhosis [33]. A retrospective study of 1077 patients with alcoholic liver cirrhosis found that elevated serum MSTN levels were significantly associated with an increased risk of developing HCC within 5 years [49]. COL18A1 is a component of the extracellular matrix that plays a crucial role in regulating glucose and lipid metabolism in the liver [50], reflecting further functional roles of the TGF‐β pathway that regulates collagen expression [45, 51]. Increased tumour burden was observed in COL18A1 knockout mice following DEN injections, indicating a potential tumour‐suppressive role for COL18A1 in HCC [52]. GLUL, also known as glutamine synthetase, plays a crucial role in nitrogen metabolism and the synthesis of glutamine, which is directly regulated by β‐catenin in the liver [53]. Aberrant activation of the WNT/β‐catenin signalling pathway promotes liver carcinogenesis, highlighting the clinical significance of GLUL overexpression in this pathological context [54]. TP63 is a member of the p53 family, which is potentially implicated in the HBV‐induced HCC development [55].

To mitigate potential multicollinearity in the logistic regression model, we took efforts to avoid it. We found that GPC3 is the most significantly increased protein in patients with HCC, regardless of stage or aetiological background. This finding provides further support for the diagnostic value of GPC3, which is abundant in liver cancer tissue but rarely detected in normal liver tissue [56]. Additionally, GPC3 is a target in multiple ongoing clinical trials (NCT05003895, NCT05103631, NCT03198546) that combine it with chimeric antigen receptor (CAR) T cells in HCC. Before these biomarkers can be incorporated into clinical practice, validation of their diagnostic usefulness in large, diverse, well‐characterised cohorts that span different demographics and aetiologies is essential to reduce the rate of false discoveries. Misdiagnosis of cancer can result in additional invasive testing, unnecessary economic loss, and patient distress. Thus, the specificity of the biomarker or biomarker panel is critical.

Strengths of this study include the use of multiple medical centers with diverse patient populations across multiple sites, and a hypothesis‐driven approach in cirrhotic patients as they progress to HCC. Limitations include a relatively small number of subjects with early‐stage HCC, which may affect our validation of probability thresholds. Independent validation of our protein panel in larger, prospective clinical studies will constitute an important test of our findings.

In conclusion, we have developed a mechanism‐based model that combines clinical and blood‐based biomarkers, potentially detecting any stage HCC in patients with cirrhosis to a degree that is significantly better than current clinical screening protocols. The extensive data obtained from this proteomic assay and multi‐layer framework, integrating heterogeneous functional biomarkers and clinical features, will provide a valuable resource for future investigations into HCC surveillance and detection biomarkers. Indeed, the minimal amount of serum sample volume required by the SOMAscan platform enables the analysis of archived samples from previous clinical trials, which can accelerate the time frame needed for further evaluation of our findings. This study thus presents exciting prospects for the early detection of HCC.

Author Contributions

R.L.A. and L.M. designed the study and provided the conceptual framework for the study. R.L.A. and X.X. performed statistical analysis and analysed data. X.X., R.L.A., and L.M. drafted the manuscript. X.X., K.S., H.Y., B.M., L.L.W., X.J.Z., S.K.S., J.M.C., P.S.L., S.H.H., S.D., A.R.H., H.H., R.L.A., and L.M. provided critical manuscript review for important intellectual content. B.M., N.N.D., L.L., F.C. and A.K.V. provided administrative, technical and material support.

Disclosure

R.L.A. reports stock ownership in Abbvie, Johnson & Johnson, Bristol Myers Squibb, and Pfizer.

Consent

The research was conducted in accordance with both the Declarations of Helsinki and Istanbul. Written informed consent was obtained from each patient. The study was approved by the Institutional Review Board at all participating institutions.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

FIGURE S1: Overall experimental design. Serum Somascan data from the matched cohort C (n = 328) was used to identify HCC‐related biomarker candidates and build predictive models distinguishing HCC+ from HCC− via multivariate logistic regression, which were 10‐fold cross validated in the entire cohort A (n = 477). TGF‐β associated protein signature was analysed in the prospective cirrhotic cohort (n = 312, Cohort B), while ELISA validations of key proteins (MSTN, TGFBR2, AFP, HMGB1, IL‐6, COL1A1, and SPTBN1) were conducted.

FIGURE S2: The flowchart illustrates the protein selection process.

FIGURE S3: Plausible pathophysiological pathways linked to HCC based on the 46 statistically dysregulated proteins. (A–D) Pathway and biology enrichment analysis via IPA (A: Top canonical pathway; B: Top upstream regulator; C: Disease; D: Molecular function). (E) Protein names are abbreviated according to the standard Entrez Gene symbol. Network and cluster analysis using the STRING database of functional and physical protein associations.

FIGURE S4: Box‐whisker plots for selected proteins in HCC+ patients (Early stage and AFP < 20 ng/mL) and HCC− patients in the entire cirrhotic cohort A (n = 477, 125 with HCC). The Kruskal Wallis test were used to compare the serum protein levels in cirrhosis and the indicated HCC groups.

FIGURE S5: Receiver operating characteristic curves comparing the logistic regression models to distinguish HCC+ from cirrhosis only patients in different subgroups from the entire cirrhotic cohort A (n = 477), filtering by low AFP (< 20 ng/mL). Basic model included body mass index ≥ 30 kg/m² (yes or no) and bilirubin. Protein model included six proteins (TGFBR2, MSTN, AFP, COL18A1, GLUL, TP63).

TABLE S1: Patient characteristics in the entire cirrhotic cohort A: including AFP clinical data and HCC stage.

TABLE S2: Somascan raw data in the whole cirrhotic cohort A: including matched (highlighted in light orange, cohort C) and unmatched patients (highlighted in light blue). HCC−: cirrhotic patients without HCC; HCC+: cirrhotic patients with HCC.

TABLE S3: Somascan raw data of TGF‐β related proteins in the prospective cirrhotic cohort B. A: without HCC progression (n = 294); B: with HCC developed during follow‐up (n = 18).

TABLE S4: ELISA assay of TGFBR2, MSTN, AFP, HMGB1, IL‐6, COL1A1, and SPTBN1.

TABLE S5: Conditional logistic regression analysis of the Somascan data in the matched cohort C.

TABLE S6:. String network analysis of the top elevated proteins.

TABLE S7: AUC values for 10‐fold cross‐validation of the panel in the entire cirrhotic cohort A (n = 477, including matched cohorts).

TABLE S8: Sensitivity for detecting HCC at a cut point that produces 90% specificity in different subgroups from the entire cohort A (n = 477, including matched cohort), filtering either by early‐stage HCC, or low AFP (< 20 ng/mL), or elevated BMI (≥ 30), or with/without HBV/HCV infection.

Xiang X., Shetty K., Yu H., et al., “Serum Proteomic Profile Based on the TGF‐β Pathway Stratifies Risk of Hepatocellular Carcinoma,” Liver International (2025): e70325, 10.1111/liv.70325.

Data Availability Statement

All proteomics data associated with this study are present in the paper or the Supporting Information Tables. Applications for clinical information of participants will be reviewed and provided upon request.