Plasma GPC3 reflects tumor GPC3 expression and predicts clinical outcomes in advanced HCC treated with atezolizumab + bevacizumab

Abstract

Background & Aims

The membrane proteoglycan glypican-3 (GPC3) is expressed in ∼70%–80% of hepatocellular carcinomas (HCCs), and its soluble fragment is released into the circulation by protease cleavage. We investigated correlations between tumor and plasma GPC3 levels and their associations with clinical outcomes in advanced HCC.

Methods

We analyzed 186 patients with advanced HCC treated between 2017 and 2023. Tumor GPC3 expression was assessed using immunohistochemistry and RNA sequencing, and plasma GPC3 by ELISA. Clinical outcomes were investigated in 106 patients receiving first-line atezolizumab plus bevacizumab.

Results

Plasma GPC3 was detectable in 59.3% of patients (median 11.4 pg/ml), and correlated with tumor expression (H-score: R = 0.40, p <0.001). Plasma GPC3 positivity (>0 pg/ml) increased across H-score strata (p <0.001) but remained undetectable in 25.9% with H-scores >200. High plasma GPC3 (>10 pg/ml) was associated with shorter progression-free survival (3.4 vs. 9.4 months, p = 0.003), overall survival (12.3 vs. 30.6 months, p <0.001) and lower objective response rates (17.9% vs. 47.9%, p = 0.001). Compared to tumor values, plasma GPC3 better predicted 6- and 12- month outcomes, and was independently associated with worse progression-free survival (hazard ratio 1.70, 95% CI 1.07–2.69, p = 0.02) and overall survival (1.96, 1.12–3.40, p = 0.02) on multivariable analyses.

Conclusions

While plasma GPC3 levels showed moderate concordance with tumor expression, meaningful discrepancies were also noted in a significant fraction of patients. High plasma GPC3 independently predicted worse survival outcomes in patients treated with atezolizumab plus bevacizumab, showing superior prognostic performance over tumor-based measures.

Impact and implications

Glypican-3 (GPC3) is a promising biomarker and therapeutic target in hepatocellular carcinoma (HCC), yet the relationship between plasma and tumor GPC3 levels remains unclear, and their clinical relevance has not been well defined, particularly in patients receiving atezolizumab plus bevacizumab. Our study provides clinical evidence supporting plasma GPC3 as a non-invasive biomarker in HCC, showing that circulating GPC3 levels are associated with tumor expression but more closely linked to clinical outcomes, serving as an independent predictor of survival and treatment response in patients treated with atezolizumab plus bevacizumab. These findings suggest that plasma-based assessment may complement tissue-based evaluation and support real-time risk stratification in advanced HCC, although further validation in diverse, prospective cohorts is warranted.

Graphical abstract

Highlights

• •Plasma GPC3 showed a significant correlation with tumor GPC3 and reflected aspects of tumor biology.
• •High plasma GPC3 was independently associated with worse outcomes in patients treated with Atezo/Bev.
• •Plasma GPC3 may serve as a non-invasive biomarker for real-time risk stratification in advanced HCC.

Introduction

Despite recent advances in systemic therapies, including immune checkpoint inhibitors and anti-angiogenic agents, the prognosis of advanced hepatocellular carcinoma (HCC) remains poor, with limited responses and a lack of validated molecular targets for precision medicine.[1], [2], [3]

Glypican-3 (GPC3) is a membrane-associated proteoglycan that is notably upregulated in approximately 70%–80% of HCCs while being barely detectable in normal liver tissue.[4], [5], [6], [7] This distinct expression pattern positions GPC3 as an attractive biomarker for HCC diagnosis and a potential therapeutic target. Accordingly, several investigational therapies – including monoclonal antibodies, bi- or multi-specific antibodies (such as 4-1BB engagers, CD3 engagers, and CD8-guided engagers), chimeric antigen receptor T cell therapies, and antibody-drug conjugates – are being evaluated in early-phase clinical trials, reinforcing the therapeutic potential of GPC3 in advanced HCC.[8], [9], [10], [11], [12]

Proteolytic cleavage of GPC3 releases its soluble fragment into the circulation, where it can be detected as plasma GPC3.^13,14 Plasma GPC3 has been investigated as a potential diagnostic and prognostic biomarker for HCC, which would add a minimally invasive, blood-based test to the clinical toolbox for assessing GPC3 levels and enable convenient and repeated measurements to facilitate disease monitoring.⁶ However, the relationship between plasma and tumor GPC3 measures remains unclear, and few studies have systematically evaluated relevant clinical implications, particularly in the context of cancer immunotherapies such as atezolizumab plus bevacizumab (Atezo/Bev).

In this study, we conducted a comprehensive analysis to assess the correlation between plasma and tumor tissue GPC3 levels and to evaluate the potential of plasma GPC3 as a surrogate marker for tumor tissue GPC3, focusing on its prognostic significance in patients with advanced HCC treated with Atezo/Bev immunotherapy.

Patients and methods

Study design and patient cohort

This study included 186 patients with advanced HCC treated at CHA Bundang Medical Center (Seongnam, Korea) between 2017 and 2023. From this population, plasma samples for 162 patients were available for GPC3 ELISA measurement, tissue samples for 150 patients were available for immunohistochemistry (IHC), and tissue samples from 99 patients were available for transcriptome profiling by RNA sequencing (RNA-seq; Fig. S1). Blood samples were collected immediately before the start of first-line systemic therapy.

Two analytical cohorts were defined in this study: (1) the correlation cohort (n = 166), comprising patients with available data for pairwise correlation analyses across the three GPC3 metrics (IHC, RNA-seq, plasma); and (2) the survival cohort (n = 106). Within the correlation cohort, 146 patients had both IHC and plasma data, 95 had plasma and RNA-seq data, and 87 had RNA-seq and IHC data. The survival cohort was a subset of the correlation cohort used to evaluate and compare the prognostic impact of GPC3 in patients treated with Atezo/Bev (from 2020 to 2023).

This study was approved by the Institutional Review Board of CHA Bundang Medical Center (IRB nos: 2017-11-052, 2017-11-054, and 2024-05-029) and conducted in accordance with institutional and ethical guidelines, including the 2013 Declaration of Helsinki and the 2018 Declaration of Istanbul. Written informed consent was obtained from all participants.

Immunohistochemical detection of GPC3 in HCC tissues

Formalin-fixed, paraffin-embedded tissue blocks identified as HCC were cut into 5 μm thick sections. IHC staining was performed using an anti-GPC3 antibody (mouse monoclonal, clone GC33, Roche), which recognizes a C-terminal epitope (residues 524–563) of GPC3. For visualization, a BenchMark ULTRA (Ventana Medical Systems, Inc., Oro Valley, AZ, USA) with heat-induced epitope retrieval (CC1 solution, Ventana) and the iView DAB detection kit (Ventana) was used.

GPC3 staining was evaluated semi-quantitatively with assessment of both the intensity and extent of staining. The intensity was scored on a scale of 0 to 3 (0, negative; 1, weak; 2, moderate; 3, strong), and the extent was estimated as the percentage of tumor cells stained (0% to 100%). The H-score was calculated by multiplying the intensity score by the extent of staining (range, 0–300).

For groupwise comparisons of GPC3 expression, the IHC H-score was categorized into four groups (0, 1–100, 101–200, 201–300) and IHC extent into three groups (0%–5%, 6%–50%, 51%–100%). Tissue GPC3 was considered detectable when the staining extent was >5%.

Quantification of plasma GPC3 by ELISA

Plasma GPC3 levels were quantified using a commercially available human GPC3 ELISA kit (R&D Systems, #DGLY30), according to the manufacturer’s protocol. This assay employs a monoclonal capture antibody and a polyclonal detection antibody in a sandwich format and primarily recognizes the C-terminal of GPC3. Briefly, plasma was separated from whole blood by dilution with PBS (1:1) and density gradient centrifugation using Ficoll (4 ml Ficoll layered with 5 ml diluted blood), followed by centrifugation at 400 × g for 30 min at room temperature. The plasma layer was collected and stored at −80 °C until analysis. For the ELISA, 100 μl of plasma sample was loaded into each well with assay diluent and incubated on a microplate shaker at 500 rpm for 2 h at room temperature. Following standard washing steps, GPC3 conjugate and substrate solution were sequentially added, and optical density was measured at 450 nm using Skanlt software v4.1 (Thermo Fisher Scientific) for the microplate reader. Plasma GPC3 positivity was defined as a detectable concentration above 0 pg/ml. Experiments were repeated twice independently, and the average of values was used for analysis.

Transcriptome profiling by 3’-QuantSeq RNA-seq

RNA-seq also was conducted using formalin-fixed, paraffin-embedded tumor tissue samples. Total RNA was extracted with Trizol reagent (Invitrogen, Carlsbad, CA, USA), and RNA quality was evaluated using the Agilent TapeStation 4000 system (Agilent Technologies, Amstelveen, The Netherlands). RNA concentration was determined using the ND-2000 Spectrophotometer (Thermo Fisher Scientific, Waltham, MA, USA), and RNA libraries were prepared using the QuantSeq 3’ mRNA-Seq Library Prep Kit (Lexogen, Vienna, Austria), following the manufacturer’s protocol. Sequencing was performed on the Illumina NextSeq 550 platform, generating 75-bp single-end reads. Raw reads were filtered for quality using BBDuk (average Q-score ≥20) and aligned to the hg19 reference genome with Bowtie2. Gene-level expression counts were obtained using bedtools. Normalization was performed using the trimmed mean of M values method implemented in the edgeR package (v4.4.1), and log₂-transformed counts per million (log₂CPM) values were used for downstream analyses, including correlation and survival analyses.

Survival analysis and prognostic modelling

Kaplan–Meier methods were used to estimate progression-free survival (PFS) and overall survival (OS), and group differences were evaluated in R using the log-rank test with the survival package and visualized using the survminer package. Optimal cut-off values for continuous variables, including GPC3 IHC extent, H-score, RNA-seq expression, and plasma concentration, were determined using the maxstat package in R based on log-rank statistics. For each variable, the cut-off value derived from PFS was used in all analyses.

To compare the prognostic accuracy of GPC3 expression measured by H-score, RNA-seq, and plasma levels, time-dependent receiver-operating characteristic (ROC) curves were generated at 6- and 12-month time points using the survivalROC package in R, applying the Kaplan–Meier method for censored survival data. The analysis used continuous GPC3 expression values as markers and reported AUC at each time point.

Univariable and multivariable Cox proportional hazards models were constructed to evaluate the association between clinical variables and survival outcomes. The following covariates were included in the models: age, sex, Eastern Cooperative Oncology Group performance status (ECOG PS), Barcelona Clinic Liver Cancer (BCLC) stage, portal vein tumor thrombosis (PVTT), extrahepatic spread (EHS), Child–Pugh score, alpha-fetoprotein (AFP), prior local therapy, and plasma GPC3 (dichotomized at 10 pg/ml using maxstat).

Statistical analysis

Pairwise correlations between GPC3 metrics were evaluated using Spearman’s rank correlation coefficient. For continuous variables, Jonckheere’s trend test was used to assess trends across ordered groups (boxplots), and the Wilcoxon rank-sum test was used for pairwise comparisons. The Cochran–Armitage trend test was applied to evaluate trends in categorical variables, including plasma GPC3 positivity rate (>0 pg/ml) across ordinal groups. Fisher’s exact test was used to compare overall response rates (ORRs) between groups. Two-sided p values <0.05 were considered statistically significant. All statistical analyses were performed using R (version 4.4.2).

Results

Baseline characteristics of patients with advanced HCC

Table 1 summarizes the population’s baseline characteristics. In the correlation cohort, the median age was 61.0 years (IQR 53.5–66.0), and 86.1% were male. According to BCLC staging, 15.7% were stage B and 84.3% stage C. PVTT and EHS were present in 27.7% and 68.7%, respectively. Most patients had preserved liver function, with Child–Pugh class A in 93.4% and class B in 6.6%. Baseline serum AFP levels were ≥400 ng/ml in 36.7% of patients. The most common etiology of HCC was HBV (71.1%), followed by alcohol (10.8%), metabolic dysfunction–associated steatotic liver disease (9.6%), and HCV (5.4%). Prior local treatment was given in 69.3% of patients, and 68.7% of patients received Atezo/Bev as first-line systemic therapy.

Table 1.

Baseline characteristics.

Characteristics	Correlation cohort (n = 166)	Survival cohort (n = 106)
Age, median (IQR)	61 (53.5-66.0)	61 (52.3-66.0)
Sex
Male	143 (86.1%)	91 (85.8%)
Female	23 (13.9%)	15 (14.2%)
ECOG performance status
0	66 (39.8%)	56 (52.8%)
1	100 (60.2%)	50 (47.2%)
BCLC stage
B	26 (15.7%)	14 (13.2%)
C	140 (84.3%)	92 (86.8%)
PVTT	46 (27.7%)	26 (24.5%)
Presence of extrahepatic spread	114 (68.7%)	77 (72.6%)
Child-Pugh class
A	155 (93.4%)	106 (100%)
B	11 (6.6%)	0 (0%)
AFP ≥400 ng/ml	61 (36.7%)	39 (36.8%)
Etiology of HCC
HBV	118 (71.1%)	74 (69.8%)
HCV	9 (5.4%)	5 (4.7%)
Alcohol	18 (10.8%)	15 (13.2%)
MASLD	16 (9.6%)	10 (9.4%)
Others	5 (3.0%)	3 (2.8%)
Prior local treatment
No	51 (30.7%)	30 (28.3%)
Yes	115 (69.3%)	76 (71.7%)
First-line systemic treatment (%)
Atezolizumab + bevacizumab	114 (68.7%)	106 (100%)
Sorafenib	19 (11.4%)	0 (0%)
Lenvatinib	11 (6.6%)	0 (0%)
Others	22 (13.3%)	0 (0%)

Quantitative variables are presented as median with IQR, and categorical variables as counts with percentages. AFP, alpha-fetoprotein; BCLC, Barcelona Clinic Liver Cancer; ECOG, Eastern Cooperative Oncology Group; HCC, hepatocellular carcinoma; MASLD, metabolic dysfunction–associated steatotic liver disease; PVTT, portal vein tumor thrombosis.

Baseline GPC3 levels in advanced HCC

Table 2 and Fig. S2 show the distribution of baseline GPC3 levels. In the correlation cohort, tumor samples from 126 of 150 patients (84.0%) had IHC-detectable GPC3 (extent >5%). Among these, IHC intensity scores were 0 in 12.7%, 1 in 13.3%, 2 in 27.1%, and 3 in 37.3%. The median IHC extent score was 80% (IQR, 20.0%–100.0%), and the median H-score was 160 (IQR, 30.0–300.0). RNA-seq–based GPC3 expression data were available for 99 patients, with a median log₂CPM of 10.1 (IQR, 8.3–11.6). Plasma samples from 96 of 162 patients (59.3%) had detectable GPC3, with a median concentration of 11.4 pg/ml (IQR, 0.0–46.5).

Table 2.

GPC3 levels by tumor IHC and RNA-seq and plasma ELISA.

Characteristics	Correlation cohort (n = 166)	Survival cohort (n = 106)
IHC intensity	n = 150	n = 93
0	21 (12.7%)	11 (10.4%)
1	22 (13.3%)	15 (14.2%)
2	45 (27.1%)	31 (29.2%)
3	62 (37.3%)	36 (34.0%)
IHC extent (%)	n = 150	n = 93
Median (IQR)	80 (20.0-100.0)	80 (20.0-100.0)
0–5%	24 (16.0%)	12 (12.9%)
6–50%	26 (17.3%)	16 (17.2%)
51–100%	100 (66.7%)	65 (69.9%)
IHC H-score	n = 150	n = 93
Median (IQR)	160 (30.0-300.0)	160 (40.0-300.0)
RNA-seq GPC3 (log2CPM)	n = 99	n = 70
Median (IQR)	10.1 (8.3-11.6)	10.0 (8.3-11.3)
Plasma GPC3 (pg/ml)	n = 162	n = 102
Median (IQR)	11.4 (0.0-46.5)	15.6 (0.0-48.8)
Positive (>0)	96 (59.3%)	64 (62.7%)
Negative (=0)	66 (40.7%)	38 (37.3%)

Quantitative variables are presented as median with IQR, and categorical variables as counts with percentages. CPM, counts per million; GPC3, glypican-3; IHC, immunohistochemistry; RNA-seq, RNA sequencing.

Comprehensive correlation analysis of plasma and tissue-based GPC3 in advanced HCC

The correlation between GPC3 levels was systematically analyzed through pairwise comparisons of IHC intensity, extent, H-score, RNA-seq expression, and plasma concentration (Fig. 1A). We first analyzed the correlation between tissue-based metrics (IHC vs. RNA-seq). IHC metrics showed strong correlations with RNA-seq GPC3 expression (R = 0.62 for intensity, 0.71 for extent, 0.75 for H-score; all p <0.001; Fig. 1A,B). When patients were stratified by H-scores (0, 1–100, 101–200, 201–300), RNA-seq GPC3 expression progressively increased across the strata (p <0.001; Fig. 1B).

Fig. 1.

Comprehensive correlation analysis of plasma and tissue-based GPC3 expression.

(A) Heatmap showing pairwise correlations among tissue-based metrics (IHC intensity, extent, and H-score), RNA-seq GPC3 expression (log₂CPM), and plasma GPC3 concentration. Levels of significance: ∗p <0.05; ∗∗p <0.01; ∗∗∗p <0.001 (Spearman’s rank correlation). (B) Scatter plot (top) showing the correlation between RNA-seq GPC3 expression and IHC H-score; boxplot (bottom) showing RNA-seq GPC3 expression stratified by H-score groups (0, 1–100, 101–200, 201–300). Levels of significance: R = 0.75, p <0.001 (Spearman’s rank correlation); p for trend <0.001 (Jonckheere’s trend test). (C) Scatter plots showing correlations between plasma GPC3 concentration and IHC metrics: intensity (left), extent (middle), and H-score (right). Levels of significance: R = 0.38, 0.39, and 0.40, respectively; all p <0.001 (Spearman’s rank correlation). (D) Boxplots showing plasma GPC3 concentration stratified by IHC metrics: intensity (0–3, left), extent (0%–5%, 6%–50%, 51%–100%; middle), and H-score (0, 1–100, 101–200, 201–300; right). Levels of significance: all p <0.001 (Jonckheere’s trend test). (E) Bar plots showing plasma GPC3 positivity (>0 pg/ml) by IHC metrics: intensity (0–3, left), extent (0%–5%, 6%–50%, 51%–100%; middle), and H-score (0, 1–100, 101–200, 201–300; right). Levels of significance: all p for trend <0.001 (Cochran–Armitage trend test). (F) Scatter plot (left) showing the correlation between RNA-seq GPC3 expression and plasma GPC3 concentration; boxplot (middle) showing plasma GPC3 concentration stratified by RNA-seq GPC3 expression quartiles (Q1–Q4); bar plot (right) showing plasma GPC3 positivity by RNA-seq quartiles. Levels of significance: R = 0.28, p = 0.006 (Spearman’s rank correlation); p for trend = 0.007 (Jonckheere’s trend test); p for trend <0.001 (Cochran–Armitage trend test). CPM, counts per million; GPC3, glypican-3; IHC, immunohistochemistry; RNA-seq, RNA sequencing.

Next, we examined the relationship between plasma GPC3 and tissue-based metrics. Plasma GPC3 levels correlated moderately with IHC metrics (R = 0.38 for intensity, 0.39 for extent, 0.40 for H-score; all p <0.001; Fig. 1A,C). In patient stratification by IHC metrics (intensity scores: 0, 1, 2, 3; extent: 0%–5%, 6%–50%, 51%–100%) and H-scores (0, 1–100, 101–200, 201–300), plasma GPC3 concentrations significantly increased with increasing strata (p <0.001 for all; Fig. 1D). Given that 40.7% of patients had undetectable plasma GPC3, we examined the positivity rate (>0 pg/ml) of plasma GPC3 across IHC groups (Fig. 1E) and found that it also increased with increasing strata (p <0.001 for all). However, even in the highest stratum of each IHC metric, 25.9%–30.2% of patients still had undetectable plasma GPC3 levels, while 28.6%–29.2% in the lowest stratum of each IHC metric had detectable plasma GPC3.

Plasma GPC3 levels correlated weakly with RNA-seq GPC3 expression (R = 0.28, p = 0.006; Fig. 1F), but when analyzed by RNA-seq expression quartiles, plasma GPC3 concentrations significantly increased across the groups (p = 0.007), a trend that was more evident in positivity rates (p <0.001).

Comparative evaluation of plasma and tissue-based GPC3 for clinical outcomes in patients treated with Atezo/Bev

We assessed the clinical relevance of baseline GPC3 levels measured by plasma, IHC, and RNA-seq in 106 patients from the survival cohort treated with first-line Atezo/Bev. Optimal cut-off values for each GPC3 metric were determined using the maximally selected rank statistic (maxstat) based on PFS. The resulting cut-offs were 70% for IHC extent, 80 for H-score, 7.9 (log₂CPM) for RNA-seq, and 10 pg/ml for plasma GPC3. These cut-offs were then used to stratify patients into high and low GPC3 groups for each metric and applied to evaluate differences in PFS, OS, and ORR. For IHC intensity scored on a 4-point scale (0 to 3), patients were stratified by low (0-1) and high (2-3) scores.

Median PFS was significantly shorter in the high compared to the low GPC3 group for most metrics (Fig. 2A–E), including IHC intensity (4.0 vs. 7.2 months, p = 0.02), IHC extent (3.9 vs. 7.2 months, p = 0.01), H-score (3.9 vs. 8.0 months, p = 0.007), and plasma GPC3 (3.4 vs. 9.4 months, p = 0.003). High log₂CPM (RNA-seq) showed a borderline association (4.0 vs. 16.5 months, p = 0.05). High vs. low plasma GPC3 levels were significantly associated with shorter median OS (12.3 vs. 30.6 months, p <0.001; Fig. 2A–E), as were high RNA-seq log₂CPM values (18.0 vs. 37.5 months, p = 0.04), while the significance for IHC extent (13.2 vs. 23.2 months, p = 0.05) was borderline. H-score and IHC intensity did not reach statistical significance for OS (p = 0.07 and p = 0.17, respectively).

Fig. 2.

Comparative evaluation of plasma and tissue-based GPC3 for clinical outcomes in patients treated with Atezo/Bev.

(A–E) Kaplan–Meier plots showing PFS (left) and OS (right) according to high vs. low GPC3 groups for each metric: (A) IHC intensity, (B) IHC extent, (C) H-score, (D) RNA-seq GPC3 expression, and (E) plasma GPC3 concentration. (F–J) Bar plots showing ORR stratified by high vs. low GPC3 groups for each metric (same order as A–E). Stacked bars indicate clinical response categories: CR, PR, SD, and PD. Group definitions were determined using the maximally selected rank statistic (details in Methods). Values of p for survival comparisons are from the log–rank test, and numbers at risk are indicated. Fisher’s exact test was used for ORR comparisons. Atezo/Bev, atezolizumab plus bevacizumab; CR, complete response; GPC3, glypican-3; IHC, immunohistochemistry; ORR, objective response rate; OS, overall survival; PD, progressive disease; PFS, progression-free survival; PR, partial response; RNA-seq, RNA sequencing; SD, stable disease.

Consistent with the survival outcomes, ORRs were numerically lower in the high compared to the low GPC3 groups across all metrics, but significantly so only for plasma GPC3 (17.9% vs. 47.9%, p = 0.001; Fig. 2F–J). Similar but non-significant trends were observed for IHC intensity (20.9% vs. 42.3%, p = 0.07), IHC extent (20.4% vs. 34.0%, p = 0.16), H-score (20.3% vs. 38.3%, p = 0.09), and RNA-seq (29.8% vs. 50.0%, p = 0.15).

For further comparison of the prognostic performance of H-score, RNA-seq, and plasma, we performed time-dependent ROC analyses at 6 and 12 months for both PFS and OS. Plasma GPC3 showed the highest AUCs for PFS at 6 months (0.68) and 12 months (0.67), followed by the H-score (0.61 at both timepoints) and RNA-seq (0.58 at 6 months, 0.53 at 12 months; Fig. S3A). For OS, plasma GPC3 also showed the highest AUCs at 6 months (0.66) and 12 months (0.70) compared to AUCs for the H-score (0.48 at 6 months, 0.58 at 12 months) and RNA-seq (0.48 at 6 months, 0.47 at 12 months; Fig. S3B). At the optimal cut-offs as determined by the Youden index, plasma GPC3 demonstrated relatively high sensitivity for predicting OS, with optimal cut-offs of 14.3 pg/ml (sensitivity 87%, specificity 45%) and 11.6 pg/ml (sensitivity 79%, specificity 39%) at 6 and 12 months, respectively.

Finally, in our analysis, several potential cut-offs for defining high plasma GPC3 levels were identified (e.g. 10.0 pg/ml from the PFS-based maxstat analysis and 11.6 pg/ml from the OS-based time-dependent ROC analysis; Fig. 2 and S3). To determine a robust and reproducible threshold, we performed a 5-fold stratified cross-validation of plasma GPC3–based survival analysis in the Atezo/Bev cohort (n = 102) (Table S1). In each training fold, the optimal cut-off was derived from PFS using the maxstat and subsequently applied to the corresponding test fold to evaluate both PFS and OS. Cross-validated cut-offs were consistent across folds (median 9.0 pg/ml, range 9.0–10.6), and high plasma GPC3 levels in the test folds consistently showed a trend toward worse outcomes (PFS: median HR 1.58, range 1.17–3.87; OS: median HR 2.49, range 1.17–9.62). Based on this reproducible range (9.0–10.6 pg/ml), we selected 10.0 pg/ml as the final cut-off for defining the high plasma GPC3 group in the subsequent analyses.

Association of plasma GPC3 levels with clinical characteristics

To investigate the association between plasma GPC3 and baseline characteristics, we evaluated its concentrations across key clinical factors (Table 3). Plasma GPC3 levels were compared according to age, sex, ECOG PS, BCLC stage, EHS, PVTT, Child–Pugh class, AFP, prior local therapy, and etiology. Among these variables, median [IQR] plasma GPC3 levels were significantly higher with vs. without PVTT (23.2 [4.8–56.7] vs. 0.4 [0.0–42.1] pg/ml; p = 0.002) and with high vs. low AFP levels (23.8 [0.0–64.8] vs. 0.0 [0.0–27.2] pg/ml; p <0.001). Additionally, patients without prior local therapy had higher plasma GPC3 levels than those who had received local therapy (20.4 [0.0–76.5] vs. 3.6 [0.0–36.4] pg/ml; p = 0.04). BCLC stage showed a borderline significant association with plasma GPC3 levels, with higher levels in stage C than in stage B (13.5 [0.0–51.8] vs. 0.0 [0.0–21.6] pg/ml; p = 0.06).

Table 3.

Comparison of plasma GPC3 levels according to clinical factors.

	Number of patients (n = 182)	Plasma GPC3 (median pg/ml, IQR)	p values
Age (years)			0.24
≤65	129	10.0 (0.0–48.0)
>65	53	0.0 (0.0–42.1)
Sex			0.34
Female	28	1.9 (0.0–29.8)
Male	154	11.4 (0.0–46.5)
ECOG PS			0.39
0	70	1.2 (0.0–47.7)
1	108	13.4 (0.0–44.6)
BCLC stage			0.06
B	27	0.0 (0.0–21.6)
C	151	13.5 (0.0–51.8)
PVTT			0.002
No	129	0.4 (0.0–42.1)
Yes	49	23.2 (4.8–56.7)
EHS			0.70
No	55	12.2 (0.0–46.0)
Yes	123	5.1 (0.0–47.5)
Child–Pugh class			0.12
A	167	5.1 (0.0–47.4)
B	11	22.0 (13.2–40.8)
AFP (≥400 ng/ml)			< 0.001
Low	113	0.0 (0.0–27.2)
High	65	23.8 (0.0–64.8)
Prior local therapy			0.04
No	51	20.4 (0.0–76.5)
Yes	127	3.6 (0.0–36.4)
Etiology			0.33
HBV	126	14.7 (0.0–53.0)
HCV	10	0.3 (0.0–14.2)
Alcohol	20	0.0 (0.0–15.9)
MASLD	18	1.5 (0.0–51.5)
Others	5	0.0 (0.0–217.6)

Comparison of plasma GPC3 levels across clinical factors. Wilcoxon rank-sum tests were used for group comparisons; for etiology, the Kruskal–Wallis test was applied. p <0.05 was considered statistically significant. Bold font indicates significant p values. AFP, alpha-fetoprotein; BCLC, Barcelona Clinic Liver Cancer; ECOG PS, Eastern Cooperative Oncology Group performance status; EHS, extrahepatic spread; GPC3, glypican-3; MASLD, metabolic dysfunction–associated steatotic liver disease; PVTT, portal vein tumor thrombosis.

Similarly, the H-score showed significant differences in association with PVTT and AFP (Table S2). Patients with PVTT had higher H-scores (200 [70–300] vs. 140 [20–300] without PVTT; p = 0.03), and those with high vs. low AFP had significantly higher H-scores (240 [130–300] vs. 120 [20–220]; p <0.001). For RNA-seq–based GPC3 expression, a significant difference was observed only by AFP level (Table S3), with higher expression in the high AFP group (10.6 [9.4–11.7] vs. 10.0 [8.1–11.4] with low AFP; p = 0.02).

Plasma GPC3 as an independent prognostic factor in Atezo/Bev–treated patients

To determine whether plasma GPC3 serves as an independent prognostic marker in patients treated with Atezo/Bev, we conducted univariable and multivariable analyses incorporating plasma GPC3 group and relevant clinical variables to obtain HRs with 95% CIs. As shown in Tables S4 and S5, univariable analysis indicated associations between shorter PFS and ECOG PS 1 (HR 1.84, 95% CI 1.20–2.82; p = 0.005), Child–Pugh score 6 (1.64, 1.05–2.56, p = 0.03), PVTT (1.78, 1.10–2.90; p = 0.02), high AFP (1.79, 1.16–2.76; p = 0.009), and high plasma GPC3 (1.90, 1.23–2.93; p = 0.004; Table S4) and between shorter OS and ECOG PS 1 (3.17, 1.93–5.21; p <0.001), Child–Pugh score 6 (2.00, 1.22–3.27, p = 0.006), PVTT (2.42, 1.42–4.13; p = 0.001), high AFP (1.71, 1.05–2.78; p = 0.03), and high plasma GPC3 (2.54, 1.51–4.26; p <0.001; Table S5). Prior local therapy (HR 0.54, 95% CI 0.32–0.90; p = 0.02) was significantly associated with improved OS.

In the multivariable analysis, with adjustment for significant factors from the univariable analysis, the following variables remained independently associated with shorter PFS: high AFP (HR 1.60, 95% CI 1.01–2.53; p = 0.04), and high plasma GPC3 (1.61, 1.02–2.54; p = 0.04; Fig. 3A, Table S4). For OS, ECOG PS 1 (2.88, 1.49–5.58; p = 0.002), and high plasma GPC3 (1.89, 1.10–3.24; p = 0.02) remained significant independent predictors (Fig. 3B, Table S5).

Fig. 3.

Multivariable analysis of plasma GPC3 and clinical factors in patients treated with Atezo/Bev.

(A) Forest plot of multivariable analysis for PFS. (B) Forest plot of multivariable analysis for OS. Statistical significance was assessed using Cox proportional hazards models, with p <0.05 considered significant. Variables that were significant in the univariable analysis were entered into the multivariable model, along with plasma GPC3 (high vs. low). Squares to the right of the vertical line indicate increased risk and squares to the left indicate decreased risk. Horizontal lines represent 95% CIs for HRs. Atezo/Bev, atezolizumab plus bevacizumab; GPC3, glypican-3; HR, hazard ratio; OS, overall survival; PFS, progression-free survival.

Discussion

This study presents a comprehensive cross-platform analysis of GPC3 expression using IHC, RNA-seq, and plasma ELISA in samples from patients with advanced HCC. Previous studies have predominantly focused on single-platform analyses in early-stage resectable HCC in the era of targeted therapies. In contrast, here we integrated tissue- and plasma-based data to assess their correlation and clinical relevance in advanced unresectable HCC treated with Atezo/Bev, reflecting the evolving landscape of immunotherapy. Our findings offer several insights with meaningful clinical implications.

First, we assessed GPC3 expression in tumor tissue using two methods (IHC and RNA-seq) and in plasma by ELISA and observed a moderate correlation among the measures. In the correlation cohort, IHC was positive for GPC3 in tumor samples from 84.0% of patients, with a median extent of 80%, indicating widespread expression. Plasma GPC3 was detectable in 59.3% of patients.

When patients were stratified by plasma detectability and H-score groups, the proportion of plasma-positive cases progressively increased with increasing H-score strata. However, 25.9% of patients with a high H-score (201–300) had undetectable plasma GPC3, and 28.6% of those with an H-score of 0 had detectable plasma GPC3. Thus, despite a degree of linear correlation between tissue expression and plasma levels, high tissue expression alone may not be sufficient for plasma detectability, implicating additional regulatory mechanisms. GPC3 is subject to various post-translational modifications, including furin-mediated cleavage or regulation by Notum and phospholipase D, which could influence the efficiency of GPC3 shedding.[15], [16], [17] Cases with high IHC expression but undetectable plasma GPC3 therefore could reflect insufficient shedding of GPC3 despite abundant tissue expression. Conversely, cases with no IHC expression but detectable plasma GPC3 might reflect spatial heterogeneity of GPC3 expression in advanced HCC. Unlike IHC, which was used to assess a single tumor region, plasma GPC3 may capture expression from unsampled sites such as intrahepatic multifocal lesions or distant metastases. This discordance also may stem from other factors such as variability in assay sensitivity, sampling bias, or differences in tumor vasculature; however, further studies are necessary to clarify these contributing mechanisms.

Second, compared to tissue-based measures, plasma GPC3 showed broader associations with advanced disease features. This finding is of special interest because tissue GPC3 expression has previously been linked to elevated AFP levels and adverse pathological characteristics such as larger tumor size, macrovascular invasion, and poor differentiation,^18,19 while the clinical relevance of plasma GPC3 has remained uncertain, with earlier studies yielding inconsistent or limited findings. In our study, plasma GPC3 levels were significantly elevated in patients with PVTT, high AFP, and no prior local therapy and showed a borderline association with advanced BCLC stage. These findings suggest that plasma GPC3 may better reflect advanced tumor biology than tissue GPC3 expression alone.

Among the clinical features evaluated here, elevated AFP levels were the only factor consistently associated with GPC3 across all three platforms. This concordance likely reflects the shared identity between AFP and GPC3 as oncofetal proteins that are co-expressed in a subset of biologically aggressive HCCs because of reactivation of fetal gene programs.^20,21 Some studies also have described etiology-related variation in plasma GPC3 expression, such as in HCV-related HCC;^22,23 we observed no such patterns, but the number of HCV-infected patients in our cohort was limited.

Last, plasma GPC3 consistently demonstrated prognostic and predictive value in patients treated with Atezo/Bev. Higher plasma levels were significantly associated with shorter PFS and OS, as well as with lower ORR, whereas tissue-based measures showed weaker or less consistent associations. Moreover, in time-dependent ROC analyses of the three methods, plasma GPC3 had the highest predictive performance for both PFS and OS at 6 and 12 months, with respective AUCs of 0.68 and 0.67 for PFS and 0.66 and 0.70 for OS. This better performance may partly reflect both spatial heterogeneity and the timing of sample collection. Tissue samples were collected at various time points prior to treatment initiation, but plasma GPC3 was measured contemporaneously with therapy and likely more accurately reflected real-time tumor burden and disease biology. Of note, at the optimal cut-off, plasma GPC3 exhibited high sensitivity but modest specificity, supporting its potential utility in identifying patients at higher risk of nearer-term mortality. Furthermore, plasma GPC3 remained an independent prognostic factor for both PFS and OS in multivariable analysis, underscoring its clinical relevance as a non-invasive biomarker reflecting systemic tumor burden and aggressiveness in advanced HCC under the treatment of Atezo/Bev.

This study has several limitations. Although blood samples were collected prospectively, the analysis was conducted retrospectively. Additionally, it was a single-center study conducted exclusively in patients of East Asian ancestry from an HBV endemic region, which may limit the generalizability of the findings. The cut-off value used to define the high plasma GPC3 group was determined using a data-driven approach and may be specific to this study cohort. While internal cross-validation supported the robustness of our prognostic results, validation in external cohorts is warranted. Because the study exclusively included patients treated with Atezo/Bev, it remains uncertain whether the observed association between plasma GPC3 and clinical outcomes is a predictive effect specific to immunotherapy or a general prognostic effect across treatment settings. Although both IHC and ELISA primarily target the C-terminal region of GPC3, minor differences in epitope recognition may have contributed to the modest tumor tissue–plasma discordance observed. Finally, the analysis did not account for molecular or spatial heterogeneity in GPC3 expression, including differences across histologic subtypes, subcellular localization patterns (e.g. membranous, cytoplasmic) or intratumoral heterogeneity. These features have been associated with distinct tumor behaviors and prognostic differences,[24], [25], [26] and may also influence tumor tissue-plasma concordance. Future studies incorporating these dimensions may provide a more comprehensive understanding of GPC3 heterogeneity and its clinical outcomes in advanced HCC.

In conclusion, our study demonstrates that plasma GPC3 levels showed a moderate alignment with tumor tissue expression, although meaningful discrepancies were also observed in a subset of patients. Importantly, high plasma GPC3 was independently associated with worse survival outcomes in patients with advanced HCC treated with Atezo/Bev, showing superior prognostic performance compared to tumor-based GPC3 measures. These findings suggest that plasma GPC3 may serve as a complementary biomarker that enables non-invasive, real-time risk stratification in the context of immune checkpoint inhibitor therapy.

Abbreviations

AFP, alpha-fetoprotein; Atezo/Bev, atezolizumab plus bevacizumab; BCLC, Barcelona Clinic Liver Cancer; ECOG PS, Eastern Cooperative Oncology Group performance status; EHS, extrahepatic spread; GPC3, glypican-3; HCC, hepatocellular carcinoma; HR, hazard ratio; IHC, immunohistochemistry; ORR, objective response rate; OS, overall survival; PFS, progression-free survival; PVTT, portal vein tumor thrombosis; RNA-seq: RNA sequencing.

Authors’ contributions

G.K., C.K., and H.J.C. designed the study. S.H., G.K., C.K., and H.J.C. acquired funding. B.K., J.S.K., G.K., H.K., and H.J.C. provided resources. G.K., H.K., C.K., B.K., and H.J.C. collected and curated the data, which were analyzed by D.J.S., S.H., W.S.L., H.Y., and S.J.K. and interpreted by all authors including S.H.L., J.H., H.Y.L., J.S., and Y.T. The article was written by D.J.S. and S.H. and was read and approved by all authors.

Data availability

De-identified patient data, including GPC3 expression profiles and relevant clinical information used in this study, will be made available to researchers upon reasonable written request to the corresponding author for scientific research purposes.

Financial support

This research was supported by a research grant from BeOne Medicines. Additional support was provided by the National Research Foundation of Korea (NRF), funded by the Korean government (MSIT) (NRF-2023R1A2C2004339) and by the Ministry of Education (NRF-2019R1A6A1A03032888). This study was also supported by the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (RS-2024-00438843).

Conflicts of interest

Hong Jae Chon consults for Eisai, Roche, Bayer, ONO, MSD, BMS, Celgene, Sanofi, Servier, AstraZeneca, SillaJen, Menarini, and GC Cell and has research grants from BeOne Medicines, Roche, Dong-A ST, and Boryung Pharmaceuticals. Chan Kim consults for Roche, ONO, MSD, BMS, Oncocross, and ViroCure and receives research funding from Boryung Pharmaceuticals, Oncocross, SillaJen, and ViroCure.

Please refer to the accompanying ICMJE disclosure forms for further details.

Supplementary data

The following are the Supplementary data to this article: