Baseline DCP May Modify Response to Atezolizumab‐Bevacizumab Versus Durvalumab‐Tremelimumab in Unresectable HCC

ABSTRACT

Background and Aims

For unresectable hepatocellular carcinoma (HCC), atezolizumab plus bevacizumab (Atez/Bev) and durvalumab plus tremelimumab (Dur/Tre) are widely used first‐line immunotherapies; however, biomarkers to guide regimen selection before treatment initiation remain undefined. We investigated whether baseline des‐γ‐carboxy prothrombin (DCP) modifies objective response rate (ORR) according to treatment regimen.

Methods

We analysed 1464 patients with unresectable HCC treated with first‐line systemic therapy between 2020 and 2025 across multiple centers. Patients were divided into training and validation cohorts. Restricted cubic spline‐based interaction models were used to evaluate the association between baseline DCP and ORR according to treatment regimen. The predicted absolute difference in ORR (ΔORR) between regimens was examined across continuous DCP values, with a prespecified 10% threshold considered clinically meaningful. Robustness was assessed using leave‐one‐center‐out and repeated random‐split sensitivity analyses.

Results

Overall ORR did not differ significantly between Atez/Bev and Dur/Tre. However, a significant interaction between baseline DCP and treatment regimen was observed in both cohorts. Atez/Bev showed a relatively stable ORR across DCP concentrations, whereas Dur/Tre showed a DCP‐dependent increase in ORR. At low DCP levels, Atez/Bev was favoured; with increasing DCP levels, the relative benefit shifted toward Dur/Tre. Sensitivity analyses confirmed a consistent DCP‐dependent transition pattern, although the DCP level at which ΔORR exceeded 10% varied across analyses, supporting a transition zone rather than a fixed cutoff.

Conclusions

Baseline DCP may modify the relative treatment benefit between Atez/Bev and Dur/Tre in unresectable HCC. Rather than defining a rigid cutoff, DCP appears to delineate a continuum in which higher levels increasingly favour Dur/Tre, providing a practical framework for individualized immunotherapy selection.

Key Points/Highlights

• Baseline DCP significantly modified the association between treatment regimen and objective response rate in unresectable HCC.

• Atezolizumab plus bevacizumab showed relatively stable objective response across the DCP spectrum.

• Durvalumab plus tremelimumab showed a DCP‐dependent increase in objective response.

• The relative treatment benefit shifted from Atez/Bev at low DCP levels toward Dur/Tre at higher DCP levels.

• Baseline DCP may serve as a practical biomarker for treatment‐effect heterogeneity between first‐line immunotherapy regimens.

Lay Summary

Baseline DCP may help doctors choose between two common first‐line immunotherapy combinations for patients with unresectable liver cancer. In this multicenter study, atezolizumab plus bevacizumab tended to be more favourable at lower DCP levels, whereas the relative benefit of durvalumab plus tremelimumab increased as DCP levels rose. Rather than acting as a strict cutoff, DCP appears to indicate a transition zone that may support more individualized treatment selection.

1. Introduction

Several combination immunotherapy regimens, including atezolizumab plus bevacizumab (Atez/Bev) and durvalumab plus tremelimumab (Dur/Tre), are currently available as first‐line treatment for patients with unresectable advanced hepatocellular carcinoma (HCC) [1]. Although these regimens have demonstrated clinical efficacy, the validated biomarkers for guiding optimal regimen selection before treatment initiation remain limited.

We previously reported the potential utility of a modified tumour marker (mTM) score combining alpha‐fetoprotein (AFP) and des‐γ‐carboxy prothrombin (DCP) to stratify treatment selection between Atez/Bev and Dur/Tre [2]. In that study, increasing DCP levels were associated with worse progression‐free survival (PFS) in patients treated with Atez/Bev, whereas the opposite trend toward improved PFS was observed in patients receiving Dur/Tre treatment, suggesting differential treatment effects according to baseline DCP. However, the mTM score could not predict objective response rate (ORR) for either regimen.

ORR is a clinically meaningful endpoint of combination immunotherapy. In an updated analysis of the HIMALAYA trial, patients who achieved an objective response to Dur/Tre treatment demonstrated prolonged treatment durability and favourable survival outcomes [3]. Similarly, in Atez/Bev‐treated patients, achieving an objective response may influence subsequent clinical decision‐making, including the feasibility of additional locoregional therapy or conversion strategies. Therefore, identifying patients who are likely to achieve an objective response before initiating treatment has clinical implications beyond those of conventional prognostic stratification.

DCP is a well‐established tumour marker associated with tumour angiogenesis and portal vein invasion in HCC [4, 5, 6, 7, 8]. However, whether DCP differentially modifies the ORR according to the treatment regimen, namely, whether a statistical interaction exists between baseline DCP and the immunotherapy regimen has not been adequately investigated.

In this study, we examined the interaction between baseline DCP and treatment regimen (Atez/Bev vs. Dur/Tre) with respect to the ORR. Furthermore, we explored the differences in the response patterns between the two regimens across the DCP spectrum.

2. Methods

We included patients with HCC who received systemic therapy between September 2020 and December 2025 at the participating institutions within a multicenter, collaborative research group. Patients enrolled in clinical trials; those with Barcelona Clinic Liver Cancer (BCLC) stage 0, A, or D disease; patients treated with durvalumab monotherapy; those with prior exposure to immune checkpoint inhibitors (ICI); and those without available baseline DCP measurements were excluded from the analysis. Patients who were positive for either anti‐HCV antibody or hepatitis B surface antigen (HBsAg) were classified as having viral liver disease, whereas those negative for both were classified as having non‐viral liver disease.

Because pathological information and imaging findings required for a formal diagnosis of cirrhosis were not uniformly available across participating centers, cirrhosis was not defined as an independent variable in this study. Instead, clinical features suggestive of cirrhosis were defined by the presence of at least one of the following: thrombocytopenia (< 100 000/μL), ascites, or hepatic encephalopathy.

The study cohort was divided into training and validation sets to assess whether the direction and pattern of the interaction between baseline DCP and the treatment regimen were reproducible across independent subsets. The training set consisted of patients treated at designated cancer care hospitals, whereas the validation set consisted of patients treated at university hospitals.

Because designated cancer care hospitals and university hospitals in Japan both function as referral‐based centers for advanced cancer care, this split was intended to assess the reproducibility of the interaction pattern across overlapping but non‐identical clinical settings, rather than to represent a strict external validation across fundamentally different healthcare systems.

2.1. Therapeutic Strategy for Hepatocellular Carcinoma

The treatment strategy for HCC was determined in accordance with the Japan Society of Hepatology guidelines [9] and finalized by the attending physician for shared decision‐making with the patient. The decision to initiate Atez/Bev or Dur/Tre treatment was made by the attending physicians at each participating institution. In this study, patients without prior TKI exposure corresponded to those who received Atez/Bev or Dur/Tre as first‐line systemic therapy, although prior locoregional treatments such as resection, ablation, or transarterial chemoembolization were not exclusion criteria.

Atez/Bev was administered as an intravenous injection of 1200 mg Atezolizumab and 15 mg/kg Bevacizumab every 3 weeks [10]. Dur/Tre was administered as a single dose of 300 mg Tremelimumab and 1500 mg Durvalumab on day 1, followed by 1500 mg Durvalumab every 4 weeks [11]. Treatment was discontinued in cases of unacceptable severe adverse events or clinically confirmed tumour progression.

2.2. Evaluation Methods of the Therapeutic Effects of Atez/Bev and Dur/Tre

The antitumor effects were assessed using dynamic computed tomography (CT) or magnetic resonance imaging (MRI) according to the Response Evaluation Criteria in Solid Tumours (RECIST version 1.1) [12] at each facility and classified as complete response (CR), partial response (PR), stable disease (SD), or progressive disease (PD). ORR was defined on the basis of the best overall response achieved during treatment. Radiological assessments were conducted every 6 to 8 weeks according to routine clinical practice at each participating institution. Central radiologic review was not performed.

Progression‐free survival (PFS) was defined as the time from the start of Atez/Bev or Dur/Tre therapy until disease progression or death from any cause, whichever occurred first. Overall survival (OS) was defined as the time from the start of Atez/Bev or Dur/Tre therapy until the death of the patient.

2.3. Statistical Analysis

Categorical variables were presented as numbers (percentages) and analysed using the chi‐squared test or Fisher's exact test. Continuous variables were expressed as medians (interquartile range [IQR]) and analysed using the Mann–Whitney U test. Survival curves were estimated using the Kaplan–Meier method and compared using the log‐rank test.

To evaluate differences in treatment effects between Atez/Bev and Dur/Tre, we performed an interaction analysis between treatment regimen and baseline DCP using a spline‐based model. The absolute difference in objective response rate (ΔORR) between Dur/Tre and Atez/Bev was prespecified as clinically meaningful when exceeding 10%. This threshold was defined as a practical benchmark for treatment selection in real‐world clinical practice and was informed by representative ORR values reported in pivotal first‐line trials for advanced HCC [11, 13]. Because these trials were not designed for direct head‐to‐head comparison, this threshold was not intended to define strict superiority between regimens, but rather to facilitate clinical interpretation of treatment‐effect heterogeneity across continuous DCP levels.

As a complementary explanatory analysis, we also performed conventional multivariable logistic regression analyses for ORR using the pooled cohort. This analysis was intended to assess whether the observed treatment‐by‐DCP interaction remained directionally consistent in a conventional parametric framework not relying on spline functions, rather than to establish a standalone predictive model. Age, sex, BCLC stage, aetiology, modified ALBI grade, log‐transformed AFP, treatment regimen, and log‐transformed DCP were included as covariates, with an additional treatment‐by‐logDCP interaction term in Model 2. Crude odds ratios were estimated using univariable logistic regression analyses. Model fit was compared using the Akaike information criterion (AIC) and Bayesian information criterion (BIC), calibration was assessed using the Hosmer‐Lemeshow goodness‐of‐fit test, and discrimination was evaluated using the area under the receiver operating characteristic curve (AUC).

To assess the robustness of the DCP‐treatment interaction against center‐specific effects, a leave‐one‐center‐out (LOCO) sensitivity analysis was performed. The spline‐based interaction model was repeatedly refitted after sequentially excluding one participating center at a time, and for each iteration, the likelihood ratio test (LRT) was used to evaluate the significance of the DCP × treatment interaction and to estimate the DCP value at which the predicted difference in ORR between regimens reached the prespecified threshold.

In addition, to evaluate the stability of the ΔORR transition under different data partitioning schemes, a repeated random‐split sensitivity analysis was conducted. The entire cohort was randomly divided into training subsets, and the spline‐based interaction model was repeatedly refitted using randomly resampled datasets. For each iteration, the DCP value at which ΔORR reached the prespecified threshold was estimated.

To adjust for imbalances in baseline characteristics between the Atez/Bev and Dur/Tre groups, we performed propensity score‐based weighting. The propensity score was estimated using a logistic regression model in which treatment assignment was the dependent variable and age, sex, BCLC stage, history of prior tyrosine kinase inhibitor (TKI) therapy, AFP (log‐transformed), and DCP (log‐transformed) were included as covariates. Considering the Atez/Bev group as the standard‐of‐care population, we applied average treatment effect in the controls (ATC) weighting, assigning a weight of 1 to patients treated with Atez/Bev and (1—PS)/PS to those treated with Dur/Tre. As a sensitivity analysis, we also performed average treatment effect in the overlap (ATO) weighting (Atez/Bev: PS; Dur/Tre: 1—PS). In the primary weighted analysis, ORR was used as the outcome and estimated using weighted regression models, and changes in the treatment‐effect difference across DCP levels were visualized using spline functions.

All reported p‐values were two‐sided, and statistical significance was defined as p < 0.05. Statistical analyses were performed using EZR version 1.68 (Jichi Medical University Saitama Medical Center, Saitama, Japan) [14].

3. Results

3.1. Overall Treatment Outcomes by Regimen

Between September 2020 and December 2025, 2001 patients with HCC received systemic therapy at the centers participating in this multicenter collaborative group. After applying the exclusion criteria, 1464 patients were included in the final analysis (Figure 1). The baseline characteristics of the Atez/Bev and Dur/Tre cohorts are summarized in Table S1.

FIGURE 1.

Patient flow.

The ORR was 30.5% in the Atez/Bev group and 26.3% in the Dur/Tre group, with no statistically significant difference between the regimens. In contrast, the disease control rate (DCR) was significantly higher with Atez/Bev (77.1%) compared to that with Dur/Tre treatment (62.4%) (Table S2).

The median PFS was significantly longer in the Atez/Bev group (6.9 months; 95% confidence interval [CI], 6.2–7.7) than in the Dur/Tre group (4.3 months; 95% CI, 3.3–5.8; p < 0.001). However, OS did not differ significantly between the regimens (Atez/Bev: 19.7 months; 95% CI, 17.8–21.4 vs. Dur/Tre: 18.7 months; 95% CI, 17.0–24.4; p = 0.359) (Figure S1).

3.2. Immune‐Related Adverse Events

Among the treatment‐related adverse events, immune‐related adverse events (irAEs) were observed more frequently in the Dur/Tre group than in the Atez/Bev group (Table S3). Similarly, grade ≥ 3 irAEs tended to occur more often in patients treated with Dur/Tre.

3.3. Patient Characteristics in the Training Set and Validation Set

The training and validation sets consisted of 722 and 742 patients, respectively.

The baseline characteristics of the training and validation cohorts are summarized in Table 1. Compared to that in the validation set, the training set included a higher proportion of male patients, an ECOG performance status of 0, a Child–Pugh score of 5, and BCLC stage B disease. In contrast, the frequency of extrahepatic metastasis (EHM) and prior treatment with TKIs was lower in the training set.

TABLE 1.

Patient Characteristics of the training and validation cohorts.

	Training set	Validation set	p
	n = 722	n = 742
Age, year*	74 [69, 80]	74 [68, 79]	0.038
Gender, n (%)			0.029
Male	599 (83.0)	582 (78.4)
Female	123 (17.0)	160 (21.6)
ECOG‐PS, n (%)			0.046
0	584 (80.9)	570 (76.8)
1	116 (16.1)	155 (20.9)
2	22 (3.0)	17 (2.3)
Liver disease, n (%)			0.117
Viral	334 (46.3)	374 (50.4)
Non‐viral	388 (53.7)	368 (49.6)
Child‐Pugh score, n (%)			< 0.001
5	436 (60.4)	377 (50.8)
6	202 (28.0)	238 (32.1)
≥ 7	84 (11.6)	127 (17.1)
ALBI score	−2.42 [−2.70, −2.07]	−2.36 [−2.70, −2.04]	0.148
modified ALBI grade, n (%)			0.262
1	256 (35.5)	240 (32.3)
2a	185 (25.6)	182 (24.5)
2b	272 (37.7)	304 (41.0)
3	9 (1.2)	16 (2.2)
Clinical features suggestive of cirrhosis, n (%)			0.715
Positive	173 (24.0)	184 (24.8)
Negative	549 (76.0)	558 (75.2)
EHM, n (%)	231 (32.0)	290 (39.1)	0.005
Intrahepatic tumour burden, n (%)			0.053
Up‐to‐7 In	122 (17.9)	144 (21.6)
Up‐to‐7 Out	396 (58.2)	339 (50.7)
MVI mild	57 (8.4)	64 (9.6)
MVI severe	105 (15.4)	121 (18.1)
BCLC stage, n (%)			0.010
B	294 (40.7)	253 (34.1)
C	428 (59.3)	489 (65.9)
Platelet, ×104/μL	14.3 [10.8, 19.6]	15.5 [11.5, 20.5]	0.051
LDH, U/L	204 [174, 240]	212 [179, 256]	0.012
CRP, mg/dL	0.32 [0.12, 0.98]	0.34 [0.12, 1.02]	0.817
AFP, ng/ml*	27.0 [5.0, 654.3]	43.0 [6.0, 822.3]	0.181
AFP‐L3, %	12.0 [1.0, 46.0]	14.0 [1.0, 53.8]	0.271
DCP, mAU/ml*	218.0 [41.0, 2771.0]	451.5 [51.3, 4599.0]	0.002
TKI history, n (%)	152 (21.1)	217 (29.2)	< 0.001
Treatment, n (%)			0.013
Atez/Bev	637 (88.2)	621 (83.7)
Dur/Tre	85 (11.8)	121 (16.3)
Observation period, months*	14.2 [7.2, 24.6]	12.3 [5.3, 21.2]	0.001

Abbreviations: Atez/Bev, Atezolizumab + Bevacizumab; Dur/Tre, Durvalumab + Tremelimumab; ECOG‐PS, Eastern Cooperative Oncology Group Performance Status; mALBI, Modified Albumin‐Bilirubin; MVI, Macrovascular Invasion; EHM, Extrahepatic Metastasis; BCLC, Barcelona Clinic Liver Cancer; AFP, Alpha‐Fetoprotein; TKI, Tyrosine kinase inhibitors; DCP, Des‐γ‐carboxy Prothrombin.

^*Median (interquartile range).

In addition, the median LDH and DCP levels were lower, and the follow‐up duration was longer in the training cohort. Baseline characteristics comparing the Atez/Bev and Dur/Tre groups within each cohort are shown in Table S4.

3.4. Interaction Between Baseline DCP and ORR According to Treatment Regimen in the Training Set

To clarify the relationship between the baseline DCP and ORR for each treatment regimen, a spline‐based interaction model was applied. A statistically significant interaction was observed between the baseline DCP and treatment regimen (deviance difference = 16.78; df = 3; p = 0.00078).

For Atez/Bev, the predicted probability of the ORR remained relatively stable across the entire range of DCP values and gradually decreased as the DCP increased. In contrast, for Dur/Tre, the predicted ORR increased progressively with increasing DCP levels and plateaued at higher DCP values (Figure 2a).

FIGURE 2.

Interaction between baseline DCP levels and treatment regimen on objective response rate (ORR). DCP values are shown on a log (1 + DCP) scale. The shaded areas represent 95% confidence intervals. (a) In the training cohort, a statistically significant interaction between DCP and treatment regimen was observed (deviance difference = 16.78; df = 3; p = 0.00078). The predicted ORR for Atez/Bev gradually decreased with increasing DCP concentration, whereas the ORR for Dur/Tre increased in a DCP‐dependent manner. (b) In the validation cohort, a similar interaction pattern was observed (deviance difference = 10.57, df = 3, p = 0.01433). ORR; DCP; Dur; Tre; Atez; Bev.

To further elucidate differences in treatment effect between regimens, the ΔORR between Atez/Bev and Dur/Tre was examined using the interaction spline model.

ΔORR changed according to baseline DCP levels: at low DCP levels, the predicted ORR favoured Atez/Bev; at intermediate levels, no substantial difference was observed; and at approximately 2000 mAU/mL, ΔORR exceeded the prespecified 10% threshold. Beyond this level, Dur/Tre demonstrated a higher predicted ORR (Figure S2a).

3.5. Interaction Between Baseline DCP and ORR According to Treatment Regimen in the Validation Set

A similar pattern of interaction was observed in the validation cohort. The interaction between the baseline DCP and treatment regimen remained statistically significant (deviance difference = 10.57; df = 3; p = 0.01433).

In the validation set, the ORR for Atez/Bev remained relatively stable across the DCP levels, whereas the ORR for Dur/Tre increased with increasing DCP values (Figure 2b).

The pattern of ΔORR was broadly similar to that observed in the training cohort; however, the DCP level at which ΔORR exceeded the 10% threshold was higher, occurring at approximately 13 000 mAU/mL (Figure S2b).

Although the specific DCP level at which ΔORR surpassed 10% differed between the training and validation cohorts, the overall interaction pattern and directional shift consistently supported a relative benefit of Dur/Tre at higher DCP levels.

3.6. Differential Treatment Response Patterns According to Baseline DCP

To explore the mechanism underlying DCP‐dependent changes in ORR, we examined the distribution of RECIST responses for each treatment regimen.

In the Atez/Bev group, increasing DCP levels were associated with a gradual decrease in ORR and SD, accompanied by a modest increase in PD. Overall, however, the response pattern remained relatively stable across the DCP levels.

In contrast, in the Dur/Tre group, ORR increased markedly with increasing DCP levels, primarily accompanied by a reduction in SD, whereas the proportion of PD did not increase (Figure 3).

FIGURE 3.

Predicted distribution of RECIST responses across baseline DCP levels according to treatment regimen. (a) In the Atez/Bev group, the proportion of ORR and SD gradually decreased with increasing DCP levels, accompanied by a modest increase in PD. (b) In the Dur/Tre group, the predicted probability of ORR increased progressively with rising DCP levels, primarily accompanied by a reduction in SD, while the proportion of PD remained relatively stable.

3.7. Conventional Logistic Regression Analysis

To further evaluate the interaction between baseline DCP and treatment regimen, we performed conventional multivariable logistic regression analyses for ORR using log‐transformed AFP and DCP values (Table 2a). In Model 1, which did not include the treatment‐by‐logDCP interaction term, logDCP was not significantly associated with ORR (OR 0.957, 95% CI 0.890–1.030, p = 0.234). In contrast, in Model 2, which included the interaction term, logDCP was significantly inversely associated with ORR (OR 0.923, 95% CI 0.875–0.974, p < 0.001), representing the DCP slope within the Atez/Bev group, while the interaction term between treatment regimen and logDCP was significantly positive (OR 1.360, 95% CI 1.190–1.540, p < 0.001). To facilitate interpretation of this effect modification, we additionally examined regimen‐specific logistic regression models (Table 2b). In the Atez/Bev group, increasing logDCP was independently associated with lower odds of ORR (OR 0.920, 95% CI 0.872–0.971, p = 0.003), whereas in the Dur/Tre group, increasing logDCP was independently associated with higher odds of ORR (OR 1.290, 95% CI 1.110–1.500, p < 0.001). These findings support the interpretation that baseline DCP acts as an effect modifier rather than a uniform predictor of ORR across regimens.

TABLE 2a.

A Conventional logistic regression models for objective response rate with and without the treatment‐by‐logDCP interaction term.

	Univariate analysis			Multivariate analysis model 1			Multivariate analysis model 2
	Crude OR	95% CI	p	Adjusted OR	95% CI	p	Adjusted OR	95% CI	p
Age	1.010	0.995–1.020	0.234	1.020	0.995–1.040	0.142	1.010	0.996–1.020	0.169
Gender (Female)	0.870	0.646–1.170	0.360	0.754	0.470–1.210	0.243	0.764	0.580–1.090	0.148
Aetiology (Nonviral)	0.930	0.739–1.170	0.539	0.839	0.596–1.180	0.315	0.931	0.734–1.180	0.555
BCLC stage (Stage C)	1.010	0.795–1.280	0.944	0.903	0.638–1.280	0.567	1.040	0.811–1.330	0.766
mALBI grade	0.869	0.763–0.989	0.034	0.939	0.774–1.140	0.527	0.882	0.770–1.010	0.068
Log AFP	0.993	0.957–1.030	0.712	1.030	0.970–1.100	0.331	1.010	0.970–1.060	0.559
Log DCP	0.962	0.921–1.000	0.081	0.957	0.890–1.030	0.234	0.923	0.875–0.974	< 0.001
Regimen (Dur/Tre)	0.831	0.593–1.160	0.282	0.668	0.383–1.170	0.156	0.110	0.041–0.297	< 0.001
Regimen×Log DCP	—	—	—	—	—	—	1.360	1.190–1.540	< 0.001
AIC	—	—	—	1838.3			1823.4
BIC	—	—	—	1870.2			1876.7
Hosmer‐Lemeshow p‐value	—	—	—	—		0.758	—		0.969
AUC				0.557			0.592

Note: Hosmer‐Lemeshow p‐values are shown as measures of model calibration.

Abbreviations: AFP, Alpha‐Fetoprotein; AIC, Akaike information criterion; BCLC, Barcelona Clinic Liver Cancer; BIC, Bayesian information criterion; DCP, Des‐γ‐carboxy Prothrombin; Dur/Tre, Durvalumab + Tremelimumab; mALBI, Modified Albumin‐Bilirubin; TKI, Tyrosine kinase inhibitors.

TABLE 2b.

Regimen‐specific conventional logistic regression models for ORR.

	Atez/Bev			Dur/Tre
	OR	95% CI	p	OR	95% CI	p
Age	1.010	0.996–1.030	0.142	1.000	0.959–1.040	0.968
Gender (Female)	0.842	0.608–1.170	0.299	0.362	0.103–1.280	0.115
Aetiology (Nonviral)	0.918	0.712–1.180	0.507	1.020	0.505–2.060	0.956
BCLC stage (Stage C)	0.924	0.711–1.200	0.552	2.980	1.290–6.910	0.011
mALBI grade	0.917	0.794–1.060	0.243	0.673	0.457–0.993	0.046
Log AFP	1.020	0.978–1.070	0.309	0.957	0.845–1.080	0.491
Log DCP	0.920	0.872–0.971	0.003	1.290	1.110–1.500	< 0.001

Abbreviations: AFP, Alpha‐Fetoprotein; BCLC, Barcelona Clinic Liver Cancer; DCP, Des‐γ‐carboxy Prothrombin; Dur/Tre, Durvalumab + Tremelimumab; mALBI, Modified Albumin‐Bilirubin; TKI, Tyrosine kinase inhibitors.

3.8. Robustness of the DCP‐Dependent Transition Pattern

To confirm that the observed interaction between DCP and treatment regimen was not driven by a specific center, we performed a LOCO sensitivity analysis. After sequentially excluding each of the 31 participating centers, a statistically significant interaction between baseline DCP and treatment regimen (Atez/Bev vs. Dur/Tre) was consistently observed in all LOCO iterations (p < 0.001 for all iterations).

We further estimated, in each LOCO analysis, the DCP value at which the predicted ΔORR reached the prespecified clinically meaningful threshold of 10% using the spline‐based interaction model. Although some variability was observed across iterations, the estimated DCP values were consistently distributed within the high‐DCP range (median 11 350 mAU/mL [IQR 10762–12 149], range 4775–26 914 mAU/mL) (Figure S3a).

In addition, to evaluate the effect of data partitioning, we conducted a repeated random‐split sensitivity analysis. Similar to that in the LOCO analysis, the DCP values at which ΔORR reached 10% were consistently located in the high‐DCP range (median 11 530 mAU/mL [IQR 8167–16 300], range 2533–48 493 mAU/mL). The minimum estimated value was approximately 2500 mAU/mL, suggesting that the relative benefit of Dur/Tre may begin to emerge at DCP levels in the several‐thousand range (Figure S3b).

3.9. Sensitivity Analysis in TKI‐Naïve Patients

The use of TKIs reportedly influences DCP levels [15]. Therefore, we examined the association between baseline DCP and ΔORR in patients without prior TKI exposure. In this study, the TKI‐naïve subgroup corresponded to patients who received Atez/Bev or Dur/Tre as first‐line systemic therapy. In this sensitivity analysis restricted to TKI‐naïve patients, a similar pattern of interaction between baseline DCP and treatment regimen was observed (Figure S4). At low DCP levels, ΔORR was negative, indicating a higher predicted ORR with Atez/Bev. As DCP levels increased, ΔORR gradually increased, and the difference between the two regimens diminished at intermediate levels. In the high‐DCP range, ΔORR became positive, and Dur/Tre demonstrated a higher predicted ORR compared to that with Atez/Bev, particularly at DCP levels exceeding several thousand mAU/mL.

3.10. Sensitivity Analysis Including Patients With BCLC Stage A Disease

As an additional sensitivity analysis, we repeated the spline‐based interaction analysis after including patients with BCLC stage A disease. The interaction between baseline DCP and treatment regimen remained significant (deviance difference = 22.83; df = 4; p = 0.00014), and the overall directional pattern was preserved, with Atez/Bev favoured in the low‐DCP range and the relative benefit shifting toward Dur/Tre at higher DCP levels (Figure S5).

3.11. Assessment of Internal Validity Using Propensity Score Weighting (ATC and ATO)

To address the possibility that the DCP‐dependent differences in treatment response observed in the unadjusted analyses were influenced by an imbalance in baseline characteristics, we examined the internal validity using propensity score‐based weighting. Covariate balance after weighting was evaluated using standardized mean differences (SMDs); the maximum absolute SMD was 0.0342 for ATC weighting and 2.1 × 10⁻⁸ for ATO weighting (Table S5).

After ATC weighting, the ΔORR trajectory across the DCP spectrum showed a directionally similar pattern to the unadjusted analysis. Atez/Bev tended to be favoured in the low‐DCP range; the differences were attenuated in the intermediate range, and the point estimates suggested a shift toward Dur/Tre in the high‐DCP range (Figure 4a). Point estimates in the binned analysis also suggested Dur/Tre favorability in the high‐DCP strata (Figure S6a). However, the 95% CIs widened at the upper end of the DCP, indicating limited precision in definitively concluding the superiority of Dur/Tre in that region.

FIGURE 4.

Propensity score‐weighted ΔORR (Dur/Tre − Atez/Bev) across baseline DCP. Restricted cubic spline interaction models were used to estimate the ΔORR between Durvalumab plus Tremelimumab (Dur/Tre) and Atezolizumab plus Bevacizumab (Atez/Bev) across the continuous baseline DCP spectrum. The solid line shows the point estimate and the shaded band indicates the 95% confidence interval. The baseline DCP was displayed on a log scale (1st–99th percentile range). Rug plots along the x‐axis show the distribution of baseline DCP in the analysed cohort. The horizontal dashed line denotes ΔORR = 0 (values > 0 favour Dur/Tre; values < 0 favour Atez/Bev). (a) ATC‐weighted estimates. (b) ATO‐weighted estimates.

In the sensitivity analysis using ATO weighting, a similar pattern was observed, with a tendency toward Atez/Bev favorability in the low‐DCP range (Figure 4b). In the high‐DCP range, ΔORR point estimates were positive; however, the CIs crossed zero, again suggesting limited precision in that region (Figure S6b).

3.12. Baseline Characteristics According to a Reference DCP Threshold

To explore potential residual confounding around the transition region, we examined baseline characteristics according to a reference DCP threshold of 2000 mAU/mL, stratified by treatment regimen (Table S6). In both the Atez/Bev and Dur/Tre groups, patients with DCP ≥ 2000 mAU/mL had poorer liver function, greater intrahepatic tumour burden, more advanced BCLC stage, and higher inflammatory and tumour‐marker profiles than those with DCP < 2000 mAU/mL.

3.13. Landmark Overall Survival According to Response at the First Radiological Assessment

A supportive landmark analysis based on the first radiological assessment also suggested that early tumour control was associated with subsequent overall survival (OS) in both the Atez/Bev and Dur/Tre groups (Figure S7).

In the Atez/Bev group, median OS was 32.2 months in patients who achieved ORR at the first radiological assessment versus 16.1 months in those who did not (p < 0.001), and the corresponding values in the Dur/Tre group were 24.3 and 16.7 months, respectively (p = 0.025).

4. Discussion

Three principal findings emerged from this study. First, DCP may function as a biomarker that contributes to the treatment selection between Atez/Bev and Dur/Tre. Second, whereas Atez/Bev demonstrated relatively stable therapeutic efficacy regardless of DCP levels, Dur/Tre exhibited DCP‐dependent efficacy, with the potential to increase the relative treatment benefit as DCP levels increased. Third, among patients treated with Dur/Tre, high DCP levels were associated with a tendency for cases that might otherwise remain as stable disease (SD) to shift toward an objective response.

Multiple combination immunotherapies are currently available for the treatment of unresectable HCC. Regimens that include anti–CTLA‐4 antibodies have achieved durable responses in a subset of responders [3, 11]. However, these regimens are associated with a higher frequency of irAEs, and only a proportion of patients achieve objective responses. Therefore, identifying patients who are more likely to benefit from treatment is clinically important.

In both the training and validation cohorts, we observed a consistent treatment–biomarker interaction between baseline DCP levels and ORR across regimens. Specifically, the predicted ORR with Atez/Bev remained relatively stable across the DCP spectrum, showing only a modest decline as DCP increased, whereas the predicted ORR with Dur/Tre tended to increase in a DCP‐dependent manner.

Consequently, the relative predicted benefit shifted from Atez/Bev at lower DCP levels to Dur/Tre at higher DCP levels. In the training set, the predicted ΔORR exceeded the prespecified 10% threshold at approximately 2000 mAU/mL, whereas in the validation set, this transition was observed at higher DCP levels.

This pattern was broadly reproduced in LOCO and repeated random split sensitivity analyses. However, the DCP value at which ΔORR reached 10% varied across analyses. The difference in the apparent DCP threshold between the training and validation cohorts may partly reflect differences in baseline patient selection across regimens. In particular, the Dur/Tre group in the validation cohort had higher baseline DCP levels, shorter follow‐up, and more unfavourable liver function–related characteristics compared with the Atez/Bev group, whereas the between‐regimen imbalance in baseline DCP was less pronounced in the training cohort. These differences may have influenced the observed transition point. Because the training and validation cohorts were defined according to hospital type rather than random allocation, this comparison should be interpreted as testing the reproducibility of the interaction pattern across distinct clinical settings, rather than as strict independent validation. Overall, Atez/Bev was associated with a relatively high predicted ORR in the low‐DCP range, whereas the relative benefit of Dur/Tre increased stepwise with increasing DCP. Therefore, these findings support the presence of a DCP‐dependent transition zone rather than a fixed cutoff.

This trend remained qualitatively similar after balancing the baseline covariates using ATC and ATO weighting. However, estimates in the high‐DCP range were less stable owing to the limited sample size, and the precision was insufficient to conclusively assert the superiority of Dur/Tre in that region.

Baseline characteristics stratified by a reference DCP threshold of 2000 mAU/mL showed that the high‐DCP subgroup represented a clinically more advanced population in both treatment groups, with poorer liver function, greater tumour burden, and higher inflammatory and tumour‐marker levels. These findings suggest that the relative benefit of Dur/Tre observed in the high‐DCP range was not simply driven by a favourable‐risk subgroup, although residual confounding cannot be completely excluded.

From a biological perspective, DCP is associated with tumour angiogenesis and portal vein invasion [4, 5, 6, 7, 8]. Elevated DCP levels may therefore reflect a biologically aggressive tumour phenotype characterized by enhanced vascular invasion, increased tumour burden, and a higher likelihood of intrahepatic spread [15, 16, 17, 18]. At the same time, such tumours may also harbour an immune milieu enriched with tumour‐associated antigens, in which CTLA‐4–mediated priming could be relatively more effective. The response pattern observed in the present study is broadly consistent with this interpretation. In the Atez/Bev group, increasing DCP levels were associated with gradual decreases in ORR and SD, accompanied by a modest increase in PD. In contrast, in the Dur/Tre group, increasing DCP levels were associated with a marked increase in ORR and a reduction in SD, without a substantial increase in PD. This pattern may suggest that, in a subset of high‐DCP tumours, the addition of anti–CTLA‐4 therapy facilitates conversion from disease stabilization to objective response, rather than simply reducing the probability of progressive disease.

However, this biological interpretation should not be regarded as exclusive. Because this was a retrospective observational study, the apparent DCP‐dependent benefit of Dur/Tre may also have been influenced by residual confounding, differences in patient selection across regimens, local radiologic assessment without central review, and shorter follow‐up in the Dur/Tre cohort. Therefore, although the present findings are supported by the observed RECIST response distribution, they remain hypothesis generating rather than mechanistically definitive. Prospective translational studies will be needed to clarify the biological basis of this interaction.

This study had some limitations. First, although this was a large multicenter study, it was retrospective, and residual confounding could not be excluded. Because treatment selection was not randomized, unmeasured factors may have influenced the choice between Atez/Bev and Dur/Tre. Moreover, as the two regimens were introduced at different time points in clinical practice, temporal bias cannot be fully excluded. In addition, radiological response assessment was performed at each participating institution as part of routine clinical practice, without central radiologic review or formal assessment of interobserver agreement. Because complete information on imaging modality at the individual assessment level was not uniformly retrievable across all centers, and radiologists were not centrally blinded to treatment assignment or biomarker changes, classification bias in ORR assessment cannot be excluded. Second, the number of patients treated with Dur/Tre was limited, potentially affecting the statistical power of this study. Third, the depth of tumour response was not evaluated. Previous reports have suggested an association between the degree of tumour shrinkage and treatment duration in Dur/Tre‐treated patients, but this was not assessed in the present study. Fourth, the study population exclusively consisted of Japanese patients; therefore, external validation in Western populations is required. Finally, the biological mechanisms underlying the observed interaction were not directly investigated, and our findings should be regarded as hypothesis generating. Prospective studies involving molecular analyses are required to confirm these observations.

5. Conclusion

In this multicenter cohort study, baseline DCP was associated with differential objective response patterns between the Atez/Bev and Dur/Tre regimens in patients with unresectable HCC. Atez/Bev showed relatively stable efficacy across the DCP spectrum, whereas the relative benefit of Dur/Tre increased with increasing DCP levels. These findings suggest that baseline DCP levels may serve as practical biomarkers for treatment‐effect heterogeneity between first‐line immunotherapy regimens. However, because the observed interaction was continuous rather than defined by a robust single cutoff, DCP should be interpreted as indicating a transition zone rather than a fixed threshold for treatment selection. Given the retrospective design and limited precision in the high‐DCP range, these results should be considered hypothesis generating and warrant prospective validation.

Author Contributions

K.T. and K.T. contributed to the concept, design, and execution of the study. K.T., K.T., A.H., T.T., M.H., K.K., J.T., M.A., K.T., E.I., S.F., K.T., T.I., K.T., H.T., H.T., C.O., T.N., T.H., S.K., K.K., A.N., H.K., T.M., Y.U., H.K., Y.Y., H.N., M.I., T.A., H.O., J.O., H.T., J.O., S.N., Y.N., O.Y., K.N., A.M., N.I., T.O., T.A., A.T., T.N., F.T., H.O., Y.K., Y.K., H.N., S.K., T.F., T.M., S.K., H.E., K.M., M.K., Y.H., M.K., and T.K. contributed to data curation. K.T. conducted the statistical analyses and interpreted the data. K.T. wrote the first draft of the manuscript. K.T. and K.T. reviewed and edited the manuscript. All authors have read and approved the final version of the manuscript.

Ethics Statement

Following official approval, this study was conducted as a retrospective analysis of database records in compliance with the Guidelines for Clinical Research of the Ministry of Health and Welfare of Japan. All the procedures adhered to the principles of the Declaration of Helsinki. Atez/Bev or Dur/Tre was administered after obtaining written informed consent from each patient. Informed consent for participation in the analysis was obtained from each patient via an opt‐out procedure for those who commenced treatment prior to the approval of the Clinical Research Ethics Committee, whilst written informed consent was obtained from patients who were introduced to treatment after approval.

Conflicts of Interest

Atsushi Hiraoka received lecture fees from Ono/Bristol Myers Squibb Co. Ltd., AstraZeneca K.K., and Chugai Pharmaceutical Co. Ltd. Toshifumi Tada received lecture fees from AstraZeneca K.K., Chugai Pharmaceutical Co. Ltd., Eisai Co., and AbbVie Inc. Joji Tani received lecture fees from AstraZeneca K.K., Hidenori Toyoda received lecture fees and research funding from Eisai, Chugai Pharmaceutical Co. Ltd., Takeda, Terumo, AbbVie, Gilead, Fujifilm WAKO, and Abbott. Satoru Kakizaki received lecture fees from AbbVie Inc. Yoshihide Ueda received lecture fees from AbbVie Inc. and Chugai Pharmaceutical Co., Kazuhito Kawata received honoraria from AbbVie Inc. Atsushi Naganuma received research funding from AstraZeneca; and lecture fees from Bristol Myers Squibb. Shohei Komatsu received lecture fees and research funding from Chugai Pharmaceutical Co., and AstraZeneca K.K., Masatoshi Kudo received lecture fees from Chugai Pharmaceutical Co. Ltd., Eisai Co. Ltd., AstraZeneca K.K., Otsuka Pharmaceutical Co. Ltd., and TAIHO PHARMACEUTICAL Co. Ltd., and received research funding from Chugai Pharmaceutical Co. Ltd. and GE Healthcare Japan Corporation.

Supporting information

Figure S1: Overall PFS (a) and OS (b).

Figure S2: Predicted difference in objective response rate (ΔORR) between Durvalumab plus Tremelimumab and Atezolizumab plus Bevacizumab according to baseline DCP levels in the training and validation cohorts.

In the training cohort (a), ΔORR exceeded +10% at approximately 2000 mAU/mL, whereas in the validation cohort (b), this transition occurred at higher DCP levels.

Figure S3: Robustness of the DCP‐dependent ΔORR transition across resampling strategies.

(A) Leave‐one‐center‐out (LOCO) analysis (n = 31), showing the distribution of DCP values at which the predicted difference in objective response rate between Durvalumab plus Tremelimumab and Atezolizumab plus Bevacizumab (ΔORR) reached the prespecified threshold of 10% after sequentially excluding each participating center.

(B) Repeated random split sensitivity analysis (n = 100), in which the same ΔORR threshold was estimated across randomly resampled training subsets.

In both panels, boxplots represent the median and interquartile range, with individual points indicating estimates from each resampling iteration. Horizontal dashed lines denote reference DCP levels (2000 and 10 000 mAU/mL). Although greater variability was observed in the repeated random split analysis, likely reflecting reduced effective sample size per iteration, the central tendency consistently localized to the high‐DCP range, supporting the robustness of the DCP‐dependent transition zone.

Figure S4: DCP‐dependent ΔORR in TKI‐naïve patients.

ΔORR increased progressively with higher baseline DCP levels, with Atezolizumab plus Bevacizumab favoured at lower DCP levels and Durvalumab plus Tremelimumab becoming increasingly advantageous in the higher DCP range. Horizontal dashed lines indicate the prespecified threshold of a 10% absolute difference in ORR. This pattern mirrors the primary analysis, indicating that the observed DCP‐dependent interaction is not driven by prior TKI exposure.

Figure S5: Sensitivity analysis including patients with BCLC stage A disease.

When patients with BCLC stage A disease were additionally included in the pooled cohort, the interaction between baseline DCP and treatment regimen remained significant (deviance difference = 22.83; df = 4; p = 0.00014). The overall directional pattern was preserved, with Atez/Bev favoured in the low‐DCP range and the relative benefit shifting toward Dur/Tre at higher DCP levels.

Figure S6: DCP‐stratified ΔORR estimated within baseline DCP bins after propensity score weighting.

Panels show the ΔORR between Dur/Tre and Atez/Bev estimated within prespecified baseline DCP categories: [0, 200], [200, 2000], [2000, 20 000], and [> 20 000] mAU/mL. Point estimates and horizontal bars represent Wald's 95% confidence intervals from the weighted regression models. The dashed vertical line indicates ΔORR = 0. The numbers on the right denote the observed sample size (n) and effective sample size (ESS) of each regimen within each DCP bin. (a) ATC weighting estimates. (b) Estimates using ATO weighting.

Figure S7: Patients who achieved ORR had significantly longer OS than those without ORR in both the Atez/Bev (a) and Dur/Tre (b) groups.

Table S1: Baseline characteristics of the Atez/Bev and Dur/Tre groups.

Table S2: Overall treatment outcome.

Table S3: Adverse events.

Table S4: Baseline characteristics of the Atez/Bev and Dur/Tre groups in the training and validation cohorts.

Table S5: Standardized mean differences after ATC and ATO weighting.

Table S6: Baseline characteristics according to a reference DCP threshold of 2000 mAU/mL, stratified by treatment regimen.

Tanaka K., Tsuji K., Hiraoka A., et al., “Baseline DCP May Modify Response to Atezolizumab‐Bevacizumab Versus Durvalumab‐Tremelimumab in Unresectable HCC ,” Liver International 46, no. 8 (2026): e70803, 10.1111/liv.70803.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.