Durvalumab plus HAIC-FOLFOX followed by maintenance durvalumab for hepatocellular carcinoma with major portal invasion: phase 2 DurHope study Authors: Junzhe Yi, Jiongliang Wang, Yimin Zhang, Xiongying Jiang, Song Chen, Jie Xu, Xintong Wu, Suixing Zhong, Qifeng Chen, Yue Hu, Yujia Song, Genjun Tan, Yunan Zhang, Jibin Li, Ming Zhao, Ning Lyu Published: 2026 ABSTRACT Hepatocellular carcinoma (HCC) patients with Vp4 portal vein tumor thrombus (PVTT) have a poor prognosis and are underrepresented in global clinical trials. We conducted a single-arm phase 2 study (DurHope) enrolling 30 patients with Vp3/4 PVTT receiving durvalumab plus hepatic arterial infusion therapy (HAIC) of FOLFOX regimen (fluorouracil, leucovorin, and oxaliplatin) as first-line treatment. Primary endpoint was 1-year overall survival (OS) rate, and secondary endpoints included progression-free survival, objective response rate per Response Evaluation Criteria in Solid Tumors version 1.1, and safety. The pre-specified study endpoints were achieved, the 1-year OS rate was 63.3%, with a median OS of 13.9 months (95% CI, 10.7–not reached [NR]) at the final analysis. In exploratory, post hoc biomolecular analyses, tumor single-nucleus RNA sequencing revealed chemotherapy resistance and immune escape signatures in nonresponders, including a MECOM+ malignant subcluster. Responders showed increased immune infiltration and stronger immune-tumor interactions. Peripheral blood dynamic single-cell RNA sequencing demonstrated expanded T-cell subsets and increased expression of cytotoxic-related genes after treatment, which was more pronounced in responders. This was accompanied by decreased nonclassical monocyte frequency and attenuated anti-inflammatory phenotype in responders. Findings were validated by immunohistochemistry and in vivo models. These data suggested the promising efficacy and generally tolerable toxicity of Durvalumab plus HAIC-FOLFOX in patients with Vp4 PVTT and provided candidate biomarkers. ClinicalTrials.gov number: NCT04945720. Introduction Primary liver cancer is the third leading cause of cancer-related mortality globally, with hepatocellular carcinoma (HCC) being the predominant type, and over half of HCC patients are diagnosed at an advanced stage1. Advanced HCC is often complicated by portal vein tumor thrombus (PVTT), with Vp4 PVTT representing the most severe subtype, which obstructs the main portal vein and leads to poor prognosis. This condition can cause portal hypertension and severe gastroesophageal varices, increasing the risk of liver function deterioration and variceal hemorrhage2. In a study evaluating untreated HCC patients, the primary causes of death were gastrointestinal bleeding (34.1%), cancer-related reasons (31.8%), and liver failure (25.0%)3. These findings underscore the need for safe and effective treatments for patients with severe portal invasion. Atezolizumab combined with bevacizumab (atezo/beva) is the preferred first-line treatment for advanced HCC, and the IMbrave150 trial showed significant overall survival (OS) benefits of atezo/beva versus sorafenib (19.2 vs. 13.4 months)4,5. However, in high-risk patients–Vp4 PVTT, tumor involvement > 50% of the liver, or bile duct invasion–the median OS decreased to 7.6 months6. The variceal hemorrhage incidence was also higher in the Vp4 PVTT subgroup compared to the overall population (14% vs. 2%)7. Given the increased risk of bleeding events associated with bevacizumab, caution is advised when considering this combination regimen for patients with Vp4 PVTT8. Tremelimumab plus durvalumab (STRIDE) is another first-line option, but the HIMALAYA trial excluded patients with Vp4 PVTT, resulting in a lack of efficacy data for the STRIDE regimen in such patients4,9. Other major phase 3 trials–including CheckMate 9DW (nivolumab plus ipilimumab vs. lenvatinib or sorafenib)10, REFLECT (lenvatinib vs. sorafenib)11, and LEAP-002 (lenvatinib plus pembrolizumab vs. lenvatinib)12–also excluded patients with Vp4 PVTT, leaving a significant gap in evidence regarding the effectiveness of these treatments in patients with advanced HCC and severe portal invasion. Hepatic arterial infusion therapy (HAIC) is a locoregional interventional procedure widely used in Asia13–15 and has been recommended as an alternative treatment for liver-confined HCC and severe PVTT. Recently, HAIC has also garnered increasing attention in Western countries as a recognized treatment for selected patients with advanced liver cancer16. The FOHAIC-1 phase 3 randomized controlled trial (RCT) demonstrated that patients receiving HAIC with the FOLFOX regimen (fluorouracil, leucovorin, and oxaliplatin) achieved significantly better OS than those treated with sorafenib (13.9 vs. 8.2 months; HR 0.41, 95% CI, 0.30–0.56)17. The superior survival benefits were also observed in the high-risk subgroup, including patients with Vp4 PVTT (OS: 10.8 vs. 5.7 months; HR 0.34, 95% CI, 0.22–0.54). Another phase 3 RCT comparing sorafenib plus HAIC-FOLFOX to sorafenib alone in patients with PVTT found that the median OS was significantly improved in the combination group (13.4 vs. 7.1 months, HR 0.35, 95% CI, 0.26–0.48)18. Stratified analysis by portal vein invasion grade showed that the combination therapy significantly prolonged OS compared to sorafenib alone in both Vp3 (13.5 vs. 6.3 months, P < 0.001) and Vp4 subgroups (9.5 vs. 5.5 months, P < 0.001). In several solid tumors, such as small-cell lung cancer and triple-negative breast cancer, the combination of chemotherapy and immune checkpoint inhibitors (ICIs) has become the standard first-line treatment19–21. Chemotherapy plays a vital role not only through its cytotoxic effects but also by modulating the immune system. Oxaliplatin, a key component of the FOLFOX regimen, has been shown to induce immunogenic cell death (ICD), enhancing the exposure of immunostimulatory molecules, including cytokines like type I interferons and damage-associated molecular patterns (DAMPs)22. This activation of both innate and adaptive immune responses provides a strong rationale for combining HAIC-FOLFOX with ICIs in the treatment of HCC. Durvalumab, an anti-PD-L1 monoclonal body, is recommended as an alternative first-line ICI treatment for advanced HCC4. The HIMALAYA study demonstrated its non-inferior efficacy compared to sorafenib, along with a favorable safety profile and a lower grade ≥ 3 treatment-related adverse events (TRAEs)9. Therefore, integrating durvalumab with HAIC-FOLFOX may offer synergistic therapeutic benefits by harnessing the immunomodulatory effects of oxaliplatin and the immune activation provided by PD-L1 blockade. In this context, we conducted a biomolecular exploratory phase 2 study to evaluate the safety, efficacy, and potential predictive biomarkers of first-line durvalumab plus HAIC-FOLFOX for HCC with major portal vein tumor thrombus (DurHope, a descriptive label for this study). Our comprehensive biomarker analyses encompassed single-nucleus RNA sequencing (snRNA-seq) and in situ analysis of tumor biopsies, dynamic single-cell RNA sequencing (scRNA-seq) of paired peripheral blood samples, and therapeutic experiments using in vivo mouse models. Results Patient characteristics Between April 2022 and July 2023, 30 patients were enrolled and treated with a trial regimen of durvalumab combined with HAIC-FOLFOX (Fig. 1). Twenty patients (66.7%) had hepatitis B virus infection and 6 patients (20.0%) had both metabolic dysfunction-associated steatotic liver disease (MASLD) and hepatitis B virus infection. Twenty-four patients (80.0%) present with Vp4 PVTT and 13 (43.3%) present with > 50% tumor volume involvement of the liver (Table 1).Fig. 1The DurHope study flow diagram and overview.The epidemiological data, clinical trial design, and key biomolecular exploratory analyses in this study are outlined. In the biomolecular exploratory, phase 2 DurHope trial, patients with Vp3 or Vp4 portal vein tumor thrombus were treated with first-line durvalumab combined with HAIC-FOLFOX chemotherapy. Pre-treatment tumor biopsies and serial peripheral blood samples from responders and nonresponders were subjected to tumor snRNA-seq, PBMC scRNA-seq, in situ multiplex immunohistochemistry to identify potential predictive biomarkers and investigate treatment-associated immune remodeling. Preclinical mouse models were used to further evaluate the synergistic and therapeutic effects of the combination regimen. This figure were created in BioRender. Lyu, N. (2026) https://BioRender.com/vhb8kv7.Table 1Patient baseline characteristicsCharacteristicTotal (N = 30)n (%)Vp4 (n = 24)n (%)Age, yearsa50 (34–65)48.5 (38–63)SexMale27 (90.0)22 (91.7)Female3 (10.0)2 (8.3)EtiologyHBV26 (86.7)22 (91.7)HCV3 (10.0)2 (8.3)Unknown or other1 (3.3)0Etiology (based on MASLD)HBV20 (66.7)17 (70.8)MASLD-HBV6 (20.0)5 (20.8)HCV3 (10.0)2 (8.3)Unknown or other1 (3.3)0Antiviral treatmentb29 (100.0)24 (100.0)Child-Pugh classA27 (90.0)22 (91.7)B3 (10.0)2 (8.3)ALBI grade117 (56.7)13 (54.2)213 (43.3)11 (45.8)Performance status024 (80.0)18 (75.0)16 (20.0)6 (25.0)Serum a-fetoprotein level, ng/mLc2666.0 (466.5–81491.3)3206.5 (489.5–94077.3)≤ 4006 (20.0)4 (16.7)> 40024 (80.0)20 (83.3)Tumor diameter, cmdMean ± SD10.9 ± 4.111.9 ± 4.3Tumor numbere1–36 (20.0)4 (16.7)> 324 (80.0)20 (83.3)Tumor involvement of the liver, %≤ 5017 (56.7)14 (58.3)> 5013 (43.3)10 (41.7)Portal vein invasionVp36 (20.0)0Vp424 (80.0)24 (100.0)Tumor involvement > 50% of the liver and/or Vp4Absent4 (13.3)0Present26 (86.7)24 (100.0)Venous invasionAbsent20 (66.7)17 (70.8)Present10 (33.3)7 (29.2)Hepatic vein9 (30.0)6 (25.0)Inferior vena cava2 (6.7)f2 (8.3)Extrahepatic spreadAbsent20 (66.7)15 (62.5)Present10 (33.3)9 (37.5)Lung6 (20.0)6 (25.0)Lymph nodes5 (16.7)4 (16.7)ALBI albumin-bilirubin, HBV hepatitis B virus, HCV hepatitis C virus, MASLD metabolic dysfunction-associated steatotic liver disease.aData is shown as a median and full range.bPatients with viral hepatitis in the general cohort (n = 29) and Vp4 group (n = 24) were included; cData is shown as median and quartile.dThe tumor diameter refers to the sum of the maximum diameters of target lesions.eThis refers to the number of intrahepatic lesionsfThe 2 patients with inferior vena cava invasion also have concurrent Vp4 PVTT. The DurHope study flow diagram and overview. The epidemiological data, clinical trial design, and key biomolecular exploratory analyses in this study are outlined. In the biomolecular exploratory, phase 2 DurHope trial, patients with Vp3 or Vp4 portal vein tumor thrombus were treated with first-line durvalumab combined with HAIC-FOLFOX chemotherapy. Pre-treatment tumor biopsies and serial peripheral blood samples from responders and nonresponders were subjected to tumor snRNA-seq, PBMC scRNA-seq, in situ multiplex immunohistochemistry to identify potential predictive biomarkers and investigate treatment-associated immune remodeling. Preclinical mouse models were used to further evaluate the synergistic and therapeutic effects of the combination regimen. This figure were created in BioRender. Lyu, N. (2026) https://BioRender.com/vhb8kv7. Table 1Patient baseline characteristicsCharacteristicTotal (N = 30)n (%)Vp4 (n = 24)n (%)Age, yearsa50 (34–65)48.5 (38–63)SexMale27 (90.0)22 (91.7)Female3 (10.0)2 (8.3)EtiologyHBV26 (86.7)22 (91.7)HCV3 (10.0)2 (8.3)Unknown or other1 (3.3)0Etiology (based on MASLD)HBV20 (66.7)17 (70.8)MASLD-HBV6 (20.0)5 (20.8)HCV3 (10.0)2 (8.3)Unknown or other1 (3.3)0Antiviral treatmentb29 (100.0)24 (100.0)Child-Pugh classA27 (90.0)22 (91.7)B3 (10.0)2 (8.3)ALBI grade117 (56.7)13 (54.2)213 (43.3)11 (45.8)Performance status024 (80.0)18 (75.0)16 (20.0)6 (25.0)Serum a-fetoprotein level, ng/mLc2666.0 (466.5–81491.3)3206.5 (489.5–94077.3)≤ 4006 (20.0)4 (16.7)> 40024 (80.0)20 (83.3)Tumor diameter, cmdMean ± SD10.9 ± 4.111.9 ± 4.3Tumor numbere1–36 (20.0)4 (16.7)> 324 (80.0)20 (83.3)Tumor involvement of the liver, %≤ 5017 (56.7)14 (58.3)> 5013 (43.3)10 (41.7)Portal vein invasionVp36 (20.0)0Vp424 (80.0)24 (100.0)Tumor involvement > 50% of the liver and/or Vp4Absent4 (13.3)0Present26 (86.7)24 (100.0)Venous invasionAbsent20 (66.7)17 (70.8)Present10 (33.3)7 (29.2)Hepatic vein9 (30.0)6 (25.0)Inferior vena cava2 (6.7)f2 (8.3)Extrahepatic spreadAbsent20 (66.7)15 (62.5)Present10 (33.3)9 (37.5)Lung6 (20.0)6 (25.0)Lymph nodes5 (16.7)4 (16.7)ALBI albumin-bilirubin, HBV hepatitis B virus, HCV hepatitis C virus, MASLD metabolic dysfunction-associated steatotic liver disease.aData is shown as a median and full range.bPatients with viral hepatitis in the general cohort (n = 29) and Vp4 group (n = 24) were included; cData is shown as median and quartile.dThe tumor diameter refers to the sum of the maximum diameters of target lesions.eThis refers to the number of intrahepatic lesionsfThe 2 patients with inferior vena cava invasion also have concurrent Vp4 PVTT. Patient baseline characteristics ALBI albumin-bilirubin, HBV hepatitis B virus, HCV hepatitis C virus, MASLD metabolic dysfunction-associated steatotic liver disease. aData is shown as a median and full range. bPatients with viral hepatitis in the general cohort (n = 29) and Vp4 group (n = 24) were included; cData is shown as median and quartile. dThe tumor diameter refers to the sum of the maximum diameters of target lesions. eThis refers to the number of intrahepatic lesions fThe 2 patients with inferior vena cava invasion also have concurrent Vp4 PVTT. Efficacy With a median follow-up of 21.6 months (interquartile range [IQR], 15.6–25.3), the 1-year OS rate was 63.3% (95% CI, 45.0–81.6%) and median OS was 13.9 months (95% CI, 10.7–not reached [NR]) at the prespecified final analysis (Fig. 2B). After extending the follow-up to 31st May 2025, the updated median OS was 15.4 months (95% CI, 10.7–23.1). The median progression-free survival (PFS) of whole disease and PFS of intrahepatic tumor and PVTT were 9.0 months (95% CI, 4.5–12.2) and 11.6 months (95% CI, 9.0–NR), respectively (Fig. 2C). The median time to progression (TTP) and intrahepatic TTP were 11.0 months (95% CI, 6.3–NR) and 13.0 months (95% CI, 11.6–NR) (Fig. 2D).Fig. 2Clinical efficacy of durvalumab plus HAIC.A Tumor size changes from baseline based on radiologic imaging. B Kaplan-Meier curves for overall survival in the study population and subgroups stratified by treatment response, by two-sided log-rank test. C, D Kaplan-Meier curves for progression-free survival, intrahepatic tumor progression-free survival, time to progression, and intrahepatic tumor progression-free survival. E, F Swimmer plots and response profiles of 30 enrolled patients and 5 patients who achieved complete response. G, H Imaging scans and tumor marker trajectories of two representative cases that achieved complete response. CR1 first documented as complete response, CR2 subsequently documented as a complete response, CR_LC last follow-up confirmed a complete response, PD progression disease, PR partial response. Clinical efficacy of durvalumab plus HAIC. A Tumor size changes from baseline based on radiologic imaging. B Kaplan-Meier curves for overall survival in the study population and subgroups stratified by treatment response, by two-sided log-rank test. C, D Kaplan-Meier curves for progression-free survival, intrahepatic tumor progression-free survival, time to progression, and intrahepatic tumor progression-free survival. E, F Swimmer plots and response profiles of 30 enrolled patients and 5 patients who achieved complete response. G, H Imaging scans and tumor marker trajectories of two representative cases that achieved complete response. CR1 first documented as complete response, CR2 subsequently documented as a complete response, CR_LC last follow-up confirmed a complete response, PD progression disease, PR partial response. Sixteen patients (53.3%) achieved objective responses (complete response [CR]: n = 5, 16.7%; partial response [PR]: n = 11, 36.7%), and 27 patients (90.0%) demonstrated disease control (Fig. 2A, E and Supplementary Table 1). The median time to response was 1.6 months (IQR, 1.3–2.8), and the median duration of response (DOR, the time from the first documented objective response to the first radiographic progression or death) was 9.1 months (95% CI, 6.7–NR). All five patients who achieved CR maintained this status from initial documentation through subsequent and most recent imaging assessments, with tumor markers remaining within the normal range throughout the post-CR period. At data cutoff, these patients were alive without tumor recurrence, with a mean DOR of 15.9 months (Fig. 2F–H; Supplementary Fig. 1–4 and Supplementary Table 2). In post hoc exploratory subgroup analysis based on the presence of MASLD, the median OS and PFS for patients in the MASLD-viral hepatitis subgroup were similar to those in the viral hepatitis subgroup (OS: 17.6 vs. 13.9 months, P = 0.89; PFS: 8.9 vs. 9.0 months, P = 0.79) (Supplementary Fig. 5 and Supplementary Table 3, 4). Additionally, after stratification by the presence of Vp4 PVTT, the median OS and PFS for the Vp4 subgroup were comparable to the Vp3 subgroup (OS: 13.1 vs. 16.9 months, P = 0.55; PFS: 10.8 vs. 6.7 months, P = 0.26) (Supplementary Fig. 6 and Supplementary Table 5, 6). In post hoc exploratory analyses according to liver function reserve, the favorable efficacy outcomes were observed in patients with Child-Pugh A liver function (OS: 16.9 month; PFS: 9.0 month), and results for patients with Child-Pugh B7 liver function were listed separately (Supplementary Fig. 7 and Supplementary Table 7). When stratified by ALBI grade, survival outcomes were similar between grade 1 and grade 2 subgroups (OS: 16.9 vs. 12.1 months, P = 0.75; PFS: 6.8 vs. 10.9 months, P = 0.23) (Supplementary Fig. 8 and Supplementary Table 7). Safety The profiles of treatment-related adverse events (TRAEs) in the study population were summarized in Table 2. The most common TRAEs included hypoalbuminemia (23 [76.7%]), reduced hemoglobin (22 [73.3%]), thrombocytopenia (21 [70.0%]), and leukopenia (21 [70.0%]). Grade ≥ 3 TRAEs occurred in 13 patients (43.3%), with neutropenia (11 [36.7%]) and thrombocytopenia (8 [26.7%]) being most frequent. Patients present with neutropenia or leukopenia generally received short courses of granulocyte colony-stimulating factor; those with thrombocytopenia were treated with thrombopoietin or thrombopoietin receptor agonists. If these hematologic TRAEs were not adequately controlled within 2 weeks or previously managed TRAEs recurred in subsequent treatment cycles, dose adjustments or delays of the study regimen were implemented. One patient experienced grade 2 treatment-related gastrointestinal bleeding, which was managed with endoscopic variceal ligation and other symptomatic treatment. The patient subsequently resumed the study regimen without any recurrent bleeding events. No treatment-related deaths were observed.Table 2Treatment-related adverse eventsEventAny graden (%)3–4 graden (%)Patients with at least one event30 (100.0)13 (43.3)Hypoalbuminemia23 (76.7)1 (3.3)Reduced hemoglobin22 (73.3)1 (3.3)Thrombocytopenia21 (70.0)8 (26.7)Leukopenia21 (70.0)4 (13.3)Neutropenia20 (66.7)11 (36.7)Elevated AST17 (56.7)3 (10.0)Fever17 (56.7)0Pain abdominal16 (53.3)0Elevated total bilirubin14 (46.7)1 (3.3)Elevated ALT12 (40.0)3 (10.0)Nausea11 (36.7)0Vomiting10 (33.3)0Elevated INR9 (30.0)0Abdominal distension8 (26.7)0Rash8 (26.7)0Diarrhea7 (23.3)4 (13.3)Ascites7 (23.3)0Constipation4 (13.3)0Pruritus4 (13.3)0Weight reduction4 (13.3)0Backache3 (10.0)0Gingival bleeding3 (10.0)0Hiccup3 (10.0)0Dizziness2 (6.7)0Elevated creatinine2 (6.7)0Pneumonia2 (6.7)0Gastrointestinal bleeding1 (3.3)0ALT alanine aminotransferase, AST aspartate aminotransferase, INR international normalized ratio. Table 2Treatment-related adverse eventsEventAny graden (%)3–4 graden (%)Patients with at least one event30 (100.0)13 (43.3)Hypoalbuminemia23 (76.7)1 (3.3)Reduced hemoglobin22 (73.3)1 (3.3)Thrombocytopenia21 (70.0)8 (26.7)Leukopenia21 (70.0)4 (13.3)Neutropenia20 (66.7)11 (36.7)Elevated AST17 (56.7)3 (10.0)Fever17 (56.7)0Pain abdominal16 (53.3)0Elevated total bilirubin14 (46.7)1 (3.3)Elevated ALT12 (40.0)3 (10.0)Nausea11 (36.7)0Vomiting10 (33.3)0Elevated INR9 (30.0)0Abdominal distension8 (26.7)0Rash8 (26.7)0Diarrhea7 (23.3)4 (13.3)Ascites7 (23.3)0Constipation4 (13.3)0Pruritus4 (13.3)0Weight reduction4 (13.3)0Backache3 (10.0)0Gingival bleeding3 (10.0)0Hiccup3 (10.0)0Dizziness2 (6.7)0Elevated creatinine2 (6.7)0Pneumonia2 (6.7)0Gastrointestinal bleeding1 (3.3)0ALT alanine aminotransferase, AST aspartate aminotransferase, INR international normalized ratio. Treatment-related adverse events ALT alanine aminotransferase, AST aspartate aminotransferase, INR international normalized ratio. The median duration of adherence to the combination regimen was 4.6 months (IQR, 2.9–6.5 months), while the median duration of adherence to at least one study agent was 7.3 months (IQR, 3.9–13.4 months). Throughout the study, a total of 248 doses of durvalumab and 157 cycles of HAIC-FOLFOX were administered, with medians of 6 doses (IQR, 4–13) and 6 cycles (IQR, 3.3–7.8) for each patient. During the FOLFOX infusion procedure, the primary complication was transient abdominal pain occurring in the late phase of oxaliplatin infusion (16 [53.3%]), which was alleviated by either slowing the infusion rate or administering medical treatment. No patients discontinued the FOLFOX regimen due to infusion-related complications. Treatment discontinuation due to TRAEs occurred in 3 patients (10.0%). Two patients discontinued the FOLFOX regimen due to chemotherapy-related myelosuppression and were managed with granulocyte colony-stimulating factor and subsequently transitioned to durvalumab maintenance. Another patient discontinued durvalumab due to pneumonia during the period of durvalumab monotherapy and was managed with symptomatic treatment (e.g., anti-infective therapy and oxygen inhalation). No patients discontinued both durvalumab and HAIC due to TRAEs. Tumor transcriptome landscape and molecular profiles associated with treatment response To identify treatment response biomarkers, we conducted exploratory analyses by performing snRNA-seq on tumor biopsy tissues from 12 participants (6 responders defined as achieving CR or PR, and 6 nonresponders defined as achieving stable or progressive disease) (Supplementary Tables 8, 9). Unsupervised clustering identified 18 distinct cell clusters, including epithelial cells, immune cells, and stromal cells (Fig. 3A, B and Supplementary Fig. 9). Differential expressed gene and pathway enrichment analyses were further performed in malignant hepatocytes. Nonresponders showed increased expression of genes associated with potential therapy resistance, including CNTN1 (cell-adhesion molecule implicated in invasion and drug resistance), IGF2 (growth factor activating tumor proliferation pathway), HMGA2 (chromatin regulator promoting epithelial–mesenchymal transition and chemoresistance), as well as additional adhesion and metabolic regulators (e.g., DACH1, ANKRD30A, MGAM2). The up-regulated genes in nonresponder were enriched in pathways known to promote tumor stress adaptation and therapy resistance, such as GTPase regulation and autophagy (Fig. 3C, D and Supplementary Table 10). Consistently, malignant cells from nonresponders exhibited lower expression levels of proliferation and immune surveillance signature genes, alongside higher expression of immune escape signature genes (Fig. 3E). Correlation analyses showed that greater tumor shrinkage was associated with lower immune-escape signature scores and higher proliferation and immune surveillance scores (Supplementary Fig. 10). Pseudotime analysis identified two distinct fates in malignant cells and suggested divergent evolutionary paths between responders (primarily in fate 1) and nonresponders (primarily in fate 2), potentially contributing to differences in treatment outcomes (Fig. 3F). Fate 1 cells exhibited high expression of S100A10, SA100A11, YBX1, and SERBP1, genes associated with immune activation and regulation of cell proliferation, suggesting increased sensitivity to chemotherapy and immunotherapy. In contrast, fate 2 cells were marked by high expression of SMYD3, CHD1L, QSOX1, and BIRC6, which are linked to tumor growth, DNA repair, and immune evasion, potentially conferring greater resistance to treatment (Fig. 3F).Fig. 3The snRNA-seq analysis of tumor biopsy tissues from the responder and non-responder group.A UMAP plot of major cell types in the tumor biopsy samples, and cells from different cell types are marked by distinct colors. B Comparison of the proportions of major cell types between the responder and nonresponder groups. C The volcano plot showing differentially expressed genes of tumor cells in the responder and nonresponder groups, and adjustments were made for multiple comparisons. D Gene Ontology (GO) analysis of differentially expressed genes of malignant cells in responder group versus nonresponder group. E Box plots showing the proliferation, immune escape, and immune surveillance scores of malignant cells from responders and nonresponders (n = 6 for each group), by two-tailed Student’s t test. F Pseudotime analysis showing two differentiation trajectories of malignant cells and a branched heatmap of dynamically regulated genes. G Stacked bar chart showing the percentages of malignant cell subclusters between responders and nonresponders. H Heatmap of selected Hallmark pathway enrichment in PEG10 + , MECOM + , and CNTN1+ malignant cells. I Kaplan-Meier survival curves for MECOM+ malignant-cell signature in the TCGA HCC cohort, by two-sided log-rank test. In each box plot, the box represents the interquartile range, and whiskers represent minima and maxima. Source data are provided as a Source Data file. The snRNA-seq analysis of tumor biopsy tissues from the responder and non-responder group. A UMAP plot of major cell types in the tumor biopsy samples, and cells from different cell types are marked by distinct colors. B Comparison of the proportions of major cell types between the responder and nonresponder groups. C The volcano plot showing differentially expressed genes of tumor cells in the responder and nonresponder groups, and adjustments were made for multiple comparisons. D Gene Ontology (GO) analysis of differentially expressed genes of malignant cells in responder group versus nonresponder group. E Box plots showing the proliferation, immune escape, and immune surveillance scores of malignant cells from responders and nonresponders (n = 6 for each group), by two-tailed Student’s t test. F Pseudotime analysis showing two differentiation trajectories of malignant cells and a branched heatmap of dynamically regulated genes. G Stacked bar chart showing the percentages of malignant cell subclusters between responders and nonresponders. H Heatmap of selected Hallmark pathway enrichment in PEG10 + , MECOM + , and CNTN1+ malignant cells. I Kaplan-Meier survival curves for MECOM+ malignant-cell signature in the TCGA HCC cohort, by two-sided log-rank test. In each box plot, the box represents the interquartile range, and whiskers represent minima and maxima. Source data are provided as a Source Data file. We further compared malignant cell subclusters between responders and nonresponders, identifying 3 subclusters (C0, C3, and C6) enriched in nonresponders (Fig. 3G and Supplementary Fig. 11A–E). These subclusters respectively exhibited high expression of PEG10, MECOM, and CNTN1, which were implicated in tumor growth, anti-apoptosis, and chemoresistance23–25. Furthermore, the increased fraction of MECOM+ malignant cells tended to associate with less tumor shrinkage (Supplementary Fig. 10). Using CancerSEA hallmark gene sets26, we found these malignant subclusters, particularly MECOM+ malignant cells, showed enrichment in pathways related to treatment resistance (Fig. 3H). Due to the limited sample size of our snRNA analysis, we extended our investigation by exploring the correlation between the signature genes of these malignant subclusters and patient prognosis in the TCGA HCC cohort. Further analysis in the TCGA HCC cohort revealed that higher expression of MECOM+ cell signature genes was significantly associated with worse OS (Fig. 3I and Supplementary Fig. 11F). Since durvalumab was part of the treatment regimen, immune cell composition and infiltration in the tumor microenvironment likely influenced outcomes. We identified 9 distinct immune cell clusters, including 3 clusters for macrophages (MSR1 + , CD109 + , and CSF3R + ), 3 clusters for T cells (SKAP1 + , IL7R + , and FYN + ), and 3 clusters for B/plasma cells (IL32 + , B cells; IGHG1 + , IGHG4 + , plasma cells) (Fig. 4A, B and Supplementary Fig. 12A, B). Within the macrophage compartment, CD109+ macrophages showed an M0-like phenotype, while CSF3R+ and MSR1+ macrophages were respectively enriched for M1-like (anti-tumorigenic) and M2-like (pro-tumorigenic) signatures. Among T cell subsets, FYN + , SKAP + , and IL7R + T cells corresponded to exhausted, cytotoxic, and effector memory phenotypes (Supplementary Fig. 12C, D). Responders showed higher immune cell proportion and greater expression of T cytotoxic and macrophage antigen-processing signature genes (Fig. 4C, D), which were positively correlated with the extent of tumor reduction (Supplementary Fig. 10). This corresponded with increased communication intensity and stronger immune-malignant cell interactions in the responder group (Fig. 4E, F). In contrast, nonresponders displayed reduced immune cell communication, particularly T cells, indicating an immunosuppressive state (Fig. 4G). Further pathway analysis revealed significant enrichment of the insulin-like growth factor, hepatocyte growth factor, platelet-derived growth factor, and tumor necrosis factor superfamily member 14 pathways in the nonresponder group (Fig. 4H). These pathways are known to promote tumor growth, invasion, and immune evasion, ultimately contributing to an immunosuppressive tumor microenvironment27–30.Fig. 4Characterization of the heterogeneity of immune cells in tumor environment.A UMAP plot of immune cell subsets in the tumor biopsy samples. B Dotplot showing the expression of marker genes for immune cell subtypes. C Proportions of immune cell subsets in responders versus nonresponders. D Box plots showing cytotoxic and exhausted gene-signature scores in T cells and antigen-processing scores of macrophages from responders and nonresponders (n = 6 for each group), by two-tailed Student’s t test. E Bar charts showing the number and strength of cell interaction between responder and nonresponder groups. F Circle plots showing the interaction network among tumor-infiltrating cell subsets in responders and nonresponders. G Scatterplot depicting the dominant senders (sources) and receivers (targets) among tumoural cell subsets. H Bar plot showing overall information flow for selected signaling pathways in responders versus nonresponders. I, J Representative multiplex immunohistochemistry images from responders (n = 11) and nonresponders (n = 7) and digital pathology quantification of selected markers by response status, by two-tailed Student’s t-test (scale bars in each image are 50 μm). The data are expressed as the mean ± SEM. In each box plot, the box represents the interquartile range, and whiskers represent minima and maxima. Source data are provided as a Source Data file. Characterization of the heterogeneity of immune cells in tumor environment. A UMAP plot of immune cell subsets in the tumor biopsy samples. B Dotplot showing the expression of marker genes for immune cell subtypes. C Proportions of immune cell subsets in responders versus nonresponders. D Box plots showing cytotoxic and exhausted gene-signature scores in T cells and antigen-processing scores of macrophages from responders and nonresponders (n = 6 for each group), by two-tailed Student’s t test. E Bar charts showing the number and strength of cell interaction between responder and nonresponder groups. F Circle plots showing the interaction network among tumor-infiltrating cell subsets in responders and nonresponders. G Scatterplot depicting the dominant senders (sources) and receivers (targets) among tumoural cell subsets. H Bar plot showing overall information flow for selected signaling pathways in responders versus nonresponders. I, J Representative multiplex immunohistochemistry images from responders (n = 11) and nonresponders (n = 7) and digital pathology quantification of selected markers by response status, by two-tailed Student’s t-test (scale bars in each image are 50 μm). The data are expressed as the mean ± SEM. In each box plot, the box represents the interquartile range, and whiskers represent minima and maxima. Source data are provided as a Source Data file. We further applied interactome analysis to investigate crosstalk between malignant subsets enriched in nonresponders (MECOM + , CNTN1 + , and PEG10+ malignant cells) and immune compartments. These malignant clusters established extensive ligand–receptor interactions with multiple immune cell populations, including T cells, macrophages, B cells, and plasma cells (Supplementary Fig. 13A–C). All T cell subsets engaged in ligand–receptor interactions with malignant cells, with the most prominent interactions observed for FYN+ T cells, which exhibited an exhausted phenotype (Supplementary Fig. 13A, B). Interactions between nonresponder-enriched malignant cells and immune cells were mediated predominantly through fibronectin 1 (FN1)–integrin (FN1–ITGA4/ITGB1, FN1–ITGAV/ITGB1) and vitronectin (VTN)–integrin (VTN–ITGA8/ITGB1) signaling pathways (Supplementary Fig. 14A–C). These identified ligand–receptor pairs have been reported as pathways exploited by tumor cells to remodel the extracellular matrix and modulate immune-cell infiltration and function state within the microenvironment31–33. Additionally, we conducted in situ immunohistochemistry (IHC) to assess PD-L1 protein expression and its correlation with clinical outcomes (Supplementary Fig. 15A, B). PD-L1 expression showed an upward trend in responders, though not statistically significant. Higher PD-L1 immune cell (IC) score of ≥ 1 % was associated with improved PFS, suggesting a potential link between PD-L1 status and clinical outcomes (Supplementary Fig. 15C). Further mIHC analysis confirmed higher abundance of total CD8 + T cells and GZMB + CD8 + T cells in responders, indicating a more cytotoxic state. Meanwhile, the proportion of PD-1 + CD8 + T cells showed no significant difference between responders and nonresponders (Fig. 4I, J and Supplementary Fig. 12E, F). The significantly increased abundance of total macrophage and HLA-DR + CD68 + M1 macrophage was observed in responders, suggesting their enhanced immune surveillance (Fig. 4I, J and Supplementary Fig. 12E, F). Dynamic scRNA reveals cellular composition and immunological changes during treatment Exploring dynamic biomarkers can help elucidate the therapeutic impact on the immune microenvironment and guide early decision-making. We conducted further exploratory analyses by performing scRNA-seq on paired peripheral blood mononuclear cells (PBMC) samples collected at baseline (cycle 1, day 1[C1D1]) and on treatment (cycle 1, day 21[C1D21]) from responders and nonresponders (Fig. 5A, B). During treatment, proliferating T cells showed an upward trend in both responders and nonresponders. In contrast, non-classical monocytes declined significantly in responders, while no significant change was observed in nonresponders.Fig. 5Overall immune landscapes and dynamic changes of T cell subsets in peripheral blood.A UMAP plot of immune cell subsets in the peripheral blood mononuclear cells (PBMC). B Box plots showing the fold change in proportions of immune cell types among paired PBMCs in responders (n = 5), nonresponders (n = 5), and both (n = 10). * P = 0.0433 (All, T_MKI67), * P = 0.0317 (Responder, Nonclassical monocytes). C Volcano plot showing the differentially expressed genes of peripheral CD8 + T cells before and during treatment (cycle 1, day 21), and adjustments were made for multiple comparisons. D GSVA analysis of T-cell functional signatures before and after treatment. E UMAP plot of T/NK cell subsets in PBMC samples. F Fold change in CD8 + T cell subset proportions in paired PBMCs from responders (n = 5), nonresponders (n = 5), and both (n = 10). G Pseudotime trajectory of CD8 + T cells inferred by Monocle2, showing differentiation from naive CCR7 + CD8 + T cells to GZMH + CD8+ effector cells and CX3CR1 + CD8+ effector memory cells. H, I Box plots showing the cytotoxic and exhausted scores in peripheral CD8 + T cells and NK cells before and after treatment, among responders (n = 5) and nonresponders (n = 5). J Fold change in proportions of CD4 + T cell subsets among PBMCs in responders (n = 5), nonresponders (n = 5), and both (n = 10). * P = 0.0433 (All, CD4_IL7R), ** P = 0.0068 (All, CD4_GZMK), * P = 0.0232 (All, Treg_CTLA4). K Pseudotime trajectory of CD4 + T cells inferred by Monocle2, showing differentiation from naive LEF1 + CD4 + T cells to GZMH + CD4+ effector cells and IL7R + CD4+ memory cells. L Box plots showing the co-stimulatory scores in peripheral CD4 + T cells and exhausted scores in Treg before and after treatment, among responders (n = 5) and nonresponders (n = 5). Statistic analyses are conducted by two-tailed Mann–Whitney U-test (B, F, J) or two-tailed Student’s t test (H, I, L). In each box plot, the box represents the interquartile range, and whiskers represent minima and maxima. Source data are provided as a Source Data file. Overall immune landscapes and dynamic changes of T cell subsets in peripheral blood. A UMAP plot of immune cell subsets in the peripheral blood mononuclear cells (PBMC). B Box plots showing the fold change in proportions of immune cell types among paired PBMCs in responders (n = 5), nonresponders (n = 5), and both (n = 10). * P = 0.0433 (All, T_MKI67), * P = 0.0317 (Responder, Nonclassical monocytes). C Volcano plot showing the differentially expressed genes of peripheral CD8 + T cells before and during treatment (cycle 1, day 21), and adjustments were made for multiple comparisons. D GSVA analysis of T-cell functional signatures before and after treatment. E UMAP plot of T/NK cell subsets in PBMC samples. F Fold change in CD8 + T cell subset proportions in paired PBMCs from responders (n = 5), nonresponders (n = 5), and both (n = 10). G Pseudotime trajectory of CD8 + T cells inferred by Monocle2, showing differentiation from naive CCR7 + CD8 + T cells to GZMH + CD8+ effector cells and CX3CR1 + CD8+ effector memory cells. H, I Box plots showing the cytotoxic and exhausted scores in peripheral CD8 + T cells and NK cells before and after treatment, among responders (n = 5) and nonresponders (n = 5). J Fold change in proportions of CD4 + T cell subsets among PBMCs in responders (n = 5), nonresponders (n = 5), and both (n = 10). * P = 0.0433 (All, CD4_IL7R), ** P = 0.0068 (All, CD4_GZMK), * P = 0.0232 (All, Treg_CTLA4). K Pseudotime trajectory of CD4 + T cells inferred by Monocle2, showing differentiation from naive LEF1 + CD4 + T cells to GZMH + CD4+ effector cells and IL7R + CD4+ memory cells. L Box plots showing the co-stimulatory scores in peripheral CD4 + T cells and exhausted scores in Treg before and after treatment, among responders (n = 5) and nonresponders (n = 5). Statistic analyses are conducted by two-tailed Mann–Whitney U-test (B, F, J) or two-tailed Student’s t test (H, I, L). In each box plot, the box represents the interquartile range, and whiskers represent minima and maxima. Source data are provided as a Source Data file. Differentially expressed genes and gene set variation analysis revealed upregulated cytotoxic gene expression and enhanced T cell function after treatment (Fig. 5C, D). T/NK cells were reclustered into 13 subclusters, including 4 CD8 + T cell subclusters (CD8 + CCR7 + , naïve T [Tn]; CD8 + GZMH + , CD8 + GZMK + , effector T [Teff]; and CD8 + CX3CR1 + , effector memory T [Tem]), 3 NK cell subclusters (FCGR3A + , GZMB + , and KLRC1 + ), and 5 CD4 + T cell subclusters (CD4 + LEF1 + , naïve T [Tn]; CD4 + GZMK + , CD4 + GZMH + , effector T [Teff]; CD4 + IL7R + , memory T [Tm]; and CD4 + CTLA4 + , Treg) (Fig. 5E and Supplementary Fig. 16A). We next performed pseudotime analysis on CD8 + T cell subsets using Monocle2, which revealed transition paths from CD8 + CCR7+ Tn to Teff (CD8 + GZMH + , CD8 + GZMK + ) and ultimately to CD8 + CX3CR1+ Tem cells. This trajectory confirmed the sequential activation and exhaustion of CD8 + T cells. In responders, the Teff cells increased after treatment, while nonresponders showed a relative increase in naïve T cells (Fig. 5F, G). Correlation analyses further showed that greater tumor shrinkage was associated with more pronounced expansion of Teff and Tem cells (Supplementary Fig. 10). Cytotoxicity-related signature genes were upregulated in CD8 + T cells and NK cells only among responders (Fig. 5H, I and Supplementary Fig. 16B, C). These findings suggest that effector T cell activation and function are limited in the non-responder group. Additionally, CD4 + GZMK+ Teff cells, CD4 + IL7R+ Tm cells and CD4 + CTLA-4+ Tregs increased after treatment, with pseudotime analysis showing transitions from CD4 + LEF1+ Tn to either CD4 + IL7R+ Tm or Teff cells (CD4 + GZMK + , CD4 + GZMH + ) (Fig. 5J, K). No significant changes were observed in the co-stimulatory score for CD4 + T cells and exhausted score for Treg (Fig. 5L and Supplementary Fig. 16D). We then focused on myeloid cells and reclustered them into 10 subclusters, including 7 monocyte subclusters (CD14 + S100A + , CD14 + IL1B + , classical monocytes; FCGR3A + APOBEC3A + , FCGR3A + C1QBA + , FCGR3A + C1QB + , FCGR3A + CXCL4 + ; FCGR3A + MS4A7 + , nonclassical monocytes), 1 megakaryocyte subcluster, and 2 dendritic cell subclusters (Fig. 6A, B). The relative abundances of FCGR3A + C1QB + , FCGR3A + CXCL4 + , and FCGR3A + MS4A7+ nonclassical monocytes decreased after treatment in responders but showed an upward trend in nonresponders (Fig. 6C), and the more pronounced reduction in these subsets correlated with greater tumor shrinkage (Supplementary Fig. 10). Pseudotime analysis revealed monocytes transitioning from classical monocytes (pro-inflammatory phenotype) to non-classical monocytes (anti-inflammatory phenotype) (Fig. 6D). After treatment, the proportion of classical monocytes did not significantly change in both clinical groups, whereas the proportion of nonclassical monocytes significantly decreased in responder group (Fig. 6E). These findings suggest a reduced anti-inflammatory phenotypic transition in responders following combination therapy. To further validate this hypothesis, we calculated the pro-inflammatory phenotype and anti-inflammatory phenotype signature scores for monocytes. Similarly, anti-inflammatory phenotype signature gene expression was significantly decreased in monocytes among responders (Fig. 6F).Fig. 6Dynamic changes of monocyte response in peripheral blood.A UMAP plot of myeloid subsets in the peripheral blood mononuclear cells (PBMC). B Dotplot showing the expression of marker genes for myeloid cell subsets. C Fold change in proportions of immune cell types among PBMCs in responders (n = 5), nonresponders (n = 5), and both (n = 10). * P = 0.016 (Responder, Mono_FCGR3A_MS4A7), * P = 0.0317 (Responder, Mono_FCGR3A_CXCL4). D Pseudotime trajectory of monocytes inferred by Monocle2, showing differentiation from CD14 + S100A+ monocytes into classical CD14 + IL1B+ monocytes and nonclassical FCGR3A+ monocytes (FCGR3A + APOBEC3A + , FCGR3A + C1QBA + , FCGR3A + C1QB + , FCGR3A + CXCL4 + ; FCGR3A + MS4A7 + ). E Proportions of classical and nonclassical monocytes, and the classical/nonclassical ratio before and after treatment, among responders (n = 5) and nonresponders (n = 5). F Box plots showing the pro-inflammatory phenotype and anti-inflammatory phenotype signature scores in peripheral monocytes before and after treatment, among responders (n = 5) and nonresponders (n = 5). G Volcano plot shows the differentially expressed genes of nonclassical monocytes on cycle 1, day 21 (C1D21) in the responder (n = 5) versus nonresponder groups (n = 5), and adjustments were made for multiple comparisons. H Chord diagram illustrating cell–cell communication between monocytes and CD8 + T cells. I Circle plots showing the contribution of latent patterns to specific cell subsets in the MIF signaling pathway network between responder and nonresponder on C1D21. Statistic analyses are conducted by two-tailed Mann–Whitney U-test (C) or two-tailed Student’s t-test (E, F). In each box plot, the box represents the interquartile range, and whiskers represent minima and maxima. Source data are provided as a Source Data file. Dynamic changes of monocyte response in peripheral blood. A UMAP plot of myeloid subsets in the peripheral blood mononuclear cells (PBMC). B Dotplot showing the expression of marker genes for myeloid cell subsets. C Fold change in proportions of immune cell types among PBMCs in responders (n = 5), nonresponders (n = 5), and both (n = 10). * P = 0.016 (Responder, Mono_FCGR3A_MS4A7), * P = 0.0317 (Responder, Mono_FCGR3A_CXCL4). D Pseudotime trajectory of monocytes inferred by Monocle2, showing differentiation from CD14 + S100A+ monocytes into classical CD14 + IL1B+ monocytes and nonclassical FCGR3A+ monocytes (FCGR3A + APOBEC3A + , FCGR3A + C1QBA + , FCGR3A + C1QB + , FCGR3A + CXCL4 + ; FCGR3A + MS4A7 + ). E Proportions of classical and nonclassical monocytes, and the classical/nonclassical ratio before and after treatment, among responders (n = 5) and nonresponders (n = 5). F Box plots showing the pro-inflammatory phenotype and anti-inflammatory phenotype signature scores in peripheral monocytes before and after treatment, among responders (n = 5) and nonresponders (n = 5). G Volcano plot shows the differentially expressed genes of nonclassical monocytes on cycle 1, day 21 (C1D21) in the responder (n = 5) versus nonresponder groups (n = 5), and adjustments were made for multiple comparisons. H Chord diagram illustrating cell–cell communication between monocytes and CD8 + T cells. I Circle plots showing the contribution of latent patterns to specific cell subsets in the MIF signaling pathway network between responder and nonresponder on C1D21. Statistic analyses are conducted by two-tailed Mann–Whitney U-test (C) or two-tailed Student’s t-test (E, F). In each box plot, the box represents the interquartile range, and whiskers represent minima and maxima. Source data are provided as a Source Data file. In nonclassical monocytes on C1D21, proinflammatory-related genes (TNF, IL1B) and chemokine-related genes (CCL3, CX3CR1) were significantly upregulated in responders (Fig. 6G). Interactome analysis between monocytes and CD8 + T cells revealed the Macrophage Migration Inhibitory Factor (MIF)-(CD74 + CD44) pathway as the most prominent signal, which has been shown to regulate monocyte migration and differentiation into an immunosuppressive phenotype (Fig. 6H)34. Among nonresponders, CD8 + CXCR3+ Tem cells provided the MIF pathway signal to non-classical monocytes, whereas no significant interaction between these subsets was found in responders (Fig. 6I). Durvalumab plus HAIC-FOLFOX in HCC mouse models To further investigate the functional role of MECOM+ malignant cells, we transfected RIL-175 cell line with MECOM overexpression or knockdown lentivirus and established orthotopic tumors in mice (Supplementary Fig. 17A–D). Treatment experiments showed that MECOM overexpression attenuated the anti-tumor efficacy of HAIC plus PD-L1 blockade, whereas MECOM inhibition enhanced the therapeutic response (Supplementary Fig. 17E–H). These results were consistent with the aforementioned snRNA-seq findings regarding the potential involvement of MECOM+ malignant cells in treatment resistance. In addition, we conducted treatment experiments using a murine HCC model driven by MYC overexpression and TP53 knockout to explore the potential synergistic effects, which is known for its resistance to immunotherapy (Supplementary Fig. 18A). After treatment, bioluminescent imaging showed significant tumor reduction in the combination therapy group, while tumor volumes in the monotherapy and control groups remained stable or increased (Supplementary Fig. 18B). Consistently, mice in the combination therapy group showed significantly improved survival compared to those in the monotherapy and control groups (Supplementary Fig. 18C). A significantly lower frequency of nonclassical monocytes (CD115+Ly6Clow) was observed in the combination therapy group compared to the monotherapy and control groups. Both the HAIC group and the combination therapy group showed an upward trend in the frequency of CD115+Ly6Chigh classical monocyte and CD3 + CD8α+ cytotoxic T cells compared to the control group (Supplementary Fig. 18D–G. Notably, the combination therapy group exhibited a significant increase in specific CD8 + T cell subsets, including GZMB + CD8 + , Ki-67 + CD8 + , and PD-1 + CD8 + T cells (Supplementary Fig. 18H). These findings align with the aforementioned peripheral blood scRNA-seq results, suggesting that combination therapy may influence peripheral monocytes, thereby enhancing anti-tumor efficacy. We further compared the anti-tumor effects of HAIC combined with anti-PD-L1 agents against standard systemic regimens (anti-PD-L1/VEGFA and anti-PD-L1/CTLA-4) in MYCOE/TP53KO murine models (Supplementary Fig. 18I). Bioluminescent imaging showed the modest tumor control in anti-PD-L1/VEGFA and anti-PD-L1/CTLA-4 groups relative to the control group, and the HAIC plus anti-PD-L1 group exhibited more pronounced tumor reduction (Supplementary Fig. 18J). Similarly, survival analysis indicated that all three treatment regimens significantly prolonged survival compared with controls, with a numerical trend toward greater survival benefits in the HAIC plus anti-PD-L1 group relative to both systemic treatment groups (Supplementary Fig. 18K). Discussion In this phase 2 study, we assessed the oncologic outcomes of durvalumab combined with HAIC-FOLFOX in advanced HCC with Vp3/4 PVTT, and explored biomolecular predictors associated with tumor response. The Durhope study demonstrated favorable survival outcomes, significant objective responses, and an acceptable safety profile using a combination of chemotherapy and anti-PD-L1 inhibitor in patients with severe portal invasion. Preliminary results indicated that tumor responders exhibited a higher abundance of infiltrating immune cells and stronger immune-malignant cell interactions. Peripheral scRNA-seq revealed expanded effector T-cell subsets and increased expression of cytotoxicity-related genes following treatment in responders. Prolonged tumor response duration is a significant clinical benefit of immunotherapy, due to its mechanism of reactivating anti-tumor immunity and maintaining immune surveillance. However, such long-term responses are rare in advanced HCC patients treated with ICIs-based systemic therapies, with only approximately 0–5.5% of patients achieving complete response35. Recent evidence suggests that locoregional therapies, such as transcatheter arterial chemoembolization (TACE) and stereotactic body radiotherapy (SBRT), can enhance the anti-tumor effects of ICIs, achieving durable tumor control in locally advanced HCC36–38. In the START-FIT phase 2 trial, Chiang et al.36 reported a median DOR of 25.7 months and a 3-year local control rate of 90.5% in a post hoc analysis of 29 patients who achieved CR after receiving combined ICIs with TACE or SBRT. Although 60.3% of patients had macorvascular invasion, the START-FIT study excluded individuals with Vp4 PVTT or distant metastasis39. Currently, prospective clinical evidence regarding outcomes of ICIs combined with locoregional therapeutic modalities (e.g., surgery, TACE, and SBRT) in patients with major PVTT remains scarce40. In the IMbrave150 study (atezolizumab plus bevacizumab vs. sorafenib), ICI-based regimen showed limited efficacy in the Vp4 PVTT subgroup, with a median OS of 7.6 months (vs. 5.5 months in sorafenib) and an objective response rate of 23% (vs. 13%)41. In the DurHope study, we focused on patients with Vp4 PVTT (80.0%) and heavy tumor burden, with a mean (± standard deviation) largest tumor diameter of 11.9 ± 4.3 cm. The combination of ICI and HAIC-FOLFOX showed a 1-year OS rate of 63.3%. Among 16 responders (53.3%), five (16.7%) attained CR with a median DOR of 15.7 months, and all five CR patients remained relapse-free at the final analysis. These preliminary observations indicated that FOLFOX chemotherapeutic agents may synergize with immunotherapy to prolong the duration of anti-tumor response, whereas the outcome of long-term tumor control should be interpreted with caution given the limited cohort size. Oxaliplatin, the key component of the FOLFOX regimen, can trigger ICD and enhance the release of DAMPs (e.g., calreticulin, HMGB1, and ATP)42, thereby promoting tumor antigen presentation and boosting T-cell-mediated anti-tumor response43. Moreover, FOLFOX chemotherapy can reduce tumor burden and deplete immunosuppressive cell populations (e.g., regulatory T cells and myeloid-derived suppressor cell44,45, fostering a more immune-active tumor microenvironment. ICIs further reinforce the anti-tumor effect by blocking PD-L1 and restoring T-cell function46, together providing a plausible explanation for the superior survival observed with HAIC-FOLFOX plus durvalumab in this study compared with HAIC-FOLFOX alone in the FOHAIC-1 trial (median OS: 15.4 months in the DurHope study population vs. 10.8 months in the high-risk subgroup of FOHAIC-1 study population)17. Biomarker analyses of tumor biopsies revealed specific T cells and macrophages subclusters enriched in responders, and they exhibited a more active immune state and stronger immune-tumor cell interactions. This is consistent with other exploratory biomarker studies evaluating regimens like atezo/beva, which showed enrichment of T cell and macrophage signatures in responders, correlating with better prognosis47. These findings highlight significant tumor microenvironment heterogeneity, with an “immune-inflamed” phenotype in responders and an “immune-excluded” phenotype in nonresponders. Additionally, we observed features of chemotherapy resistance and immune evasion in nonresponders, particularly involving a cluster of MECOM+ malignant cells. The MECOM+ cluster was enriched for pathways related to gene silencing, stemness, immune evasion, and metabolic processes like antioxidant and lipid metabolism. These pathways are plausibly associated with treatment resistance through enhanced DNA repair, anti-apoptotic signaling, and autophagy48–50. Moreover, the upregulation of angiogenesis-related pathways in MECOM+ cells raises the hypothesis that incorporating anti-angiogenic agents or tyrosine kinase inhibitors (TKIs) could potentially improve therapeutic efficacy. Dynamic scRNA-seq of paired PBMC samples revealed increased CD8+ effector T cells and higher cytotoxic gene expression in responders after therapy, indicating enhanced CD8 + T cells activation. These findings align with previous studies demonstrating correlations between CD8 + T cell characteristics and treatment outcomes in HCC patients receiving PD-1 inhibitors, alone or combined with TKIs51,52. We also identified distinct monocyte dynamics; responders showed decreased proportions of nonclassical monocytes and reduced expression of anti-inflammatory signature genes, indicating an immunoreactive state, while nonresponders experienced a slight increase in nonclassical monocytes. Similarly, increased frequencies of CD14 + FCGR3A+ nonclassical monocytes were linked to poor response to nivolumab in sorafenib-resistant HCC patients52. Previous studies have reported that peripheral monocytes can differentiate into tumor-associated macrophages, and infiltrating monocytes may modulate the immunosuppressive tumor microenvironment53–55. Early decreases in the monocyte-to-lymphocyte ratio have been associated with better nivolumab efficacy, potentially serving as a surrogate marker for determining ICIs therapy continuation56. Overall, these results suggested a potential relationship between dynamic changes in peripheral immune cells and treatment response, and whether these exploratory findings could serve as early predictors of therapeutic efficacy requires validation in independent cohorts. Safety is crucial in patients with Vp3/4 PVTT due to their heightened risk of gastrointestinal bleeding and liver failure. HAIC-FOLFOX uses a safe and effective dosage established in preclinical studies and phase I trials57,58, and has demonstrated with lower systemic toxicity compared to intravenous FOLFOX59,60. Based on favorable safety data from previous studies–8.2% grade ≥ 3 TRAEs with durvalumab alone (HIMALAYA9) and 20.3% with HAIC-FOLFOX (FOHAIC-117)–we utilized durvalumab in combination with HAIC-FOLFOX. In our study, the grade ≥ 3 TRAEs rate was 43.3%, with only one patient (3.3%) experiencing grade 2 treatment-related gastrointestinal bleeding, underscoring the importance of routine endoscopy examination. For patients with bleeding risks, such as esophagogastric varices, prophylactic interventions–including medication or endoscopic ligation–are strongly recommended. Collectively, durvalumab plus HAIC-FOLFOX showed a generally tolerable safety profile, even in patients with severe portal vein invasion. This study had certain limitations. First, the absence of a control group may limit the interpretation of efficacy specific to the combination therapy. Since standard treatment options for advanced HCC with Vp4 PVTT lack relevant data due to the exclusion of this population—except for atezolizumab plus bevacizumab showing limited efficacy—we did not include these treatments as control groups. Thus, the interpretation of our study outcomes relies on historical data from prior trials such as FOHAIC-1 and IMbrave150. Although such cross-trial comparisons suggest that durvalumab plus HAIC-FOLFOX might offer potential benefits, the baseline population characteristics (e.g., intrahepatic tumor burden, extent of vascular invasion, extrahepatic disease, and liver function) vary across studies. Moreover, we conducted in vivo experiments using a murine HCC model to further assess the regimen’s therapeutic effects. The results showed that the combination treatment provided superior anti-tumor activity compared with HAIC alone, PD-L1 blockade alone, or systemic agents (including PD-L1 plus CTLA-4 blockade and PD-L1 plus VEGFA blockade). While the MYC/TP53-driven model used in this study provides biological plausibility for the observed anti-tumor effects in an aggressive HCC context, its relevance to patients with Vp3/Vp4 PVTT is indirect, as it reflects an unresectable invasive stage but does not fully reproduce the biology of macroscopic PVTT. Consequently, the efficacy of the trial regimen in HCC with Vp3/4 VPTT requires further validation in future RCTs. Second, the patient population in our study was recruited from a single Chinese center and predominantly HBV-infected. Exploratory subgroup analyses in our study suggested that survival outcomes were numerically similar between patients with MASLD–HBV etiologies and those with HBV alone, whereas the small sample size limits definitive interpretation. In Western HCC patients, the FOLFOX regimen has shown favorable efficacy in non-HBV patients when combined with sorafenib in a phase 2 trial59, warranting further studies to investigate the feasibility of FOLFOX plus durvalumab in Western HCC population. Thirdly, the multiple comparisons in single-cell differential expression testing, pathway enrichment, and interactome analyses carry an inherent risk of false positives. This means that certain individual genes, pathways, or cell–cell interactions may still be prone to type I error, highlighting that our findings are hypothesis-generating and require further functional validation. In conclusion, the DurHope trial suggests that combining durvalumab with HAIC-FOLFOX may offer favorable anti-tumor efficacy and a generally tolerable safety profile in advanced HCC patients with Vp4 PVTT. Biomarker analyses identified potential predictors of response and resistance, providing insights into clinically relevant immune responses. These findings support durvalumab plus HAIC-FOLFOX as a feasible treatment strategy for severe portal invasion, warranting further validation in larger RCTs. Methods Study design and participants The open-label, single-arm, phase 2 DurHope trial was approved by the Ethical Review Committee of Sun Yat-sen University Cancer Center (approval number: B2021-111) and performed following the Declaration of Helsinki (enrolled from April 22, 2022, to July 6, 2023). All patients provided written informed consent before enrollment. Main inclusion criteria included a diagnosis of Barcelona Clinic Liver Cancer stage C HCC per European Association for the Study of the Liver (EASL) guideline, the presence of Vp3 or Vp4 PVTT, the presence of at least one measurable lesion according to Response Evaluation Criteria in Solid Tumors version 1.1 (RECIST v1.1), no prior systemic treatment, liver function classified as Child-Pugh class A–B7, Eastern Cooperative Oncology Group (ECOG) performance status of 0–1, and adequate organ function. The full list of inclusion and exclusion criteria is presented in trial protocol in the supplementary notes. This trial was registered with ClinicalTrials.gov identifier NCT04945720 (https://clinicaltrials.gov/study/NCT04945720; first posted on June 22, 2021). Study procedure Eligible patients received HAIC of FOLFOX regimens (oxaliplatin 130 mg/m2, leucovorin 200 mg/m2, fluorouracil 400 mg/m2, and fluorouracil 2,400 mg/m2) on day 1–3 plus durvalumab (1120 mg) intravenously on day 7 ( ± 2) of each 21-day cycle. HAIC-FOLFOX chemotherapy was given for up to 8 cycles. Once the patient becomes intolerant to chemotherapy or completes 8 cycles, durvalumab (1500 mg) would continue as maintenance therapy every 28 days until unacceptable toxicity, disease progression, or death. Radiologic response was assessed using abdominal CT/MRI and chest CT scan every 6–8 weeks, according to RECIST v1.1 criteria. Complete and partial responses were required to be confirmed with a follow-up scan at the next scheduled assessment following the initial documentation. Details of HAIC procedure and criteria of FOLFOX dose modification or discontinuation were described in our previous studies61,62. The decision to delay or discontinue durvalumab adhered to local standards of care and the approved product label. Baseline gastroscopy was mandatory for all enrolled patients to assess the presence and severity of gastroesophageal varices. Generally, patients without recent variceal bleeding (<6 months) were required to receive gastroscopy within 3 months before treatment initiation; those who experienced variceal bleeding (<6 months) needed to undergo gastroscopy within 1 month. In accordance with international gastroesophageal varices management guidelines and local standards of care63,64, patients with absent or mild varices were eligible for study participation and were routinely monitored for coagulation status during each treatment cycle; patients with moderate varices underwent prophylactic management (non-selective β-blockers or endoscopic ligation) before treatment initiation; and those with severe varices underwent variceal management and symptomatic treatment followed by re-evaluation within the one-month screening window to determine eligibility for study enrollment. Follow-up endoscopy was performed every 4 HAIC courses (i.e., approximately every 3 months) and every 3–6 months thereafter during durvalumab monotherapy. Newly developed gastroesophageal varices were handled following similar management principles. In case of severe varices occurring during the study period, patients were allowed to receive adequate variceal management and re-evaluation for continuing the trial regimen. Assessments and endpoints The primary outcome was the 1-year OS rate, defined as the proportion of patients alive at 12 months. Secondary outcomes included PFS (the interval from treatment initiation to either disease progression or death from any cause), objective response rate (the proportion of patients achieving a complete or partial response) based on the RECIST v1.1 criteria, and safety profiles according to Common Terminology Criteria for Adverse Events, version 5.0 (CTCAE 5.0). The prespecified date for final analysis was set as 5th July, 2024, corresponding to one year after the enrollment of the last patient. Exploratory analyses focused on identifying biomarkers associated with treatment efficacy. All human samples used for exploratory biomarker analysis were collected prospectively following the study protocol. Sample collection Enrolled patients optionally underwent tumor biopsy before study treatment. The biopsy specimens were either frozen for subsequent transcriptomic analysis or preserved in formalin and processed into 4-μm-thick paraffin sections for IHC analysis. Paired peripheral blood samples were collected before treatment and on the 21st day of each treatment cycle. PBMCs were isolated from whole blood through density gradient centrifugation, and scRNA-seq was performed on PBMC before treatment and on C1D21. Pre-processing and initial clustering of single-cell RNA sequencing data Frozen biopsy samples and PBMCs samples were processed following standard protocols. Cell or nucleus suspensions were filtered using a 40 µm nylon cell strainer, stained with 0.4% Trypan blue to assess viability, and then loaded into the Chromium Single-cell Controller (10X Genomics) to generate gel beads in emulsion. After exposure to cell lysis buffer, the beads were retrieved for reverse transcription, followed by cDNA amplification and library preparation using the Single Cell Reagent Kit (10X Genomics). Sequencing was performed on an Illumina NovaSeq platform with 150 bp paired-end reads (Shanghai Biotechnology Corporation, Shanghai, China). Demultiplexing, barcode processing, alignment to the human reference genome (hg38), and generation of gene expression matrices were conducted using CellRanger (v5.0.0). Subsequent analysis of gene expression data was performed using Seurat (v4.0). Cells with over 5% mitochondrial gene expression or more than 4500 detected genes were filtered out, and doublets were identified and removed using the demuxlet package. Batch correction was applied using the FindIntegrationAnchors function, followed by principal component analysis for dimensionality reduction. Unsupervised clustering was performed using the shared nearest neighbor graph, based on the top 50 principal components (resolution = 0.2), with uniform manifold approximation and projection (UMAP) used for visualization. Characterization of each subcluster of scRNA-seq data To identify marker genes, differentially expressed genes for each cluster were selected using the FindAllMarkers function, applying the following parameters: log fold change > 0.25, min.pct > 0.1 (the minimum fraction of test genes detected in cells from other clusters), and false discovery rate (FDR)-adjusted P < 0.05 (Supplementary Data 1). To assess quantitative changes, we calculated and compared the proportion of each cell type between subgroups of responders and nonresponders. Fold enrichment in cell type proportions was determined by dividing the fraction of each cell type at C1D21 by that at C1D1, followed by log2 transformation. Copy number variation estimation and identification of malignant cells To identify malignant cells from epithelia, we performed the CopyKAT algorithm (v0.1.0) to estimate the copy number variation (CNV), including amplifications and deletions in the single-cell level based on the scRNA-seq data65. The fibroblasts and endothelial cells were selected as normal reference cells, with default parameters applied. The CNV score for each cell was determined by summing the calculated CNV values for all genes in that cell. Differential expression and gene ontology enrichment analysis To identify marker genes for each cluster, differential expression analysis was performed using the function FindAllMarkers and FindMarkers to detect positive enriched (overexpressed) markers in each population. Genes with |log2 fold-change | ≥ 0.25, adjusted P < 0.05, and detected in ≥ 10% of cells per group were defined as differential expression genes. P values from differential expression testing were corrected for multiple comparisons using the Benjamini–Hochberg (BH) procedure. Gene Ontology (GO) and pathway enrichment analyses of differential expression genes were performed using the “clusterProfiler” package. GO biological process terms with BH-adjusted P values < 0.05 were considered significantly enriched. Gene module enrichment analysis To estimate the proliferation, immune surveillance, and immune escape signatures in malignant cells, we calculated enrichment scores for each cell using the AddModuleScore function in Seurat66. Similarly, we assessed cytotoxicity and exhaustion in CD8 + T and NK cells by computing cytotoxic and exhaustion scores based on canonical markers. Using the same approach, we calculated pro-/anti-inflammatory phenotypes for macrophages/monocytes, antigen presentation and phenotype scores for macrophages, costimulatory scores for CD4 + T cells, and exhaustion scores for Tregs. Details of the marker genes used in each gene set were provided in Supplementary Table 11. The mean module score for each cell cluster (≥ 10 cells) was calculated to represent the signature level in individual samples. Pseudotime trajectory analysis and cell-cell interaction analysis The cell lineage trajectory was inferred using the Monocle2 R package67. The differentiation trajectory was generated using Monocle’s default parameters, following dimensionality reduction and cell ordering. Intercellular communication among different cell types was analyzed using the CellChat package (version 0.5.0)68. A CellChat object was constructed from the normalized count data using the createCellChat function. Overexpressed genes and interactions between cell types were then identified using the identifyOverExpressedGenes and identifyOverExpressedInteractions functions. Additionally, the computeCommunProb, computeCommunProbPathway, and aggregateNet functions were employed, utilizing the CellChatDB database. PD-L1 IHC Formalin-fixed paraffin-embedded (FFPE) tumor samples were sectioned into 4-µm slides and baked at 65 °C for 4 h. The slides underwent dewaxing and rehydration using xylene and graded ethanol solutions. Heat-induced antigen retrieval was performed with EDTA (pH 9.0, ZSGB-BIO, China, Cat# ZLI-9069). To block endogenous peroxidase activity and reduce nonspecific staining, 0.3% hydrogen peroxide and 10% goat serum (Beyotime, China, Cat# C0265) were applied. Primary antibodies were incubated overnight at 4 °C and then rewarmed for 30 minutes. Following TBST washing, slides were incubated with an HRP-conjugated secondary antibody (Beyotime, China, Cat# A0216 and A0208) for 30 minutes. The antigen-antibody reaction was visualized using 3,3-diaminobenzidine (Elabscience, China, Cat# E-IR-R101). The slides were then counterstained with hematoxylin (JLB, China, Cat# JLM-111), dehydrated, cleared in xylene, and sealed. PD-L1 IHC assay was conducted using anti-PD-L1 monoclonal antibodies, and results were assessed based on PD-L1 expression in both tumor cells (TC) and tumor-infiltrating immune cells (IC). The TC score indicated the percentage of PD-L1-positive TCs, while the IC score represented the tumor area occupied by PD-L1-positive ICs. Multiplex immunohistochemistry FFPE human sections were collected for analysis. Multiplex immunofluorescence staining was performed using the PANO 7-plex IHC kit (Panovue, China; Cat# 10004100050) following the manufacturer’s instructions. Primary antibodies were appropriately diluted, and incubation was carried out under optimized conditions for each antibody, with additional details available in Supplementary Table 12. Tissue sections were labeled with Opal fluorophores 480, 520, 570, and 650, and incubated at room temperature for 10 minutes. Antigen retrieval was conducted in a microwave using EDTA buffer (pH 9.0, ZSGB-BIO, China, Cat# ZLI-9069) for 15 minutes. Imaging was performed using the Vectra Polaris Automated Quantitative Pathology Imaging System (Akoya Biosciences, USA), and positive staining quantification, as well as spatial cell analysis, were conducted using HALO image analysis software (Indica Labs, USA, version 2.3.2089.34). Cell culture The murine HCC cell line RIL-175 (kindly provided by Prof. Jun Yu’s group, The Chinese University of Hong Kong) was cultured in DMEM (GIBCO, USA; Cat# C11995500BT) supplemented with 10% fetal bovine serum (FBS, Bio-Channel, China, Cat# BC-SE-FBS07C) and 1% penicillin–streptomycin (NCM, China, Cat# C100C5) at 37 °C in a humidified incubator with 5% CO2. Cells were routinely screened for mycoplasma contamination using a single-step PCR assay and were used for experiments only when they tested negative. Establishment of stable MECOM knockdown and overexpression cell lines Short-hairpin RNA targeting Mecom (shMecom) and the corresponding control vector (shCtrl), as well as the MECOM overexpression plasmid (Mecom) and its empty-vector control (Ctrl), were packaged into lentiviral particles in HEK293T cells. Briefly, HEK293T cells were co-transfected with the transfer plasmids together with the packaging plasmids psPAX2 and pMD2.G using a commercial transfection reagent (CYTOCH, China, Cat# CT0005) according to the manufacturer’s protocol. After 24 h, culture supernatants containing lentiviral particles were collected, filtered, and used to infect RIL-175 cells in the presence of polybrene (MCE, China, Cat# HY-K6005). Infected cells were selected with puromycin (MCE, China, Cat# HY-B1743) to establish stable lines. MECOM knockdown or overexpression efficiency was confirmed by western blotting. Western blotting Cells were washed three times with PBS and lysed in ice-cold lysis buffer (NCM, China, Cat# WB3100) for 30 minutes on ice. Lysates were sonicated and centrifuged at 10,000 × g for 15 min at 4 °C, and the supernatant containing total protein was collected. Protein concentrations were determined using a BCA assay (CWBIO, China, Cat# CW0014S) with BSA as the standard, and samples were adjusted to equal protein amounts. Proteins were mixed with loading buffer, boiled at 100 °C for 10 min, separated by SDS–PAGE, and transferred onto PVDF membranes. Membranes were blocked with 5% BSA and incubated overnight at 4 °C with primary antibodies against MECOM or α-tubulin. After incubation with HRP-conjugated secondary antibodies (Beyotime, China, Cat# A0216 and A0208), bands were visualized using an enhanced chemiluminescence (ECL, NCM, China, Cat# P10300) detection kit. The membranes were then stripped with a stripping buffer and subsequently reblocked, incubated with additional primary and secondary antibodies, and detected again by ECL following the same procedure. Mice used for in vivo experiments Wild-type male C57BL/6 mice were purchased from the Animal Facility of Sun Yat-sen University Cancer Center, and used for treatment experiments involving arterial FOLFOX regimen and/or monoclonal antibodies. All animal procedures were conducted following institutional guidelines and approved by the Sun Yat-sen University Cancer Center ethics committee (approval number: L025501202309007). The mice were housed under specific pathogen-free conditions, five per cage, with a 12-hour light-dark cycle and ad libitum access to food and water. Experiments were performed on mice aged 6 to 8 weeks. Mice were sacrificed at designated time points, and their livers and peripheral blood were collected for analysis. Mice were monitored three times per week for general health and euthanized early based on defined endpoint criteria, including ascites, lethargy, or other signs of sickness or distress. If animals appeared moribund or the diameter of the tumors reached 15 mm, the mice were sacrificed. In some cases, the maximal tumor burden permitted has been exceeded the last day of measurement and the mice were immediately euthanized. Hydrodynamic gene transfer HCC mouse model and treatment The three gene constructs included: (1) a human c-Myc gene linked to luciferase, (2) a gRNA guide targeting p53 alongside the Cas9 nuclease, and (3) a Sleeping Beauty transposase expression cassette. All constructs were verified by nucleotide sequencing, and vector integrity was confirmed by restriction enzyme digestion. Then hydrodynamic gene transfer was performed by rapidly injecting a solution containing the gene constructs through the lateral tail vein of C57BL/6 J mice. After 3 weeks, noninvasive in vivo multispectral fluorescence imaging was conducted on all mice using the IVIS Spectrum system (PerkinElmer, USA) with Living Image software (version 4.4) to evaluate tumor number and extent. The mice were randomly assigned to treatment and control groups, with each group having an equivalent average luciferase signal. For the intra-arterial oxaliplatin (MCE, China, Cat# HY-17371) and fluorouracil (MCE, China, Cat# HY-90006) injections, the following procedure was performed: (1) the mice were anesthetized with avertin and positioned in the lateral recumbent position, followed by a 1 cm incision below the rib cage; (2) under microscopic guidance, sterile sutures were used to ligate the pancreatic-splenic artery and splenic-gastric artery to temporarily occlude blood flow to the abdominal artery; (3) the abdominal artery was punctured with a syringe for drug infusion; (4) after drug perfusion was completed, the syringe was removed; and finally, (5) the surgical incision was sutured and disinfected. The anti-PD-L1 (STARTER, China, Cat# S0B0593)/CTLA-4 (MCE, China, Cat# HY-P99132)/VEGFA antibody (MCE, China, Cat# HY-P990148) or saline was administered every 3 days, while the FOLFOX regimen was administered every 7 days for a total of 2 doses. After each treatment, dynamic multispectral fluorescence imaging was performed on all mice to track changes in fluorescence signals. Three days after the final dose, the mice were euthanized using carbon dioxide asphyxiation, followed by necropsies. Body and liver weights were recorded, and peripheral blood samples were collected for flow cytometry analysis. Orthotopic mouse model and treatment Stable RIL-175 cells (1 × 106 cells per mouse) were resuspended in 25 µL PBS and mixed 1:1 with Matrigel (MCE, China, Cat# HY-K6005). The cell suspension was injected under the hepatic capsule to establish orthotopic liver tumors. Seven days after inoculation, tumor-bearing mice were randomized into treatment and control groups. The anti-PD-L1 antibody or saline was administered every 3 days for a 7-day period, and the FOLFOX regimen was administered as a single dose. At the experimental endpoint, mice were sacrificed, liver tumor issue as well as total body weight were recorded, and the ratio of liver tumor weight to body weight was used for data analysis. Statistical analysis and sample size calculation The efficacy and safety analysis were performed in patients who received at least one dose of the trial regimen. Survival outcomes were analyzed using the Kaplan-Meier method and compared by log-rank tests. Group differences in continuous variables were assessed using two-tailed Student’s t test or Mann-Whitney U test (for two-group comparisons) and one-way ANOVA analysis (for multiple-group comparisons), and differences in categorical variables were evaluated using the χ2 test or Fisher exact test. Statistical analyses were performed using SPSS version 26.0 (IBMCorp., Armonk, NY), Prism version 9.0 (GraphPad Software, SanDiego, CA), and R version 3.4.2 software. The differences are considered statistically significant at P value < 0.05 or FDR-adjusted P value < 0.05 for analyses involving multiple comparisons. To determine the appropriate sample size, we used a one-stage binomial design with a one-sided alpha level of 0.05 and statistical power of 80%. Previous studies have reported a median OS of approximately 7.6 months in high-risk HCC patients treated with atezolizumab plus bevacizumab, corresponding to an estimated 1-year OS rate of 35%6,69. We hypothesized that the combination of hepatic arterial infusion chemotherapy (HAIC) and durvalumab would improve the 1-year OS rate to 60%. Assuming an anticipated dropout rate of 10%, the target sample size was set at 30 patients. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Reporting summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Reporting summary Transparent Peer Review file Source data source data source data