Exploring clinically anchored therapeutic stratification and tumor-stroma heterogeneity using patient-derived 3D bioprinted primary liver cancer models

Abstract

Background

The marked heterogeneity of primary liver cancer (PLC), together with the dynamic influence of the tumor microenvironment (TME), remains a major challenge to precision oncology. As currently available therapies provide limited benefit for many patients and robust preclinical platforms for rapid, patient-specific treatment stratification are lacking, there is an urgent need for models that faithfully preserve individual tumor architecture and cellular diversity.

Methods

We established three-dimensional bioprinted primary liver cancer (3DP-PLC) models comprising 61 patient-derived monocultures and 34 co-culture models with patient-matched cancer-associated fibroblasts (CAFs). To explore the correlation between ex vivo models and clinical drug response, we developed a clinically anchored hybrid stratification framework for response classification. In parallel, the 3DP-PLC/CAF co-culture biobank was used to investigate stromal regulation of therapeutic response.

Results

These 3DP-PLC constructs preserved the histological features, biomarker expression patterns, mutational landscapes, and transcriptomic profiles of their corresponding parental tumors. Drug-sensitivity profiling across the biobank revealed substantial intertumoral heterogeneity in therapeutic responses. The clinically anchored hybrid stratification framework integrating ex vivo pharmaceutical profiles with external clinical benchmark data enabled interpretable response classification and demonstrated clinical relevance in patients receiving neoadjuvant or adjuvant targeted therapy. In parallel, the 3DP-PLC/CAF co-culture biobank, combined with single-cell transcriptomic analysis, recapitulated fibrous ring-like architecture and revealed CAF-associated drug-resistant states.

Conclusions

These findings support 3DP-PLC as a high-fidelity and scalable platform for personalized therapeutic stratification and mechanistic investigation of tumor–stroma interactions in PLC.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12943-026-02751-1.

Background

Primary liver cancer (PLC) remains a leading cause of cancer-related death worldwide and comprises predominantly hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (iCCA) [1]. Although surgical resection, locoregional therapies, molecularly targeted agents, and immune checkpoint inhibitors have expanded the therapeutic landscape, [2, 3] durable clinical benefit remains limited for many patients with advanced disease. Multikinase inhibitors, including sorafenib [4], lenvatinib [5], regorafenib [6], and apatinib, [7, 8] continue to serve as important systemic options, but their efficacy is limited by profound intertumoral and intratumoral heterogeneity. [9, 10] Molecular diversification across HCC, [11, 12] iCCA [13, 14], and combined hepatocellular cholangiocarcinoma (CHC) [15] further complicates treatment selection and contributes to variable therapeutic sensitivity and resistance. These problems highlight the need for rapid, patient-derived systems capable of anticipating therapeutic response and guiding treatment selection before therapy begins.

Beyond tumor intrinsic diversity, the tumor microenvironment (TME) actively governs liver cancer progression, metastasis, and therapeutic response [16–18]. Among stromal populations, cancer-associated fibroblasts (CAFs) constitute a dominant compartment and are thought to originate, at least in part, from hepatic stellate cells [19, 20]. Through cytokine secretion, extracellular matrix (ECM) remodeling, and sustained paracrine signaling, CAFs establish spatially and functionally heterogeneous stromal niches [21]. Studies have shown that CAFs adjacent to tumor cells are associated with TGF-β activation and collagen deposition, while more distal CAFs promote hyaluronic acid deposition and reshape the TME through the sustained release of IL-6 and other inflammatory mediators [22, 23]. Notably, a fibrous ring (FR) formed by CAFs at the tumor margin has been associated with paradoxical prognostic outcomes in PLC [24]. However, the mechanisms that drive FR formation, the features of spatially distinct CAF subsets, and their contribution to drug resistance remain insufficiently defined. Faithfully modeling these tumor-stroma architectures is therefore essential for resolving how microenvironmental heterogeneity shapes therapeutic response.

Existing preclinical models incompletely address these problems. Conventional two-dimensional (2D) cultures lack ECM complexity and fail to reflect the biological behavior observed in patients [25]. Patient-derived xenograft (PDX) models preserve tumor heterogeneity and can inform drug sensitivity, but their low establishment efficiency, high cost, and limited clinical turnaround restrict real-time therapeutic application [26]. Patient-derived tumor organoids (PDTOs) have improved the modeling of tumor heterogeneity and drug response in PLC [27–29], yet their translation remains hindered by protocol variability, long-term culture complexity, Matrigel batch effects, and manual inconsistencies. Moreover, drug-response data from patient-derived tumor models often lack standardized computational frameworks that convert ex vivo pharmacologic profiles into clinically interpretable response categories [30]. Three-dimensional bioprinting (3DP) offers a complementary strategy by enabling rapid, spatially controlled assembly of living tumor and stromal cells within tunable bioinks. Accumulating evidence supports the potential of 3DP for high-throughput pharmacological screening and therapeutic prediction [31–36]. For short-term therapeutic application, 3D bioprinting offers several advantages over conventional organoid generation. Rather than relying on spontaneous aggregation and variable growth of patient-derived cells, bioprinting enables controlled deposition of cells within a defined bioink, with reproducible cell density, construct geometry, and spatial organization [37]. This standardization is particularly important for ex vivo drug testing, where differences in model size, cell number, and cell distribution may influence pharmacological readouts. [38, 39] We previously established patient-derived 3D-bioprinted PLC models that preserved key histologic features and mutational landscapes of parental tumors [40]. Several studies have shown that 3DP technology allows for the precise spatial arrangement of tumor and stromal components, [41, 42] providing a unique opportunity to reconstruct native tumor-stroma architectures, such as the fibrous ring observed in PLC.

In this study, we developed the 3DP-PLC platform consisting of 61 patient-derived monocultures and 34 direct co-culture models established with patient-matched CAFs. The monoculture platform retained key biomarker and genomic features of parental tumors and revealed marked interpatient heterogeneity in drug sensitivity. To enhance clinical interpretability, we developed a hybrid response stratification framework that integrates pharmaceutical profiles with external clinical benchmark data, enabling clinically anchored classification of drug sensitivity. Subsequently, follow-up data further confirmed distinct disease-free survival (DFS) patterns across subgroups, thereby supporting the clinical relevance of this stratification. Building on this monoculture platform, we further established 3DP co-culture models that recapitulate the fibrous ring-like architecture, allowing for the investigation of tumor-stroma interactions, their impact on drug response, and associated single-cell transcriptional changes. Collectively, these findings validate the 3DP-PLC model as a powerful platform for personalized treatment strategies and for exploring tumor-stroma interactions in PLC.

Methods

Please see the supplementary information for detailed materials and methods.

Tissue specimen acquisition and clinical data collection

Primary liver cancer specimens were procured from patients undergoing curative hepatectomy who had been enrolled in an investigator-initiated, observational clinical trial (ClinicalTrials.gov identifier: NCT07437859) conducted at Peking Union Medical College Hospital. A postoperative histopathological examination confirmed the diagnosis of PLC in every case. Comprehensive clinicopathological characteristics of the patients are cataloged in Table S1 and S2. Eligible patients were required to meet the following criteria: (1) age 18 years or older; (2) radiologically suspected or histologically confirmed primary liver cancer, including hepatocellular carcinoma or intrahepatic cholangiocarcinoma; (3) suitability for curative-intent hepatectomy based on multidisciplinary evaluation; (4) Child-Pugh class A liver function; (5) Eastern Cooperative Oncology Group performance status of 0 or 1; (6) availability of fresh tumor tissue sufficient for 3DP-PLC model generation and downstream analyses; and (7) provision of written informed consent. Patients were excluded if they had metastatic liver tumors, recurrent liver cancer, insufficient or severely necrotic tumor tissue for viable cell isolation, uncontrolled infection or severe systemic comorbidity, incomplete essential clinical information, or withdrawal of informed consent.

Resected tumor tissues were immediately immersed in ice-cold Advanced DMEM/F-12 (Gibco, Billings, MT, USA) supplemented with 0.2% Primocin (InvivoGen, San Diego, CA, USA), 1% antibiotic–antimycotic (Gibco) and 10 µM Y-27,632 (Sigma–Aldrich, St Louis, MO, USA), and transported on ice to the laboratory without delay. Each region of the tissue sample was dissected into several parts according to size for primary cell isolation, WES, RNA sequencing or histopathological analysis.

Tumor isolation and culture of cancer-associated fibroblasts

Tissue fragments designated for tumor cell isolation were sequentially rinsed in ice-cold PBS containing 3% (v/v) antibiotic–antimycotic (Thermo Fisher Scientific; 3 × 5 min). While maintained on ice, the tissue was dissected into < 1 mm³ pieces under sterile conditions. The fragments were collected in a 15 mL centrifuge tube, centrifuged at 100 × g for 3 min, and subjected to enzymatic dissociation at 37 °C for 30–60 min in Advanced DMEM/F-12 supplemented with 1.5 mg mL⁻¹ collagenase IV, 1 mg mL⁻¹ dispase II, 100 µg mL⁻¹ DNase I, 10 µM Y-27,632, 1% antibiotic–antimycotic and 0.2% Primocin on an orbital shaker. Manual agitation was performed every 10 min to enhance digestion. The resultant cell suspension was passed through a 70-µm strainer, centrifuged at 300 × g for 5 min at 4 °C, and, when erythrocytes were evident, incubated for 10 min at 4 °C in 3 mL erythrocyte-lysing buffer (Servicebio) after resuspension in 1 mL ice-cold medium. Following a second centrifugation (300 × g, 5 min, 4 °C), viable cells were counted and resuspended in 3D culture medium for subsequent bioprinting. For CAF isolation, cells were cultured in 3DP culture medium containing Advanced DMEM/F12 (Gibco), penicillin/streptomycin (1%), 2 mM GlutaMAX (Gibco), 10 mM HEPES (Gibco), 1.25 mM N-acetyl-L-cysteine (Sigma-Aldrich), 10 nM Recombinant human (Leu15)-Gastrin I (Sigma-Aldrich), 50 ng mL − 1 recombinant EGF (PeproTech, Rocky Hill, NJ, USA), 100 ng/mL Recombinant Human FGF-10 (PeproTech, Rocky Hill, NJ, USA), 25 ng/mL Recombinant Human HGF (PeproTech, Rocky Hill, NJ, USA), 5 µM A8301 (Tocris Bioscience, Bristol, UK), 10 µM Y-27,632 (Sigma-Aldrich), 25 ng mL − 1 Recombinant Human Noggin (PeproTech, Rocky Hill, NJ, USA), 500 ng mL − 1 Recombinant Human R-Spondin1 (Novoprotein), 100 ng mL − 1 Recombinant Human Wnt3a V3 (Novoprotein), 1× B27 supplement (Life Technologies, Carlsbad, CA, USA), 1× N2 supplement (Life Technologies), 10 µM Forskolin, 10mM Nicotinamide (Sigma-Aldrich).

Primary CAFs were generated by seeding the freshly isolated tumor-cell suspension in Advanced DMEM/F-12 (Gibco) supplemented with 20%(v/v) fetal bovine serum (Gibco), 2 mM GlutaMAX (Gibco), 10 mM HEPES (Gibco), 10 ng mL⁻¹ recombinant human FGF-10 (Gibco) and 1% penicillin–streptomycin (Gibco). The success rate for CAF culture establishment was 90%. The cultures were then incubated undisturbed for 1 week to facilitate selective adhesion and migration of fibroblasts. Thereafter, medium was exchanged twice weekly, and the first passage was performed 14–21 days after seeding. CAF identity was verified by morphology and positive immunofluorescence for α-SMA and vimentin. All experimental assays were conducted with cells between passage 4 and passage 6.

Bioprinting

Various compositions of biomaterials were tested to formulate the most biocompatible bioink for liver cancer cells. Finally, we chose a composite bioink mixed with the cell suspension, 5% (w/v) GelMA (Engineering For Life), 10% (w/v) LAP (Engineering For Life), 0.5% (w/v) HAMA (Engineering For Life) and 0.5% (w/v) AlgMA (Engineering For Life), resulting in a final concentration of tumor cells of 8 × 10⁶ cells mL^− 1. The mixture was loaded into a 3 mL syringe (BD, Franklin Lakes, NJ, USA) equipped with a 23G needle. The syringe was incubated at 4 °C for 20 min to facilitate gelation and then loaded into a 3D bioprinter (SUNP Biotech) for printing, with the nozzle and chamber maintained at 15 °C and 10 °C, respectively. The 3DP-PLC construct was configured as a stratified lattice (6 mm × 6 mm × 0.8 mm) comprising four layers (layer height 0.2 mm) with a line width of 0.8 mm. Extrusion was performed at 1.5 mm³ s⁻¹ while the print head translated at 6 mm s⁻¹. 3DP structures were constructed in 48-well plates on the chamber platform, and the printing process took approximately 2 h. Following bioprinting, the constructs were photo-cross-linked under 405 nm light (20 mW cm⁻²) for an optimized exposure time of 20 s to induce GelMA–HAMA polymerization. Thereafter, the 3DP models were transferred to culture medium at 37 °C under 5% CO₂. The medium was exchanged at 72-hour intervals throughout the culture period.

Drug testing

Drug testing in the 3DP models was conducted on day 5. Nine antitumor agents used in clinical practice, five chemotherapeutic drugs including 5-FU (Selleck), Cisplatin (MCE), Gemcitabine (MCE), Epirubicin (MCE) and Oxaliplatin (Selleck) and four tyrosine kinase inhibitors including Lenvatinib (Selleck), Regorafenib (Selleck), Sorafenib (Selleck), and Apatinib (Selleck) were assembled for high-throughput drug sensitivity screening in 3DP models after bioprinting. All drugs were dissolved in dimethyl sulfoxide (DMSO) or sterile deionized water to yield concentrated stock solutions and maintained at − 20 °C under desiccated conditions until use. Culture medium was aspirated and replaced with fresh culture medium containing six concentrations of each drug. Targeted agents were tested at 50, 10, 5, 1, 0.1, and 0 µM (vehicle control), whereas chemotherapeutic agents were tested at 100, 50, 10, 1, 0.1, and 0 µM (vehicle control). Final DMSO concentration did not exceed 1% (v/v) in any condition. Cell viability was quantified using the CellTiter-Glo 3D assay (Promega, Madison, WI, USA) following the manufacturer’s protocol after 72 h of drug exposure, and results were normalized to corresponding vehicle-treated controls. The luminescence values were recorded on a microplate reader (Synergy H1, USA). Each concentration was assayed in a minimum of three technical replicates. Pharmacodynamic parameters (AUC, IC50 and Log₁₀IC50) for each drug in 3DP models were calculated using GraphPad Prism v10 and were summarized in Table S3.

To assess whether immune checkpoint inhibitor exposure affected the viability of 3DP-PLC models during the short-term screening window, pilot antibody-treatment experiments were performed using tislelizumab. 3DP-PLC constructs were treated with tislelizumab at 0, and 100 µg/mL for 72 h. After treatment, overall viability was assessed by C-AM/PI staining, and flow cytometry quantified epithelial populations and CD107a-positive CD8 T cells.

Externally Calibrated Jenks-Merging Stratification (ECJMS)

The ECJMS framework was designed as an outcome-independent threshold-anchoring procedure rather than as a supervised patient-level prediction model. External clinical trial response compositions were used only as population-level anchors to select among candidate AUC partitions generated from ex vivo drug-response distributions. No patient-level clinical outcomes from the present cohort were used to derive the AUC cutoffs. Patient-level clinical response and postoperative DFS were reserved for subsequent exploratory evaluation of clinical relevance.

Individual AUC values were normalized to the maximum area observed within their respective concentration ranges, yielding decimal AUCs (ranging from 0 to 1) for cross-sample comparison. The normalized AUC values of the 3DP-PLC models were then ranked and categorized using the Jenks natural breaks algorithm, which identifies all cutoff values by minimizing within-group variance and maximizing between-group variance. For each candidate partition, models were initially stratified into multiple subgroups, which were subsequently merged with adjacent subgroups to form three consolidated groups. The proportion of patients in each consolidated group was calculated. Pre-specified criteria were applied to reduce overfitting and define the candidate set for subsequent analyses: (1) goodness of variance fit (GVF) ≥ 90%; (2) incremental GVF gain ≥ 1%; and (3) a minimum group-size constraint.

During the initial exploration, all partitioning schemes with K ≥ 3 were evaluated. Since K = 3 did not meet the pre-specified adequacy criteria (GVF ≥ 90%) and lacked a meaningful contiguous merging step, it was excluded from the subsequent merge-based bootstrap stability analysis. For the retained solutions, the subgroup with the lowest AUC values was designated as Sensitive (S), the intermediate subgroup as Relatively sensitive (RS), and the subgroup with the highest AUC values as Resistant (R) [43].

Subsequently, the class-prevalence vectors derived from all retained solutions were projected into three-dimensional space and benchmarked against RECIST 1.1 response categories reported in external clinical studies, namely complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD), with S mapped to ORR, RS to SD, and R to PD [44]. To ensure comparability between the external clinical benchmark and the three-category composition used in this study (ORR/SD/PD), the external anchor was recalculated after excluding unevaluable cases [6, 7, 45, 46]. The external anchor points for the four targeted therapies are shown in Table. S5. The discrepancy between the two was quantified using Euclidean distance, and the cutoff pair with the smallest distance was selected as the final lower cutoff (S–RS cutoff) and upper cutoff (RS–R cutoff). The formula is as follows:

To assess the internal stability of cutoff selection, we performed bootstrap full-pipeline resampling [47]. For each sample, we calculated the probabilities of assignment to the S, RS, and R classes across bootstrap resamples, using the maximum assignment probability to evaluate sample-level stability. Samples with a maximum assignment probability < 0.80 were classified as grey-zone cases, indicating uncertain class membership near the class boundaries. To further evaluate sampling uncertainty in the external benchmark, we extended the external benchmark from a fixed point estimate to a Dirichlet posterior [48, 49]. If the optimal cutoff pair obtained under the Dirichlet-based benchmark showed little or no shift compared to the point-estimate benchmark, the selected cutoffs were considered robust against benchmark uncertainty.

Because the external clinical response benchmarks used for ECJMS calibration were derived from HCC studies, the calibration was interpreted as HCC-focused. The iCCA-derived models were retained for descriptive presentation of model establishment and drug-response heterogeneity but were not used to support entity-level inference for iCCA. A sensitivity analysis was performed by repeating the ECJMS calibration after excluding the iCCA-derived models. The association between drug sensitivity testing and clinical outcome was evaluated by using the Jenks classifier implemented in the classInt R package and Euclidean distance to stratify normalized AUC values into optimal response groups. All statistical analyses were carried out in R v4.3.2 (www.r-project.org).

Correlation between 3DP-PLC models and clinical outcomes

Following multidisciplinary team discussion, patients were administered standard therapy based on clinical guidelines for primary liver cancer, adjusted to individual clinical conditions and tumor-related characteristics. Some participants received neoadjuvant targeted treatment before resection, whereas others proceeded directly to surgery and, when indicated, were offered adjuvant targeted treatment after surgery. For patients with neoadjuvant targeted treatment, clinical response to treatment was evaluated based on RECIST 1.1 criteria independently by three board-certified radiologists. TKI efficacy in PLC patients was graded as CR, PR, SD, or PD according to RECIST 1.1. Based on the established correlation with ex vivo drug sensitivity, CR/PR and SD were classified as drug-sensitive and relatively sensitive responses, respectively. Together, these two categories constitute disease control. In contrast, PD was defined as indicative of drug resistance, which corresponds to disease progression.

In accordance with standard clinical practice, postoperative management for patients with PLC was individualized by the treating physician, with regimens selected based on each patient’s unique clinical presentation. Clinical treatment data were retrospectively collected, with DFS defined as the primary endpoint for assessing postoperative outcomes. DFS was calculated from the date of adjuvant lenvatinib initiation to the first occurrence of radiologically confirmed intrahepatic or extrahepatic recurrence, appearance of a new lesion, or death from any cause. Patients without an event were censored at the date of last follow up.

Statistical analysis

Wherever relevant, the analytical method is detailed in the corresponding figure legends. All statistical analyses were performed using R software and GraphPad Prism v10 (GraphPad Software Inc., San Diego, CA). Between-group comparisons were conducted using an unpaired or paired two-tailed Student’s t-test for two groups, and one-way analysis of variance (ANOVA) for comparisons involving three or more groups. Statistical significance was defined as P < 0.05 (*P < 0.05, **P < 0.01, ***P < 0.001; ns, not significant).

Data availability

The data supporting the findings of this study are available within the Article and its Supplementary Information. The single-cell RNA sequencing, RNA sequencing, and whole exome sequencing data generated in this study will be deposited in the Genome Sequence Archive (GSA, HRA019374). Data and analytical methods are available to other researchers upon request to the corresponding author.

Results

Successful establishment of patient-derived 3DP-PLC models

In this study, we obtained 71 surgical tissue samples from patients who underwent resection and were histologically diagnosed with PLC. Among these, 61 3DP-PLC models were successfully established in vitro, corresponding to an overall establishment rate of 85.9% (Fig. 1A, Table S1). We further compared neoadjuvant treatment history between successfully established and failed cultures. Neoadjuvant targeted therapy was numerically more frequent among failed tissues than among successfully established 3DP-PLC models [4 of 10 (40.0%) vs. 17 of 61 (27.9%)], but this difference was not statistically significant (Fisher’s exact test, P = 0.468). The establishment rate was found to be associated with multiple factors, particularly sample processing time, cell viability, and the method of tissue digestion. The successfully established biobank comprised 51 HCC and 10 iCCA cases (Fig. 1B). Within this subset, 43 of 61 patients (70.5%) were hepatitis B virus-positive, 13 of 61 (21.3%) had elevated serum alpha-fetoprotein (AFP) levels (> 400 ng/mL), 17 received neoadjuvant therapy before surgery, and 18 received adjuvant therapy after resection (Fig. 1C).

Fig. 1.

Establishment of patient-derived 3DP-PLC models. A Workflow of patient enrollment, 3DP-PLC model establishment, and downstream analyses. A total of 71 patients with primary liver cancer were enrolled, from whom 61 successful 3DP-PLC models were established and 10 samples failed to generate successful models. Drug sensitivity screening was performed in 61 3DP-PLC models. Paired sequencing analyses were performed using primary tumor samples and matched 3DP-PLC models, including whole-exome sequencing in 19 pairs, RNA sequencing in 12 pairs, and single-cell RNA sequencing in 2 pairs. B Distribution of Samples Derived from 61 patients, Stratified by Histological Subtype (HCC, n = 51; iCCA, n = 10). C Pie charts depicting the clinical characteristics of the patient cohort (n = 61) with successfully established 3DP-PLC models, stratified by AFP level, viral infection status, and treatment history. D Patient-derived PLC cells maintain high viability in 3D-bioprinted GelMA/AlgMA/HAMA constructs. Cells were assessed by Live/Dead staining over 15 days. The accompanying histogram shows quantitative analysis of cell viability calculated from three independently acquired microscopic fields per time point and presented as mean ± SD. Scale bars, 200 μm (overview images) and 40 μm (magnified views). E Viability of patient-derived 3DP-PLC models over 15 days in culture, assessed by the CellTiter-Glo^® 3D assay. Data represent mean ± SD. F Bright-field microscopy revealed the self-assembly of cells into organoid-like structures within the 3DP-PLC models at day 10 post-printing. Scale bar: 50 μm (G) H&E staining images of 3DP-PLC models and corresponding primary tumor tissues. Scale bar: 50 μm. H Immunofluorescence staining of AFP, CPS1, EpCAM, CK7, CK19, Ki-67, β-catenin in 3DP-PLC models and corresponding primary tumors. Scale bars: 100 μm (tumor tissue sections) and 40 μm (3DP-PLC models)

To generate the 3DP-PLC models, we optimized the bioink composition on the basis of prior established protocols. As the structural backbone of the printed constructs, hydrogels are critical for recapitulating extracellular matrix-like features of the tumor microenvironment and thereby maintaining cellular viability and growth. We therefore screened multiple bioink formulations to identify a system with favorable printability and biocompatibility. Cell viability, assessed by calcein-AM/propidium iodide staining, served as the principal selection criterion and varied substantially across formulations (Fig. 1D, Fig. S1A). Although the GelMA-HAMA formulation supported rapid spheroid formation, viability declined after approximately 10 days, limiting its utility for longer-term culture and drug-sensitivity testing. By contrast, incorporation of AlgMA improved long-term survival and better preserved long-term culture performance. Across models derived from multiple PLC specimens, cell proliferation stabilized after day 3, and viability remained above 60% throughout 15 days of culture (Fig. 1E). Morphologically, 3DP-PLC models retained patient-specific histoarchitectural heterogeneity, including compact solid growth patterns resembling HCC and irregular cystic structures consistent with iCCA (Fig. 1F, Fig. S1B).

3DP-PLC models recapitulated histological, genomic, transcriptomic features and cell types of liver cancer

3DP-PLC models recapitulated the morphological diversity observed across the clinical spectrum of primary liver cancer. During in vitro culture, HCC models predominantly formed compact solid structures, frequently with pseudoglandular rosette-like architecture, whereas iCCA models exhibited prominent glandular organization and preserved key pathological features of the parental tumors, including intraluminal invasion and cribriform growth patterns (Fig. 1G, Fig. S2A). To further assess pathological fidelity, we performed immunofluorescence analyses in 3DP-PLC models and their matched primary tumor tissues. Expression patterns of key lineage and proliferation markers were largely concordant between models and parental tumors, including HCC-associated markers (AFP, CPS1, GPC3, and β-catenin), iCCA-associated markers (CK7, CK19, and EpCAM), and the proliferation marker Ki-67 (Fig. 1H, Fig. S2B and C).

We next examined whether 3DP-PLC models retained the molecular features of their parental tumors. Genomic profiling was performed using paired whole-exome sequencing in 19 tumor-model pairs and paired RNA sequencing in 12 tumor-model pairs, with samples collected on day 10 after bioprinting. At the single-nucleotide level, the models reproduced the mutational landscape of the corresponding tumor tissues with high fidelity. Analysis of single-nucleotide variants (SNVs) and insertion–deletion events showed broadly concordant mutational distributions between models and matched tumors, including preservation of major mutational signatures (Fig. 2A). Driver gene analysis based on a curated panel of more than 60 significantly mutated genes [14, 15, 50, 51] in primary liver cancer further demonstrated substantial genomic concordance between patient-matched tumors and 3DP-PLC models (Fig. 2B, Fig. S2D). Across the 19 sequenced pairs, TP53 was the most frequently altered gene (47%), followed by APOB (32%) and CTNNB1 (26%). Notably, complete (100%) mutational concordance was observed for 28 genes, including canonical drivers such as TP53, PTEN, ERBB2, AXIN1, STAT3, and PREX2. High concordance was also evident for genes related to metabolic regulation (ALB, APOB, KEAP1, and NFE2L2) [15], chromatin remodeling (ARID1A and ARID2) [15], and signaling related to treatment resistance (JAK1 and STAT3) [52]. Beyond point mutations, comparative copy-number variation (CNV) analysis revealed highly concordant genomic profiles, characterized by similar patterns of chromosomal gains and losses between models and primary tumors (Fig. 2C). At the transcriptome level, analysis of 12 matched pairs demonstrated a strong global gene expression correlation, further affirming the models’ molecular fidelity (Fig. S3).

Fig. 2.

Histopathological, genomic, transcriptomic, and cellular recapitulation of parental tumors in 3DP-PLC models. A Comparison of major mutational signatures between primary tumor tissues and matched 3DP-PLC models based on paired whole-exome sequencing (WES; n = 19 tumor-model pairs). B Oncoplot showing frequently mutated driver genes in primary PLC tissues and matched 3DP-PLC models based on paired WES analysis (n = 19 tumor-model pairs). Rows indicate driver genes, and columns indicate paired tumor-model samples. C Concordance of genome-wide copy-number variation profiles between primary tumor tissues and matched 3DP-PLC models inferred from paired WES data (n = 19 tumor-model pairs). D UMAP plots split by sample condition, showing the distribution of major cell populations in primary tumor tissues and matched Day 10 3DP-PLC models from two matched tumor-model pairs (n = 2 pairs; 4 samples). E Stacked bar plots showing the relative proportions of major cell types in the indicated primary tumor tissues and matched Day 10 3DP-PLC models based on scRNA-seq analysis (n = 2 matched tumor-model pairs; 4 samples). F Flow cytometry comparison of freshly dissociated primary PLC tissues and matched Day 10 3DP-PLC models (n = 2 matched tumor-model pairs). The analysis was used to assess retention of epithelial, stromal, and immune cell populations after 10 days of 3D culture

To evaluate whether 3DP-PLC models preserve the cellular heterogeneity of the parental tumors, we performed single-cell RNA sequencing (scRNA-seq) on two matched tumor-model pairs. Integrated analysis identified the major tumor-associated populations, including HCC-associated cells, fibroblasts, macrophages /monocytes, T-cell subsets, NK cells, dendritic cells, B cells, and plasma cells (Fig. S4A and S4B). Cell identities were supported by canonical markers, including ALB, APOB, TF, TTR, GPC3 and AFP for HCC-associated cells, COL1A1, FAP, ACTA2 and DCN for fibroblasts, CD68, CD163, C1QA and CSF1R for myeloid cells, CD3D, CD3E, TRAC, CD8A, CD79A and MS4A1 for lymphoid populations. UMAP and cell-type composition revealed a striking overlap between cells derived from the primary tissues and the 3DP-PLC models (Fig. 2D and 2E, Fig. S4D). The 3D-bioprinted constructs robustly maintained the diverse subpopulations of malignant hepatocytes present in the parental tumors, indicating that the bioprinting process and the tailored bioink microenvironment effectively sustain patient-specific tumor architectures ex vivo. Because matched adjacent non-tumor tissues were not profiled by scRNA-seq, normal hepatocytes could not be directly identified from paired non-tumor tissue. We therefore used canonical hepatocyte-lineage markers to describe hepatocyte-lineage programs and inferCNV to support malignant-cell assignment. Using T and B cells as inferCNV reference cells, 9,620 marker-defined cancer cells entered CNV scoring, and 6,372 cells were retained as CNV-defined malignant HCC cells for downstream analysis (Fig. S4C). Program-based subclustering of these CNV-defined malignant cells resolved hepatocyte-like, progenitor/ductal-like, cycling, inflammatory/acute-phase, hypoxia/stress, EMT/ECM-invasive and UPR/immediate-early HCC states (Fig. S4E). Matched tumors and 3DP-PLC models showed substantial overlap in the malignant-cell transcriptional landscape (Fig. S4E). Importantly, patient-specific state composition was maintained, with PLC24 enriched for hypoxia-like and UPR-high states and PLC45 dominated by cycling states in both parental tumors and matched models (Fig. S4F).

Finally, we assessed the preservation of primary tumor features at the cellular population level by comparing flow cytometry of freshly dissociated tumor tissues collected with the corresponding 3DP-PLC models after 10 days of culture. We examined the relative abundance of CD326-positive cell populations, FAP-associated stromal compartments, and immune cell populations. The Day 10 3DP-PLC models retained portions of the epithelial-/stem-like, stromal, and immune cellular architecture of the corresponding parental tumors, despite measurable shifts in the proportions of individual cell subsets (Fig. 2F).

We further evaluated whether these components could be retained during short-term culture by immunofluorescence staining. CPS1/CD45 staining was performed on day 1 and day 10 in 3DP-PLC constructs generated with GelMA/HAMA and GelMA/HAMA/AlgMA bioinks. CD45-positive immune cells were detectable on day 1 and remained detectable at day 10 in both bioink conditions, although their abundance was reduced compared with day 1. (Fig. S5A) To further validate stromal and immune-cell retention, 3DP-PLC models were stained for CD3/CD45, CD11b/CD45, and CPS1/α-SMA at day 1 and day 10. These analyses confirmed the presence of residual T-cell-associated, myeloid-associated, and fibroblast-associated components in day 10 3DP-PLC models. Together, these findings support partial short-term retention of tumor-associated stromal and immune components. (Fig. S5B)

Personalized screening of therapeutic drugs via 3DP-PLC models

In this study, we screened nine therapeutically relevant agents in 61 3DP-PLC models. These included first-line targeted therapies (lenvatinib and sorafenib), second-line targeted therapies (regorafenib and apatinib), and commonly used chemotherapeutic agents (oxaliplatin, cisplatin, epirubicin, gemcitabine, and 5-fluorouracil). In the representative model derived from PLC 14, differential sensitivity to the four targeted agents was readily apparent in bright-field imaging after treatment. Increasing drug concentrations were associated with progressive morphological changes, including central necrosis, cellular disintegration, and spheroid contraction, indicating distinct drug-specific effects on tumor viability and structural integrity (Fig. 3A). Consistent with these observations, calcein-AM/propidium iodide staining performed after 72 h of exposure showed heterogeneous viability responses across the concentration gradient (Fig. S6). To quantify these differences systematically, dose–response profiling was performed for all nine agents in the 61 3DP-PLC models, and the normalized area under the dose–response curve (AUC) was used as a standardized measure of drug sensitivity (Fig. 3B and C, Fig. S7A, S7B and S7C, Table S3). Marked interpatient heterogeneity was observed across both targeted and chemotherapeutic agents. In the iCCA models, drug-response profiles were heterogeneous across the four tested agents (Fig. 3D). Models derived from PLC48 showed broad resistance to multiple agents, particularly oxaliplatin, cisplatin, and lenvatinib, whereas PLC59 displayed relative sensitivity to several chemotherapeutic agents. In HCC models, targeted therapy response profiling revealed substantial interpatient variability across sorafenib, regorafenib, apatinib, and lenvatinib (Fig. 3E). Models derived from PLC03 and PLC18 showed broad resistance to multiple targeted agents, whereas models from PLC08 and PLC16 displayed relative sensitivity to several targeted therapies. Other models showed drug-selective response patterns, supporting individualized rather than uniform sensitivity to tyrosine kinase inhibitors. Chemotherapeutic response profiling also revealed heterogeneous sensitivity among HCC models (Fig. S7D). These findings indicate that 3DP-PLC models capture substantial interpatient variation in ex vivo drug-response profiles across both HCC and iCCA models.

Fig. 3.

Assessment of 3DP-PLC Models in Response to Anticancer Drugs. A Dose-dependent morphological changes in 3DP-PLC models. Representative bright-field images taken 7 days post-printing after 72 h of drug treatment. B Dose-response curves of seven drugs (lenvatinib, apatinib, regorafenib, sorafenib, cisplatin, epimycin, and 5-FU) for 3DP-PLC models (n = 61 models). Each colored line represents the dose-response curve of an individual patient-derived 3DP-PLC model, and cell viability (mean ± SD) was normalized to the corresponding vehicle-treated control. C Distribution of normalized AUC values across 3DP-PLC models (n = 61 models). Violin plots display the data spread; dashed and dotted lines represent the mean and interquartile range (upper and lower quartiles). D AUC heatmap of drug sensitivity in 10 iCCA 3DP-PLC models. E AUC heatmap of Targeted drug sensitivity in 51 HCC3DP-PLC models

Because immune checkpoint inhibitors are increasingly used in HCC and iCCA, we performed a pilot experiment to examine whether PD-1 blockade affected 3DP-PLC viability during the short-term screening window. One iCCA model and one HCC model were treated with anti-PD-1 antibody at 100 µg/mL and compared with untreated controls. Live/dead staining showed that PD-1 treatment did not cause extensive construct disruption, although PI-positive cells were modestly increased after treatment. (Fig. S7E) Flow-cytometry analysis showed a modest reduction in tumor-cell populations and a slight increase in CD107a-positive CD8 T cells after PD-1 exposure. (Fig. S7F) However, the magnitude of these changes was limited.

Because preoperative systemic therapy may affect tumor viability and ex vivo drug-response profiles, we performed an additional HCC-only sensitivity analysis focusing on neoadjuvant lenvatinib exposure. After excluding iCCA cases from the original drug-screened cohort, 51 HCC models were included in this analysis, including 12 models from patients who had received neoadjuvant lenvatinib and 39 models from patients who had not. Normalized lenvatinib AUC values were comparable between the two groups [median, 0.724; IQR, 0.626–0.874 versus median, 0.758; IQR, 0.701–0.815; Mann–Whitney U test, P = 0.747] (Table S4). Moreover, neoadjuvant lenvatinib-treated samples were distributed across Sensitive, Relatively sensitive, and Resistant classes rather than being uniformly enriched in the Resistant class.

Development of a therapeutic response stratification system based on drug testing in 3DP-PLC models

When drug-sensitivity testing was performed in 61 3DP-PLC models, normalized AUC values showed marked interpatient heterogeneity across both targeted and chemotherapeutic agents. Because the external clinical benchmark used for response calibration was derived from published HCC targeted therapy trials, iCCA drug-response data were visualized separately and were not included in the HCC-focused stratification analysis. In the HCC cohort, normalized AUC values for the four targeted therapies showed a largely continuous distribution with occasional breakpoints. On this basis, we established a HCC-focused stratification system that integrates breakpoint identification by the Jenks natural breaks method with calibration against clinically reported ORR, SD, and PD in HCC trials, thereby converting continuous ex vivo AUC values into clinically interpretable response classes (Fig. 4A). Importantly, cutoff derivation was performed without using patient-level clinical outcomes from the present cohort, which were reserved for subsequent exploratory evaluation of clinical relevance. Within this system, the Jenks natural breaks method was applied to the normalized AUC distribution to identify all potential breakpoints. Adjacent subclasses defined by these breakpoints were then systematically merged to generate three final categories. This procedure produced the complete set of initial grouping solutions (Fig. 4B, Fig. S8A and Table S5). These solutions were subsequently filtered according to pre-specified criteria, including a goodness of variance fit (GVF) of at least 90% and an incremental GVF gain of at least 1% (Fig. 4C, Fig. S8B). The retained solutions were then advanced as candidate partitions for downstream analyses. For each retained candidate partition, the resulting class-prevalence vector was projected into a three-dimensional response space and compared with the recalculated external benchmark composition (ORR/SD/PD), which was derived from four published clinical trials of targeted therapies after excluding unevaluable cases (Table S6). To enable this comparison, the three ex vivo response classes were mapped to the corresponding clinical response categories according to AUC magnitude: the subgroup with the lowest AUC values was assigned to ORR, the intermediate subgroup to SD, and the subgroup with the highest AUC values to PD (Fig. 4D, Fig. S8C). The discrepancy between each candidate partition and the external benchmark was then quantified using Euclidean distance (Table S7). For each of the four drugs, the cutoff pair yielding the minimal distance was selected as the final cutoff pair and the corresponding grouping scheme (Fig. 4E).

Fig. 4.

Development of an HCC-focused therapeutic response stratification system based on drug testing in 3DP-PLC models. Because the external clinical response benchmarks were derived from HCC studies, ECJMS calibration was performed using HCC-derived 3DP-PLC models after excluding the three iCCA-derived models (n = 51). A Schematic overview of the algorithm. B Initial grouping of 3DP-PLC models according to lenvatinib response. Breakpoints were identified using the Jenks natural breaks method (K ≥ 3), followed by contiguous merging of adjacent subclasses into three final groups. C Goodness-of-fit curves based on 3DP-PLC models and lenvatinib response were used to define the candidate range of interval numbers for Jenks classification; the region between the two dashed lines was retained for subsequent analyses. D Three-dimensional matching of candidate lenvatinib stratification schemes to the external benchmark composition. The optimal cutoff pair and corresponding grouping scheme were defined by the minimum Euclidean distance. E Optimal grouping schemes and corresponding cutoff pairs for the four targeted therapies. N indicates the number of classes. F Joint stability of the bootstrap-derived lower and upper cutoffs for lenvatinib. Grey dots indicate cutoff pairs obtained from bootstrap full-pipeline resampling, and dashed contour lines show the density distribution of locally optimal solutions. The blue vertical dashed line and red horizontal dashed line represent the locked lower and upper cutoffs, respectively. The intersection of the two lines marks the final locked cutoff pair for lenvatinib. G Sample-level assignment stability across bootstrap resamples of 3DP-PLC models according to lenvatinib response. For each sample, the probability of assignment to the three classes was calculated, and the maximum assignment probability was used to assess classification stability. Samples with lower assignment confidence represent boundary or grey-zone cases. H Dirichlet-based uncertainty propagation of the external benchmark for lenvatinib. Scaled density curves show the posterior distributions of the recalculated external benchmark composition (ORR/SD/PD) after exclusion of unevaluable cases. Vertical dashed lines denote the observed ex vivo class proportions generated by the locked cutoff pair. The close correspondence between the observed proportions and the posterior-supported benchmark range indicates robustness of the stratification system to benchmark uncertainty. I Violin plots depicting the optimal stratification of four targeted drugs. The two dashed lines indicate the established grouping boundaries for therapeutic response classification

To further assess the internal stability of this system, we performed bootstrap full-pipeline resampling. The resulting cutoff pairs did not form a diffuse distribution; instead, they clustered within a limited neighborhood of local optima, indicating structured local uncertainty rather than overall instability (Table. S8). Notably, both the externally calibrated lower and upper cutoffs fell within these locally optimal regions (Fig. S9A and S9B), and their joint stability was preserved (Fig. 4F, Fig. S9C), supporting the overall robustness of the system. At the sample level, class assignment remained highly reproducible across bootstrap resamples, with uncertainty largely confined to a small subset of boundary cases (Fig. 4G, Fig. S10A). Given potential differences between the external benchmark population and our study cohort, we further evaluated the robustness of the calibration procedure by modelling the external benchmark as a Dirichlet posterior rather than a fixed point estimate. Under this uncertainty-aware calibration system, the optimal cutoff pair remained unchanged (Table. S9). Moreover, the observed ex vivo class composition consistently remained within the posterior-supported range of the external benchmark distribution (Fig. 4H, Fig. S10B). Together, these findings support the robustness of the prevalence-matching system to benchmark uncertainty.

Following application of the locked cutoffs, continuous AUC values were stratified into three ordered response classes, namely Sensitive, Relatively sensitive, and Resistant, which exhibited clear monotonic separation across all four drugs (Fig. S10C). Violin plots confirmed that these strata preserved the expected gradient of drug sensitivity, with the Sensitive class showing the lowest AUC values and the Resistant class the highest, while remaining broadly concordant with the recalculated external benchmark composition (Fig. 4I).

Association of 3DP-PLC model drug sensitivity with clinical response and postoperative outcome

We next examined whether drug-sensitivity classes derived from 3DP-PLC models were concordant with clinical treatment responses. Among the 12 patients who received neoadjuvant lenvatinib, normalized AUC values from patient-matched 3DP-PLC models differed significantly across clinical response categories (Fig. 5A). Application of the optimal cutoffs defined by the stratification system partitioned all patients into three response classes, whose corresponding normalized AUC values remained clearly separated (Fig. S11A). These findings indicate that the classification system captured clinically meaningful heterogeneity in therapeutic response. When neoadjuvant treatment responses were mapped onto the classification, complete concordance was observed: all patients with PR were assigned to the Sensitive class, those with SD to the Relatively sensitive class, and those with PD to the Resistant class (Fig. 5B). Notably, patients who achieved PR had received lenvatinib monotherapy, whereas some patients with SD or PD had received combined regimens including TACE and/or PD-1 blockade. However, because of the limited sample size and heterogeneity of combined neoadjuvant regimens, these results should not be interpreted as definitive evidence of lenvatinib monotherapy prediction and require validation in larger prospective cohorts. A similar pattern was observed in the neoadjuvant apatinib setting, in which the patient with PR was classified as Sensitive (Fig. 5C).

Fig. 5.

Concordance between drug sensitivity profiles of 3DP-PLC models and corresponding patient clinical outcomes. A Comparison of normalized AUC values among 3DP-PLC models stratified by clinical response to neoadjuvant lenvatinib therapy (n = 12 patients; PR, n = 4; SD, n = 4; PD, n = 4). Statistical significance was determined by one-way ANOVA (PR vs. SD, P = 0.0056; PR vs. PD, P < 0.0001; SD vs. PD, P = 0.0013). **P < 0.01, ****P < 0.0001. B Correlation between normalized AUC values derived from drug sensitivity testing and disease control status in PLC patients receiving neoadjuvant lenvatinib therapy (n = 12 patients). Individual AUC values are plotted as circles in descending order; the two dashed lines denote the optimal cutoff values for response stratification. C Correlation between normalized AUC values derived from drug sensitivity testing and disease control status in PLC patients receiving neoadjuvant apatinib therapy. (D) This table summarizes the drug testing results of representative 3DP-PLC models and the corresponding clinical responses of patients during follow-up. E Treatment process and clinical outcome of patient PLC01. F Magnetic resonance imaging (MRI) of target lesions in patient PLC01 before and after therapy. G Swimmer plots depicting DFS differences among 15 patients stratified into three groups based on drug sensitivity testing results. H Kaplan-Meier curves showing OS differences in TCGA for patients enriched in differentially expressed genes of 3DP-PLC models. I Kaplan-Meier curves showing OS differences in ICGC for patients enriched in differentially expressed genes of 3DP-PLC models. J KEGG pathway enrichment analysis of differentially expressed genes between lenvatinib-sensitive and lenvatinib-resistant 3DP-PLC models

The association between clinical response and ex vivo drug responsiveness in patient-matched 3DP-PLC models among patients who received both neoadjuvant and adjuvant lenvatinib is shown in Fig. 5D. This concordance was further supported at the individual-patient level. In patient HCC 01, a liver mass had been detected three years earlier and was radiologically suspected to represent HCC, as shown in Fig. 5E. The patient initially received prior systemic therapy; however, the assessment after two months of therapy indicated PD. Following discontinuation of the trial regimen, first-line lenvatinib was initiated. This treatment induced modest tumor shrinkage, corresponding to a best response of SD, and rendered the patient eligible for surgical resection. Subsequent in vitro drug-sensitivity testing of the patient-derived 3DP-PLC tumor model indicated relative sensitivity to lenvatinib (Fig. 5E). Longitudinal MRI monitoring documented dynamic changes in tumor burden throughout the treatment course (Fig. 5F). Guided by both the 3DP drug-sensitivity result and the observed neoadjuvant clinical response, lenvatinib was resumed one month after surgery, and the disease has since remained stable for months before recurrence was detected. Additional representative cases, including PLC 17 and PLC 42, demonstrated similar concordance between ex vivo sensitivity class, treatment response, and postoperative clinical course, further supporting the translational relevance of the stratification system (Fig. S10B, Fig. S11B).

We further examined the association between 3DP-PLC drug-sensitivity stratification and postoperative clinical outcome in patients who subsequently received lenvatinib. Follow-up analyses were performed in 15 patients who underwent adjuvant lenvatinib treatment (Fig. 5G). On the basis of in vitro drug-sensitivity profiles generated from patient-matched 3DP-PLC models, these patients were stratified into three categories. Of the six patients classified as Resistant, five reached the DFS endpoint within 6 months. Among the seven patients classified as Relatively sensitive, only two reached the DFS endpoint, both after more than 12 months. The remaining two patients were classified as Sensitive, and neither experienced disease progression during follow-up. No clear outcome pattern was evident in the Sensitive group, largely owing to the limited sample size and shorter follow-up duration. Nevertheless, the overall trend was directionally consistent with the ex vivo drug sensitivity classification, with poorer postoperative outcomes concentrated in the Resistant group and comparatively more favorable outcomes observed in the Relatively sensitive group. Taken together, the concordance between neoadjuvant treatment response, postoperative clinical outcome, and drug-sensitivity classification derived from 3DP-PLC models provides both biological and clinical support for the translational relevance of this system.

Comparative transcriptomic analysis of resistant and sensitive phenotypes in 3DP-PLC models

Using the Jenks natural breaks method, 3DP-PLC models were stratified into lenvatinib-sensitive, relatively sensitive, and resistant categories. Comparative transcriptomic analysis of the sensitive and resistant groups identified differentially expressed genes associated with treatment response (Fig. S12A). The transcriptional profile of lenvatinib-resistant 3DP-PLC models was characterized by upregulation of canonical resistance-associated drivers, including EGFR and MET, together with additional genes implicated in resistance, such as NEK7, APC, ABCB1, ATF2, FZD4, PIK3CA, and MAST4. In contrast, genes functionally linked to apoptosis and ubiquitination, including ZBP1, BAK1, and DUSP4, were downregulated (Fig. S12B).

To assess the clinical relevance of this transcriptional profile, we evaluated the prognostic associations of the differentially expressed genes in the TCGA cohort using univariable Cox regression analysis, with the results summarized in Table S10. Using a significance threshold of P < 0.01, we identified candidate prognostic genes associated with overall survival (OS). These genes were subsequently subjected to LASSO regression for dimensionality reduction (Fig. S12C, Fig. S12D), followed by multivariable Cox regression, which yielded a five-gene prognostic signature comprising GNL2, SLC1A5, EIF5B, PLOD2, and CHAC2 (Fig. S12E). Based on the regression coefficients derived from the multivariable model, a risk score was calculated for each patient, and the TCGA cohort was stratified into high-risk and low-risk groups. The risk distribution plot showed that death events increased with rising risk score (Fig. S12F). Kaplan–Meier analysis demonstrated significantly poorer survival in the high-risk group than in the low-risk group (P < 0.001; Fig. 5H). The prognostic performance of the five-gene signature was further validated in the ICGC cohort, in which patients classified as high risk similarly showed worse OS (P = 0.0023; Fig. 5I).

KEGG analysis further supported the biological plausibility of the resistant phenotype. This enrichment analysis showed significant activation of multiple oncogenic pathways in the lenvatinib-resistant group, including the MAPK, TGF-β, mTOR, and ErbB signaling pathways, as well as the EGFR tyrosine kinase inhibitor resistance pathway (Fig. 5J, Table S11). Consistently, GO enrichment analysis indicated downregulation of apoptosis-related processes, suppression of oxidative phosphorylation, and reduced ROS production in the lenvatinib-resistant group (Fig. S12G and S12H, Table. S12). Collectively, these findings suggest that 3DP-PLC models preserve resistance-associated and prognostically relevant molecular features observed in patient tumors.

3D co-cultures of matched 3DP-PLC models and CAFs

To investigate stromal contributions to primary liver cancer biology, we established a 3D bioprinted co-culture platform incorporating tumor cells and CAFs (Fig. 6A). Following tissue dissociation, CAFs were enriched by plastic adherence and expanded in culture, whereas the remaining cell populations were cryopreserved. The isolated CAFs displayed the expected elongated spindle-shaped morphology, and immunofluorescence analysis confirmed strong expression of α-SMA and vimentin, consistent with retention of canonical CAF features under in vitro conditions (Fig. S13A, Fig. S13B) [53].

Fig. 6.

Establishment of patient-derived 3DP-PLC co-culture models incorporating matched CAFs.A Schematic illustration of the fabrication strategy and characterization of patient-derived 3DP-PLC co-culture models. B Comparative morphological analysis of Hep3B cells in 3DP monoculture and co-culture models. Representative bright-field micrographs captured at days 1, 2, 4, 6, and 8 demonstrate the temporal evolution of cellular architecture. Scale bar: 50 μm. C Quantification of spheroid formation in 3DP Hep3B models at day 1 and day 8 post-bioprinting. D Diameter distribution of cellular spheroids within 3DP Hep3B models at day 1 and day 8 post-bioprinting. E Longitudinal assessment of cell viability and self-assembly in 3DP Hep3B monoculture and co-culture models via Live/Dead staining. Representative fluorescence images (live cells, green; dead cells, red) were captured at days 1, 2, 4, 6, and 8 post-bioprinting. Scale bars, 200 μm (overview images); 40 μm (magnified views). F Longitudinal proliferation analysis of Hep3B cells in 3DP monoculture and co-culture models over 8 days. G Longitudinal proliferation analysis of Huh-7 cells in 3DP monoculture and co-culture models over 8 days. Scale bar: 50 μm. H Representative bright-field images of 3DP-PLC monoculture and co-culture models. Red arrows indicate fibroblast-rich ring structures surrounding the tumor cell clusters in co-culture conditions. Scale bar: 40 μm. I Representative immunofluorescence images of 3DP-PLC monoculture and co-culture models. Red arrows indicate fibrous ring. J Drug testing results of four TKIs (n = 34). Statistical significance was assessed using the Wilcoxon matched-pairs signed-rank test (*P < 0.05, **P < 0.01, ***P < 0.001). K Heatmap of differential drug sensitivity in the presence of CAFs (AUCco/AUCmono). Data are sorted according to the CAF drug influence, representing the average relative change of all four treatments (n = 34)

We next examined the functional effects of CAFs using co-culture models containing Huh-7 or Hep3B cells. Bright-field imaging documented the growth dynamics of monoculture and co-culture constructs on days 1, 2, 4, 6 and 8. By day 8, evident spheroid formation was observed in both conditions (Fig. 6B, Fig. S13C). Quantitative analysis further showed that spheroid number and diameter were significantly greater in the co-culture models than in the monoculture controls at day 8 (Fig. 6C and D, Fig. S13D and S13E), indicating that CAFs promote tumor cell growth and spheroid formation. Consistently, live/dead staining performed on days 1, 2, 4, 6 and 8 revealed progressive cellular aggregation and sustained viability in both monoculture and co-culture constructs (Fig. 6E, Fig. S13F). Proliferation assays further demonstrated that CAFs enhanced the proliferative capacity of Huh-7 and Hep3B cells (Fig. 6F and G). Together, these data indicate that CAFs promote tumor cell expansion and spheroid development in 3D bioprinted liver cancer models.

We then generated 34 pairs of monoculture and CAF co-culture 3DP-PLC models derived from PLC spanning multiple anatomical sites and disease stages. Bright-field imaging showed close spatial proximity between tumor cells and CAFs in the co-culture constructs (Fig. 6H, Fig. S13G). To further validate direct tumor-CAF contact, tumor cells and CAFs were labeled with DiO and DiI, respectively, before co-bioprinting. Dual-fluorescence imaging demonstrated direct contact and spatial organization between tumor cells and CAFs within the 3D constructs (Fig. S13H). In a subset of cases, tumor clusters were surrounded by a distinct fibrous ring-like structure, a feature that was further supported by immunofluorescence analysis (Fig. 6I, Fig. S13I). Drug-sensitivity testing was subsequently performed in 34 paired monoculture and CAF co-culture 3DP-PLC models using four tyrosine kinase inhibitors, namely lenvatinib, apatinib, regorafenib and sorafenib (Fig. S14, Table S13). Compared with matched monocultures, CAF co-culture increased AUC in 97 of 133 evaluable drug-model comparisons (72.9%), supporting an overall stromal-associated reduction in ex vivo drug sensitivity. However, the magnitude and direction of this effect varied across both patients and drugs. The proportion of models showing increased AUC ranged from 82.4% for lenvatinib to 63.6% for sorafenib, indicating that CAF-mediated drug modulation was not uniform. This heterogeneity was reflected in the normalized AUC profiles and enabled derivation of a CAF-dependent drug-response score for each case (Fig. 6J, K).

Single-cell dissection reveals CAF-driven remodeling toward a resistant tumor state

The tumor microenvironment, particularly CAFs, plays a pivotal role in liver cancer progression and therapeutic recalcitrance. To dissect this stromal-epithelial crosstalk, we performed scRNA-seq on six samples derived from two independent HCC cases, including CAF monoculture, HCC monoculture and HCC-CAF co-culture conditions for patient PLC31 and PLC42. Integrated analysis resolved a complex cellular atlas composed of multiple malignant HCC states, four CAF states, and several immune populations (Fig. 7A, Fig. S15A-C). The malignant compartment segregated into epithelial, mesenchymal, stressed, cycling, hepatocyte-like and metabolic HCC states, whereas the stromal compartment comprised myCAFs, vCAFs, dCAFs and pericyte-like CAFs. In addition, CD8⁺ T cells, Treg cells, macrophages, plasma cells and B-cell populations were retained in the integrated dataset (Fig. 7B, Fig. S15D).

Fig. 7.

Single-cell dissection reveals CAF-driven remodeling toward a resistant tumor state (n = 2 patients, 6 samples). A UMAP visualization of integrated single-cell transcriptomes from matched HCC–CAF models, colored by annotated cell type. B Stacked bar plots showing the cellular composition of each sample from the two matched HCC patients. C UMAP plots colored by sample type and by patient identity. D Feature plots showing the expression of representative marker genes used for lineage annotation. E Heatmap summarizing Hallmark-based module scores associated with tumor aggressiveness across the indicated sample groups. Colors indicate row-wise z-scores. F Violin plots comparing stemness score and drug resistance score among CAF monoculture, co-culture, and HCC monoculture conditions. G GSEA analysis of cancer cells comparing co-culture versus HCC monoculture. H Circle plot of cell-cell communication network. The width of edges represents the strength of communication. I Heatmap summarizing secretory and matrix-related functional programs across CAF subsets identified under different culture conditions

Overlay of sample type on the UMAP showed that the co-culture condition occupied a transcriptional space intermediate between HCC monoculture and CAF monoculture, while also extending into distinct regions of the integrated manifold not prominent in either single-component condition alone, which revealed a distinct phenotypic shift in tumor cells co-cultured with CAFs compared to their mono-cultured counterparts. (Fig. 7C and D, Fig. S15E and F). To determine whether this variability was associated with CAF-state heterogeneity, we performed CAF reclustering of the CAF compartments from PLC31 and PLC42 CAF monoculture and co-culture samples (Fig. S16A). This analysis resolved myCAF, iCAF, dCAF and vCAF CAF states. CAF monocultures exhibited patient-specific state composition, with PLC31 enriched for myCAF and dCAF states and PLC42 containing prominent dCAF, vCAF and iCAF populations (Fig. S16B & S16C). Recovered co-culture CAFs were predominantly iCAF-like, with a minor vCAF component in PLC31. Because the number of recovered co-culture CAFs was limited, these results were interpreted descriptively rather than as definitive evidence of CAF-state remodeling.

We next examined malignant cells only to avoid confounding malignant-cell programs with CAF transcriptional profiles (Fig. S16D). CAF co-culture increased malignant-cell TKI-resistance, stemness and EMT/invasive scores in a patient-specific manner. PLC31 showed a prominent increase in the TKI-resistance program, whereas PLC42 showed larger increases in stemness and EMT/invasive programs. These findings indicate that CAF co-culture generally reduces ex vivo drug sensitivity, but the direction and magnitude of this effect are shaped by patient-specific CAF-state composition and CAF-to-malignant communication programs.

To further define the biological basis of this transition in cancer cells, we performed GSEA comparing co-culture with HCC monoculture. The co-culture condition was enriched for pathways linked to xenobiotic and drug metabolism, including cytochrome P450-related programs, metabolism of xenobiotics, steroid hormone biosynthesis and ABC transporters, together with complement and coagulation cascades, cytokine-cytokine receptor interaction, cell adhesion molecules and tight junction pathways (Fig. 7G). Furthermore, tumor cells in the co-culture models upregulated gene signatures associated with multidrug resistance and metabolic reprogramming.

Because the phenotypic effect of CAF incorporation is likely mediated by active intercellular communication, we next reconstructed ligand–receptor interaction networks across the integrated co-culture atlas. The resulting communication map showed dense bidirectional interactions centered on malignant HCC cells and major CAF populations, with additional contributions from macrophages and lymphoid populations (Fig. 7H). The secretome heatmap further suggested that vCAF, myCAF, pericyte-like CAF and dCAF populations contributed distinct secretory programs involving cytokines, growth factors, ECM components, MMPs and exosome-related signals (Fig. 7I).

Taken together, single-cell profiling of matched CAF co-culture models provides mechanistic support for the phenotypic drug-resistance data obtained in the larger paired biobank. Specifically, CAF co-bioprinting drives a coordinated remodelling of tumor cell state, stromal composition and intercellular signaling that converges on stemness-associated and drug-resistant transcriptional programs.

Integrative multi-omics analysis links tumor-intrinsic pathway activity, stromal modulation and ex vivo TKI response

To determine whether the genomic, transcriptomic, single-cell and pharmacological data converged on coherent biological patterns, we performed an overlap-aware integrative analysis across the available molecular and drug-response datasets (Fig. S17) [54, 55]. Because WES, bulk RNA-seq, scRNA-seq and drug-response measurements were not available in an identical set of patients, we used an overlap-aware design in which each analysis used the maximal valid cohort, without imputing unmeasured modalities or fitting a multivariable predictor (Fig. S17A). Tumor-model fidelity was evaluated in 19 WES pairs, 12 bulk RNA-seq pairs and two scRNA-seq pairs. PLC48, the only iCCA case within the WES/bulk-RNA overlap, was retained only for descriptive fidelity analysis and excluded from HCC-specific molecular-pharmacological tests. Under the strict somatic-variant filter, the median exact-variant Jaccard index was 0.556 in 18 matched HCC WES pairs (interquartile range, 0.508–0.604), compared with a median of 0 among mismatched tumor-model combinations (model-label permutation P < 0.0001; Fig. S17B). The matched model ranked first among all candidate models for every HCC tumor. In 11 HCC bulk-RNA pairs, the median matched tumor-model Spearman correlation across the 5,000 most variable expressed genes was 0.678, compared with 0.485 for mismatched models (model-label permutation P = 0.0005; Fig. S17C). Thus, matched models retained patient-specific molecular identity above mismatched baselines, although this did not imply complete preservation of every parental molecular state.

We next defined a strict complete-case subset of eight HCC models with paired WES, paired bulk RNA-seq and complete monoculture AUC measurements for lenvatinib, apatinib, regorafenib and sorafenib: PLC33, PLC38, PLC41, PLC43, PLC45, PLC47, PLC49 and PLC51. The aligned map visualized retained WES driver-pathway status, 3DP bulk-RNA ssGSEA pathway activity and monoculture TKI AUC at the patient level without using the eight cases as a training cohort (Fig. S17D) [56]. WES pathway status was tested against lenvatinib AUC in 18 HCC models and against the other TKI AUCs in 16 HCC models; no WES pathway-drug association survived multiple-testing correction (minimum BH-FDR = 0.109). In the 11 HCC models with 3DP bulk RNA-seq and monoculture TKI data, higher EMT/TGF-beta activity was associated with higher lenvatinib AUC (Spearman rho = 0.773, permutation P = 0.00735, patient-bootstrap 95% CI = 0.327–0.972), as was hypoxia activity (rho = 0.700, P = 0.0200, 95% CI = 0.271–0.886). Both associations retained a positive direction in all leave-one-out analyses, but neither remained significant after correction across 32 pathway-drug tests (BH-FDR = 0.235 and 0.320, respectively; Fig. S17E).

Stromal pharmacology was integrated using the patient-matched monoculture and CAF co-culture cohort (Fig. S17F). Lenvatinib had 34 evaluable pairs, whereas apatinib, regorafenib and sorafenib each had 33. Among the 33 models evaluable for all four TKIs, the median patient-level mean log2(AUC co-culture/AUC monoculture) was 0.0818 (patient-bootstrap 95% CI = 0.0496–0.1824; two-sided Wilcoxon signed-rank P = 0.000462). Twenty-six of 33 models (78.8%) showed an overall resistance shift and seven (21.2%) showed an overall sensitization shift. Drug-specific median effects were positive for lenvatinib, apatinib, regorafenib and sorafenib, with adjusted FDR values of 0.008, 0.008, 0.013 and 0.047, respectively. Individual model-level effects nevertheless occurred in both directions for every TKI.

The scRNA-seq datasets provided cellular-context anchors for the molecular and pharmacological observations (Fig. S17G). The tumor-model fidelity dataset comprised PLC24 and PLC45, whereas the CAF co-culture dataset comprised PLC31 and PLC42. PLC45 overlapped with the eight-model WES/bulk-RNA/drug subset, providing one four-modality case; no population-level four-modality analysis was performed. In PLC24 and PLC45, matched malignant-cell pseudobulk correlations were 0.978 and 0.977, respectively, compared with cross-patient correlations of 0.849 and 0.857 (Fig. S17G1). PLC31 and PLC42 showed different CAF-state compositions (Fig. S17G2). Following CAF co-culture, PLC31 showed the larger increase in the malignant-cell TKI-resistance score (+ 0.75), PLC42 showed the larger increase in stemness (+ 0.38), and both systems showed increased EMT/invasion scores (+ 0.54 and + 0.46; Fig. S17G3). Focused CAF-to-malignant communication patterns also differed between the two systems (Fig. S17G4). Because each scRNA-seq analysis comprised only two patient systems, these observations were treated as cellular-context anchors rather than population-level validation.

Discussion

Interpatient variability in therapeutic response remains a key obstacle to precision treatment in PLC [44, 57]. This variability reflects both tumor-intrinsic heterogeneity and the dynamic influence of the TME [24]. Although clinical decision-making increasingly incorporates molecular and clinicopathologic information, these approaches often provide limited guidance for selecting systemic therapy at the individual level [58]. Patient-derived ex vivo platforms, including PDX models [30] and organoids [27, 44], have advanced liver cancer modeling but remain constrained by cost, culture complexity, reproducibility, standardization, and the lack of clinically interpretable response frameworks [59]. Building on our previous proof-of-concept work [40], the present study establishes 3DP-PLC as a tunable patient-derived platform that preserves tumor features while extending beyond model establishment toward clinically anchored therapeutic stratification and tumor-stroma heterogeneity analysis. In this context, 3D bioprinting provides a complementary short-term modeling strategy that enables standardized construct geometry, controlled cell density, tunable bioink composition, and reproducible tumor-stroma assembly for rapid ex vivo drug-response profiling.

A major strength of this platform is the biological adaptability of the optimized bioink. PLC comprises histologically distinct categories, most prominently HCC and iCCA, which differ in lineage features, architecture, biomarker expression, and stromal elements. The ability of a single optimized bioink system to support ex vivo growth of both HCC and ICCA models demonstrates its broad histologic compatibility. 3DP-PLC models retained lineage-consistent morphology, biomarker expression, driver gene alterations, and transcriptomic features of their parental tumors. The paired WES and RNA-seq analyses supported molecular preservation of primary tumor features in 3DP-PLC models, but the models were not identical to the parental tissues. Differences between paired samples may reflect sampling heterogeneity, tissue dissociation, short-term culture pressure, selective survival of specific tumor-cell states, and changes in stromal or immune-cell composition. These factors likely account for the limited molecular divergence observed between primary tumors and matched 3DP-PLC constructs and should be considered when interpreting ex vivo molecular profiles. Drug profiling across therapeutically relevant agents further revealed substantial interpatient heterogeneity, supporting the utility of 3DP-PLC for individualized pharmacologic assessment. However, the current platform also has limitations for immunotherapy-response modeling. Although immune cell populations were partially retained in day 10 3DP-PLC constructs and pilot PD-1 blockade experiments did not markedly compromise overall construct viability, the immune-cell abundance decreased during culture and the observed immune activation was modest. Therefore, the present 3DP-PLC system should not be considered a fully functional immune checkpoint inhibitor testing platform. Most ex vivo immunotherapy-response models require additional immune reconstitution, such as autologous PBMC or TIL co-culture [60–62], and optimized cytokine support [63]. Future studies incorporating these immune-preserving or immune-reconstituted strategies will be required to establish robust checkpoint inhibitor testing in 3DP-PLC models.

The central translational advance of this study is the development of a clinically anchored three-tier response stratification system. This concept builds on recent progress in the strategy reported by Yang et al., in which clinical ORR was used to define ex vivo sensitivity [44]. Rather than relying on arbitrary cutoffs, we integrated ex vivo AUC profiles with external clinical trial benchmarks to generate standardized, clinically interpretable response categories. To minimize data leakage, all response cutoffs were locked using external clinical trial benchmarks before correlation with patient radiological and follow-up data, enabling 3DP-PLC AUC profiles to be translated into independent and clinically interpretable response strata. This framework provides direct therapeutic guidance. Models classified as sensitive identify patients for whom the corresponding drug is strongly recommended. Relatively sensitive models define patients with marginal but potential benefit, making them prime candidates for rational combination strategies, such as targeted therapy plus immune checkpoint blockade. Resistant models identify patients unlikely to benefit from the tested drug. In this group, even targeted-immunotherapy combinations may provide limited efficacy, allowing clinicians to avoid ineffective treatment and unnecessary toxicity. Thus, the three-tier system better reflects the graded nature of clinical response and is more actionable than conventional binary classifications.

The analytical strategy is also highly generalizable. Because the classification is anchored to external clinical trial benchmarks rather than internally derived assay distributions alone, it can be adapted to other agents and tumor types when appropriate clinical response data are available. Bootstrap full-pipeline resampling and Dirichlet-based uncertainty propagation supported the overall robustness of the stratification schemes, although some drug-specific cutoffs showed local instability. Importantly, the system was more stable at the level of final response-group assignment than at the level of individual cutoff values, suggesting that category-level interpretation may be more clinically reliable than reliance on a single numeric threshold. In patients receiving neoadjuvant or adjuvant targeted therapy, 3DP-PLC classification showed clinical relevance, including concordance with neoadjuvant response and a directional association with DFS. Transcriptomic analyses further supported the resistant phenotype, linking resistance-associated programs to adverse prognosis, TKI resistance, pro-survival signaling, and suppression of apoptosis.

Beyond drug response intrinsic to cancer cells, our study also addressed the contribution of stromal context by establishing a patient-matched 3DP-PLC/CAF co-culture biobank. Although residual CAF/fibroblast populations were detectable in some primary tumor-derived 3DP-PLC monoculture models, these stromal components reflected the native cellular composition of individual tumors and were not experimentally controlled in abundance, activation state, or spatial organization. In contrast, the patient-matched CAF co-culture platform enabled controlled reintroduction of isolated and expanded CAFs, allowing paired assessment of CAF-mediated effects on tumor growth, spatial architecture, and therapeutic response [64–66]. Our paired co-culture drug screening data indicate that CAFs generally decrease ex vivo drug sensitivity in 3DP-PLC models; however, this effect was not uniform across patients or drugs. This observation is consistent with the increasingly recognized molecular and functional heterogeneity of CAFs [67]. We further evaluated whether these CAF-only single-cell analysis in PLC31 and PLC42 further suggests that CAF-mediated drug modulation may reflect patient-specific stromal states and CAF-to-malignant signaling programs rather than a single universal CAF mechanism. The marked separation between PLC31 and PLC42 in the co-culture scRNA-seq UMAP highlights the substantial inter-patient variability of primary liver cancer ecosystems. We therefore interpret the two-patient scRNA-seq dataset as a mechanistic, patient-resolved anchor rather than as evidence for a universal CAF-induced state. Patient-stratified analyses showed that PLC31 and PLC42 differed in CAF-state composition, malignant-cell resistance/stemness/EMT shifts and CAF-to-HCC signaling programs. These findings are consistent with previous liver cancer single-cell studies showing strong patient-specific malignant and microenvironmental states. Thus, our data support a context-dependent model in which patient-matched CAFs can remodel malignant-cell programs through distinct stromal states and ligand-receptor circuits. Nevertheless, the single-cell component of this analysis was limited to two mechanistic anchor cases, and the recovered co-culture CAF compartment was modest in size. Future experiments involving CAF-state isolation, ligand perturbation or spatially resolved validation will be required to define the causal contribution of individual CAF subsets to therapeutic resistance.

The overlap-aware analysis connected molecular fidelity, tumor-cell state and stromal pharmacology without assuming complete four-modality coverage. The higher matched than mismatched WES and transcriptomic similarities support preservation of patient identity in the 3DP-PLC models, whereas the absence of an FDR-significant WES pathway-drug association argues against a single genomic determinant of TKI response in this cohort. The nominal associations of 3DP EMT/TGF-beta and hypoxia activity with lenvatinib AUC instead indicate that ex vivo response may reflect model-state programs that are not captured by a single driver alteration. The predominantly positive, but bidirectional, CAF-associated AUC shifts further suggest that stromal context modifies this tumor-intrinsic response in a patient- and drug-dependent manner. The distinct CAF compositions, malignant-cell program shifts and communication patterns in PLC31 and PLC42 are consistent with this context-dependent interpretation rather than a uniform CAF mechanism.

Several boundaries qualify this interpretation. The modalities were only partially overlapping across the cohort, and PLC45 was the sole case represented in WES, bulk RNA-seq, scRNA-seq and drug-response datasets. The 8 HCC complete-case subset was therefore used for aligned visualization, not model training or formal four-modality inference. Sampling heterogeneity, tissue dissociation, short-term culture selection and shifts in cellular composition could also contribute to the observed molecular and pharmacological relationships. Fig. S17 should therefore be interpreted as an exploratory cross-modal evidence bridge, not as a causal chain or validated multi-omic predictor. Larger cohorts with same-patient profiling across all modalities, together with targeted perturbation experiments, will be required to test causality and clinical generalizability.

Conclusion

In this study, we established 3DP-PLC as a high-fidelity, scalable, and clinically relevant ex vivo platform for functional drug testing in PLC. By integrating patient-derived bioprinting, a standardized and clinically anchored response stratification workflow, and tumor-stroma co-culture modeling, our system enables the investigation of therapeutic heterogeneity at both the tumor cell and microenvironmental levels. More importantly, it offers a practical analytical system for translating continuous ex vivo drug-response data into reproducible and clinically interpretable categories. In this sense, the major contribution of our work is not only the development of a liver cancer model, but the establishment of a potential route by which functional drug testing may be incorporated into future prospective translational studies and, ultimately, into clinically meaningful treatment stratification strategies for PLC.

Abbreviations

2D

Traditional two-dimensional

3DP

Three-dimensional bioprinting

3DP-PLC

Three-dimensional bioprinted primary liver cancer

AFP

Alpha-fetoprotein

AlgMA

Alginate Methacrylate

ANOVA

One-way analysis of variance

AUC

Area under the dose-response curve

CAFs

Cancer-associated fibroblasts

CAM

Calcein-AM

CHC

Hepatocellular cholangiocarcinoma

CNV

Copy-number variation

CR

Complete response

ECM

Extracellular matrix

ECOG

Eastern Cooperative Oncology Group

FR

Fibrous ring

GelMA

Gelatin Methacryloyl

GVF

Goodness of variance fit

HAIC

Hepatic arterial infusion chemotherapy

HAMA

Methacrylate Hyaluronic Acid

HBV

Hepatitis B virus

HCC

Hepatocellular carcinoma

HCV

Hepatitis C virus

ICCA

Intrahepatic cholangiocarcinoma

MRI

Magnetic resonance imaging

OS

Overall survival

PD

Progressive disease

PDTOs

Patient-derived tumor organoids

PDX

Patient-derived xenograft

PFA

Paraformaldehyde

DFS

Disease-free survival

PI

Propidium iodide

PLC

Primary liver cancer

PR

Partial response

RECIST 1.1

Response Evaluation Criteria in Solid Tumors, version 1.1

SD

Stable disease

SNVs

Single-nucleotide variants

TACE

Transarterial chemoembolization

TKIs

Tyrosine kinase inhibitors

TME

Tumor microenvironment

Authors’ contributions

Z.K. conducted experiments and wrote the manuscript. S.M.H, S.X.Y., J.B., D.S.D. and Z.H.T. prepared and processed the clinical samples. F.X., Z.J.G., Y.H.Y., L.X.Y., D.J.X., L.Y.C., S.H., D.L.W., P.M.C and J.S.Z collected the patients information. Z.K., D.L.W., Y.H.Y, M.Y.L, and X.H.F conceived the study. Z.K. and Y.H.Y supervised the process of the study and revised the manuscript. Z.K. conducted experiments and wrote the manuscript.

Funding

This work was supported by the National Key R&D Program of China (No. 2024YFB4607800), National High Level Hospital Clinical Research Funding (2025-PUMCH-C-030) and Peking Union Medical College Hospital Talent Cultivation Program (Category C; No. UBJ10452).

Declarations

Ethics approval and consent to participate

The study was approved by the Institutional Review Board of Peking Union Medical College Hospital (approval No. I-26PJ1022) in strict accordance with the Declaration of Helsinki. Written informed consent was obtained from every participant before tissue resection.

Consent for publication

Not applicable. The manuscript does not contain any identifiable individual data in any form (including any individual details, images or videos).

Competing interests

The authors declare no competing interests.