ABSTRACT
-
Background/Aims
Liver transplantation (LT) following total hepatectomy is a life-saving treatment for hepatocellular carcinoma (HCC). The HCC recurrence after LT hinders the effectiveness of the procedure. The objective of this study is to develop a pre-operative risk stratification model based on a liquid biopsy.
-
Methods
We conducted a comprehensive multi-omics study of 260 HCC patients from three centers, including clinical data, low-coverage whole-genome sequencing of cell-free DNA (cfDNA) from plasma, as well as whole-exome, single-nucleus RNA, and spatial transcriptomics from matched tumor and non-tumor tissues.
-
Results
We identified cfDNA-derived copy number alteration (CNA) signatures associated with post-transplant recurrence. By integrating cfDNA-derived CNA profiles with single-cell transcriptomic data, we traced recurrence-associated cfDNA to a distinct subpopulation of malignant cells within the primary tumor. These cells were embedded in a pro-metastatic microenvironment of specialized endothelial subtypes and cancer-associated fibroblasts. Notably, most recurrence-associated lesions were detectable in cfDNA prior to liver transplantation (LT). Building on these insights, we developed the ZJU Criteria based on CNA fragments and tumor markers, a pre-LT risk prediction tool that integrates conventional clinical factors with cfDNA-derived CNA signatures, and validated it using internal and independent external cohorts.
-
Conclusion
Our findings suggest that post-transplant recurrence commonly originates from advanced subclones that emerge late during tumor evolution. The ZJU Criteria provides an accurate, non-invasive strategy that significantly improves pre-LT risk stratification and clinical decision-making for patients with HCC.
-
Keywords: Hepatocellular carcinoma; Liver transplantation; Cell-free DNA; Liquid biopsy; Multi-omics
Study Highlights
• cfDNA uncovers early genomic instability that precedes post-liver transplantation relapse.
• Recurrence-associated signatures from cfDNA can be traced to subpopulations in the primary tissue.
• Specific endothelial and CAFs programs characterize a metastatic-permissive microenvironment in which VEGFA and NOTCH increase HCC cell motility.
• A clinically applicable nomogram provides an individualized estimation of recurrence-free survival before liver transplantation.
Graphical Abstract
INTRODUCTION
Hepatocellular carcinoma (HCC) is the third leading cause of cancer deaths worldwide, with a 5-year survival rate of approximately 18% [
1]. Total hepatectomy followed by liver transplantation (LT) is the only effective treatment for patients with failing livers [
2,
3]. The procedure is intended to remove all malignant cells along with the original liver while the transplant restores liver functions, suitable for patients who have yet to develop vascular invasion or extrahepatic spread. The canonical Milan criteria set a threshold on the tumors of eligible patients [
4]. Extended criteria, such as the UCSF [
5] and Hangzhou [
6], similarly relied on morphological and traditional clinical parameters. Under current eligibility-test selections, recurrence still developed at significant rates [
7-
15]. Conversely, a proportion of patients whose condition exceeded common selection criteria but nonetheless received LT were cancer-free, suggesting that current preoperative risk assessments are inadequate.
The main cause for post-transplantation recurrence is occult metastases, which occurs when undetected cancer cells disseminate from the primary site to enter the bloodstream before seeding and growing at distal tissues. Surgical removal of the tumor-burdened liver is only curative when remote seedings have yet to take place. The exact timing of tumor cell dissemination and the duration of metastatic dormancy are difficult to determine and may vary among different cancer types [
16,
17]. Our previous work retrospectively profiled tumors from patients beyond the Milan criteria and showed that recurrence is fueled by a cancer-associated fibroblasts (CAFs)-governed microenvironment [
18]. But leveraging these findings for patient stratification would require invasive sampling methods.
Here, we took an agnostic approach to trace the cellular origins of disseminating HCC cells that lead to metastasis. We applied the term “circulating tumor DNA” to refer to the tumor-derived fraction of cell-free DNA (cfDNA). We used circulating tumor DNAs as a proxy of tumor dissemination [
19] and cross-examined the genetic lesions present in the primary, circulation and metastatic sites. We analyzed preoperative samples from HCC patients and found that circulating tumor DNAs, the tumor-derived fraction of cfDNA, recapitulate copy number alteration (CNA) patterns of the corresponding primary tumors. Through comparatives of the primary tumor and the metastatic tumor, we found that the metastatic tumor originated from a subclone in the primary tumor. By correlating cfDNA-detected and single-nuclei RNA sequencing (snRNA-seq)-inferred CNAs events, we traced circulating tumor cells to sub-populations in the primary tumor mass.
Previous studies showed that metastatic HCCs often bear similarities to the primary tumor [
20,
21], but the causality and hence the utility of primary genetic lesions in predicting recurrence propensities remained elusive. In the context of post-LT recurrence, we identified recurrence-associated lesions by WES and snRNA-seq data. We then showed that distinct endothelial cells and CAFs constitute a niche to facilitate metastatic malignancy by single-nuclei and spatial transcriptomics. Using integrated cfDNA-derived CNA data and clinical records with external cohorts’ validation, we hereby propose the ZJU Criteria to predict the recurrence risk. Collectively, our work combines the strengths of liquid biopsy and solid tumor multiomics to shed light on the underlying genomic lesions of recurrent HCC following LT.
MATERIALS AND METHODS
Ethical approval
We followed the World Medical Association’s Declaration of Helsinki. The research procedure was approved and supervised by the Ethical Review Committee of The First Affiliated Hospital, School of Medicine, Zhejiang University (20191421) and Huashan Hospital & Institute of Organ Transplantation of Fudan University (2024595) and the Ethics Committee of Clinical Research at Shulan (Hangzhou) Hospital (KY2021014). Written informed consents were signed by all participants.
RESULTS
cfDNA revealed genome instability preceding post-LT recurrence
The derivation cohort involved 122 patients from Shulan Hangzhou Hospital and validation cohort consisted of 138 patients from three centers (44 from Shulan Hangzhou Hospital, 43 from the First Affiliated Hospital, Zhejiang University and 51 from the Huashan Hospital & Institute of Organ Transplantation of Fudan University). The graphical abstract was shown (
Fig. 1A). Clinical information before LT in the derivation cohort was shown in
Figure 1B. Pre-transplantation plasma, primary tumor, paired non-tumor, and metastatic tissues were collected for multiomic analyses. To determine whether pre-transplant liquid biopsy was informative in the context of HCC recurrence, we collected pre-transplantation plasma of patients in the derivation cohort (recurrent, n=40; non-recurrent, n=82) and compared the genetic lesions captured by cfDNA. The proportions of patients with gain or loss in genome-wide fragments for recurrent and non-recurrent groups were shown (
Fig. 2A). Proportion of patients with CNAs in each group was quantified separately, with the ratio of these proportions presented (
Supplementary Fig. 1). Of note, amplifications in chr. 1, chr. 5 and chr. 7, along with deletions in chr. 9 and 16, were significantly more prevalent in the recurrent group compared to non-recurrent group patients, while amplifications of chr. 19 and 22q were observed in both groups, whose frequencies showed no significant differences (
Fig. 2A,
Supplementary Fig. 1). Sex chromosomes were excluded from analysis due to the male predominance in our patient cohort. Estimated tumor fraction of the cfDNAs was also significantly higher in the recurrent than the non-recurrent group (
P<0.01, Wilcoxon,
Fig. 2B). The CN heterogeneity score, a measurement of intertumoral heterogeneity and genomic instability [
22,
23], was also higher in the recurrent than the non-recurrent group (
P<0.05, Wilcoxon,
Fig. 2C). Collectively, recurrent group patients demonstrated more chromosomal gain/loss, indicating greater genomic instability compared to non-recurrent group patients. These data suggested that genomic CNA events could be detected in cfDNA samples and leveraged for recurrence prediction.
To determine whether cfDNA can be the proxy of disseminating tumor cells, the primary tumor and post-transplantation liver metastatic tissue from the same patient (P0158) were subjected to 200× coverage WES and compared to the low coverage sequencing results from cfDNA. We reasoned that mutational events shared by the primary tumor (
P) and the metastatic site (
M) could be traced to the disseminating cells in the blood stream, which later gave rise to the metastatic tumor (
Fig. 3A, top). Hence the overlapped events (
P∩
M) might be detected in the cfDNA and harbor predisposing factors of metastatic recurrence. Events only present in the primary tumor (
P＼
M) were from non-metastatic cells. Conversely, additional mutational events acquired by the metastatic tumor (
M＼
P) suggested continuous somatic evolution of post seeding. We found 335 mutations shared by the primary tumor (398 mutations, VAF cutoff=0.01) and the metastatic tumor (1,373 mutations, VAF cutoff=0.01) (
Fig. 3A, bottom) including known HCC trunk mutations (
Supplementary Table 1) reported by previous studies [
24,
25]. A fraction of these mutations was detected in cfDNA (27/335 mutations, 8.1%). The clonal history reconstructed based on a Bayesian clustering method [
26] showed that the primary and metastatic tumors were indeed derived from the same ancestral clone that harbored mutations in
BRCA2,
EP300,
KMT2D,
TCF7L2, and
BRD4 (
Fig. 3B,
Supplementary Table 2).
Like the mutational profiles, CNA profiles demonstrated shared ancestry of the primary and metastatic tumor samples (
Fig. 3C–
3E). Prominent CNA events were consistently detected in both primary and metastatic tumors, such as frequent amplification events in chr8q and deletion events across chr. 4 and chr. 13 (
Fig. 3C). The CNA trends (grouped by cytoband segments) in both samples were highly correlated (
Fig. 3E,
P<0.0001). Metastatic tumors exhibited increased genomic instability developed from the primary tumor (
Fig. 3E, r=0.5668), consistent with the notion that tumor cells progressively accumulate genetic lesions as they acquired metastatic malignancy to invade the metastatic site.
Our data showed that cfDNAs were more reliable in detecting dissemination-related CNA events (
Fig. 3D) than mutational events (
Fig. 3A). A similar overlapping pattern was found in the CNA profiles of the primary and metastatic tumors (7,944 primary; 9,377 metastatic; 3,889 overlapped events. Cutoff, |log
2R|>0.3). But a larger number and a higher fraction of the intersecting CNA events were detected in cfDNA (736/3889 CNAs, 18.9%; cfDNA CNA cutoff, |log
2R|>0.1, see
Supplementary Table 3). While these data suggest a clonal relationship between the primary tumor and metastatic lesions, the limited availability of metastatic samples necessitates that these findings be viewed as preliminary. Nonetheless, our results indicate that the CNA profile of cfDNA may potentially serve as a surrogate for circulating tumor cells.
Circulating cancer cells account for only a small proportion of the primary tumor mass. We hypothesized that the shedding primary HCC cells were not randomly picked but evolved from subclones of the primary tumor and could be lineage-traced. To reveal the cellular composition of the primary HCC tumors, we constructed single nuclei profiles of 12 total-hepatectomy specimens. 118,200 single nucleus passed quality controls (
Supplementary Fig. 2A–
2E). Nine major cell clusters were identified by the Uniform Manifold Approximation and Projection (UMAP) (
Fig. 4A and
Supplementary Fig. 2F). We assumed that the hepatocytes (25.88%), cholangiocytes (2.17%), and malignant (45%) clusters were potential tumor-initiating populations and that the B cells (0.33%), NK/T cells (3.57%), myeloid (6.90%), CAFs (4.90%), endothelia (8.73%), and other (2.52%) clusters were non-tumor lineages. All cell types were present in each patient (
Supplementary Fig. 2G,
2H) from both non-recurrent and recurrent groups (
Fig. 4A, right). The high expression levels of hepatocyte genes, such as
TTR and
APOA2, in the malignant cluster were consistent with the HCC diagnoses. The malignant cells were regrouped into to seven unbiased sub-clusters (Malignant 1–7) (
Fig. 4B, left), each characterized by differentially expressed genes (
Fig. 4C). Of note, these sub-clusters were unevenly distributed in non-recurrent and recurrent groups (
Fig. 4B, right). The recurrent group was most prominently represented by Malignant 3, followed by Malignant 5. Malignant 3 cells were characterized by
VEGFA,
MAPK9,
ASPM, and
CDC27. Malignant 5 were characterized by
DYNC1I1,
DPP4,
CDK6,
SDHAF3, and
PHF14. Malignant 2 enriched hepatocyte-like metabolic genes
APOH,
CYP3A4,
APOA1, and
AMBP. These malignant single nucleus transcriptomes showed a crude correlation with the bulk RNA-seq-based subtyping of HCC (
Supplementary Fig. 2M). Interestingly, the VEGFA
+ Malignant 3 cells loosely resembled the Hoshida subclass S1, which exhibited more vascular lesions and satellite lesions, and was considered invasive/disseminative [
27]. Enrichment analysis of differential expression genes in malignant subtypes was performed using the Kyoto Encyclopedia of Genes and Genomes (KEGG). Malignant 3 subtype showed elevated cell cycle, DNA replication, proliferation and EMT genes (
Fig. 4D,
4E), suggesting that tumor cells from malignant 3 subtype had stronger proliferation and growth ability and were more likely to invade and metastasize.
To trace the cells in the primary tumor that harbored the circulating recurrence-associated genetic lesions, we first reconstructed the CNA profiles of individual cells with the snRNA-seq data (
Fig. 5A) using inferCNV [
28]. The inferred CNA pattern resembled the pattern generated by WES (
Supplementary Fig. 3) of the bulk primary tissue. The CNA comparisons of primary tumor WES data across samples are provided in the
Supplementary Table 4.
We next filtered cfDNA-derived binwise CNA data and recalculated those within the coding region on a per-gene basis for comparison. Pearson correlation coefficients were calculated between the CNA patterns of single nuclei and the corresponding cfDNA-derived CNA pattern from the same patient (cutoff:
P<0.01) and then mapped to the UMAP space of malignant populations from transcriptome clustering (
Fig. 5B). We found that the Malignant 3 subcluster was most correlated to the tumor fraction of cfDNA. cfDNAs were largely composed of cells from Malignant 3 (38.5%) and Malignant 5 (26.5%) (
Fig. 5C). These data demonstrated the consistency between cfDNA-based CNA profiling and gene expression pattern of the putative metastatic population, suggesting that a group of malignant subclone within the primary tumor were more likely to enter the bloodstream and metastasize to other sites, ultimately leading to recurrence.
Interaction between tumor and niche cells is critical for malignant and metastatic potencies [
29,
30]. Based on single nuclei transcriptomic profiling (
Fig. 4), we examined major niche cell types, including the immune cells, endothelia, and CAFs. A total of ten immune clusters emerged, including Kupffer cells, macrophages (cluster 1 and cluster 2), conventional dendritic cells, T cells (effector CD8
+ T and exhausted CD4
+ T), NK (regulatory NK and cytotoxic NK), cells, plasma B cells and B lymphocytes (
Fig. 6A and
Supplementary Fig. 4B). All the immune subtypes were present across patients from both recurrent and non-recurrent groups (
Supplementary Fig. 4C). Overall, the fraction of myeloid was larger than that of lymphoid cells, and Kupffer cells were the largest subtype in each sample. The proportion of different immune cells in each sample was shown (
Supplementary Fig. 4D,
4E). A total of six endothelial clusters emerged (
Fig. 6B), all of which were shared across patients, albeit at different proportions (
Supplementary Fig. 5A–
5D). The endo2 and endo5 sub-clusters expressed high levels of
CDH13, which was known to drive angiogenesis and promote tumor-stromal crosstalk [
31]. The CLEC4G
+ endo1 and RHOJ
+ endo6 clusters showed features of liver sinusoidal endothelial cells (LSECs), whereas PECAM1
+CDH13
+ endo2, PRCP
+ endo3, RAMP3
+ endo4 and CDH13
+CASC5
+ endo5 closely resembled liver vascular endothelial cells (LVECs) (
Supplementary Fig. 5B). LSECs were the major endothelial cell type in all normal tissues, whereas LVECs became the predominant population specifically in the recurrent tumor group, indicative of occult vascular invasion (
Supplementary Fig. 5E).
To identify niche signals stimulating metastasis, we evaluated tumor-niche cell interactions by CellPhoneDB [
32] with single nuclei RNA-sequencing data. Known ligand-receptor pairs were used to infer interacting populations. Malignant cells of the recurrent group showed prominent interactions with endothelial and CAFs clusters, compared to their non-recurrent counterparts (
Fig. 6C and
Supplementary Fig. 6). Among all malignant cells, Malignant 3 interacted more strongly with CAFs and endothelial cells than with other Malignant subsets. Notably, VEGFA
+ Malignant 3 cells received strong NOTCH-activating signals from endothelial cells and CAFs via the DLL4 and JAG1/2 ligands in the recurrent group, respectively (
Fig. 6D). By contrast, the interaction of Malignant 3 cells and CAFs was weaker in the non-recurrent group.
We verified these interactions with spatial transcriptomics with 10×Genomics Visium platform after sufficient quality control (
Supplementary Fig. 7). Based on unbiased clustering, spots were annotated (
Fig. 6E left, F left, and
Supplementary Fig. 8A–
8D). The vicinity of spots was measured by the Neighborhood score, defined as the ratio of endothelial spots surrounding a given VEGFA
+ Malignant spot over total surrounding spots (
Fig. 6E, middle, Neighborhood=
Nendo/
Ntotal), as well as the Proximity score, defined as the distance between the endothelial spot closest to a given VEGFA
+ malignant spot (
Fig. 6E, middle, Proximity=1/[1+Distance]), as in a previous study [
33]. The recurrence group demonstrated increased CAFs neighborhood, CAFs proximity, endothelial neighborhood, and endothelial proximity scores in contrast to the non-recurrent group (
Fig. 6E,
6F, right). The VEGFA
+ NOTCH2
+ malignant and DLL4
+ endothelial spots were in proximity, specifically in the primary tumors of the recurrent group (
Supplementary Fig. 8E–
8G), suggesting that malignant cells gained metastatic potency via niche NOTCH signaling. A hallmark component of tumor microenvironment is CAFs, which exhibited high levels of plasticity during cancer development [
34]. Of note, PDGFRB
+ CAFs were known to play a prominent role in promoting angiogenesis in tumors [
35]. Using spatial transcriptomics, we confirmed interactions between CAFs and HCC cells, as well as between endothelial cells and HCC cells (
Fig. 6E,
6F, and
Supplementary Fig. 8). The vicinity of malignant spots was measured by the neighborhood score and proximity score (
Fig. 6E,
6F), many malignant spots having high CAFs neighborhood and proximity scores. Spatial mapping revealed that the VEGFA
+NOTCH2
+ malignant cells and PDGFRB
+JAG1
+ CAFs were in proximity, especially in primary tumors of the recurrent group, indicating potential pro-metastatic interactions between these cells via the NOTCH2-JAG1 signaling pathway (
Supplementary Fig. 8F). Many malignant spots having high endothelial neighborhood and proximity scores (
Fig. 6E,
6F) and VEGFA
+NOTCH2
+ malignant and DLL4
+ endothelial cells were in proximity (
Supplementary Fig. 8G), indicating potential pro-metastatic interactions between these cells via the NOTCH2-DLL4 signaling pathway. Taken together, the Malignant 3 subset showed the strongest interaction among all malignant cells and endothelia, and CAFs were the most significant niche cell types that promoted recurrence-associated malignancy.
To validate our findings from transcriptomic sequencing, we employed transwell migration assays. These assays demonstrated that co-culture with CAFs significantly enhanced the migration of human HCC cell lines HuH-7 and SNU-449 (
Fig. 7A: a and e,
Fig. 7C: a and e). Consistently, RT-qPCR analysis revealed a marked upregulation of
NOTCH1,
NOTCH2, and
NOTCH3 expression in both HuH-7 and SNU-449 cells under CAFs co-culture conditions (
Fig. 7B top and
Fig. 7D top). Our prior transcriptomic analysis had identified a tumor cell subpopulation (Malignant 3) characterized by high VEGFA expression and increased metastatic potential. To functionally validate this, we added recombinant VEGFA protein to HCC cells. This treatment resulted in a significant increase in cell migration (
Fig. 7A: a and c,
Fig. 7C: a and c). In contrast, treatment with the NOTCH pathway inhibitor DAPT significantly reduced the migratory ability of HCC cells (
Fig. 7A: a and b,
Fig. 7C: a and b). Notably, DAPT treatment also effectively abrogated the enhanced migration induced by CAFs co-culture (
Fig. 7A: e and f,
Fig. 7C: e and f). Similarly, the increased migratory capacity of HCC cells induced by VEGFA stimulation was significantly suppressed by DAPT treatment (
Fig. 7A: c and d,
Fig. 7C: c and d). The quantitative results of the transwell assay are shown (
Fig. 7B bottom and
Fig. 7D bottom). Collectively, these findings indicate that interactions between CAFs and HCC cells promote tumor cell migration. This pro-migratory phenotype involves the activation of VEGFA and NOTCH signaling, thereby providing experimental validation for the metastasis-associated programs identified in our transcriptomic analysis.
In the derivation cohort, we integrated clinical records and cfDNA data to predict recurrent risks pre-operatively. The workflow of model generation was shown (
Supplementary Fig. 9A). For cfDNA data, CNA fragments were first filtered during data preprocessing, and the retained candidate fragments were subsequently subjected to feature selection using LASSO regression. Two CNA fragments highly associated with post-LT recurrence emerged (see
Fig. 8A legends). The details of the two fragments were in
Supplementary Table 5. In spatial transcriptomics data, genes located within fragment 1 of chr. 7, such as
TTYH3,
MRM2,
EIF3B, and
ITFG1, showed higher expression in the recurrence group (
Supplementary Fig. 10A–
10C). Further validation using single-cell transcriptomics revealed that a subset of genes from chr. 7 fragment 1 (
SNX8,
TTYH3,
AMZ1,
GNA12 and
NUDT1) were highly expressed in Malignant 3. Conversely,
ITFG1 and
PHKB from chr. 16 fragment 2 were downregulated in the Malignant 3 subpopulation (
Supplementary Fig. 10D).
VEGFA and
NOTCH2 exhibited higher expression in the recurrence group, with particularly elevated levels in the Malignant 3 subpopulation (
Supplementary Fig. 10E,
10F). These findings are consistent with our previous observations. To identify important clinical variables predicting recurrence after liver transplantation for hepatocellular carcinoma, we performed univariable Cox regression analysis. Among the available clinical variables, only five (PIVKA-II, AFP, ALP, tumor number and creatinine) exhibited a hazard ratio [HR]>1 with statistical significance (
P<0.05) in univariable Cox regression analyses (
Supplementary Fig. 9C). Consequently, all factors significant in the univariable analysis were included in a subsequent multivariable Cox regression model to adjust for potential interferences. In the multivariable analysis, only fragment 1, fragment 2, PIVKA-II and AFP remained significant predictors of recurrence. The forest plot represented multivariable Cox regression results of the two cfDNA CNA fragments and two key clinical variables (PIVKA-II and AFP), with the hazard ratios and their 95% confidence interval with
P-values shown (
Fig. 8A, left).
We hereby propose the ZJU Criteria based on a nomogram providing a visual representation of the predictive model and facilitating clinical implementation in personalized risk assessment (
Fig. 8A, right). Specifically, we suggest the nomogram score ≤100 as the cut-off for selecting HCC patients for LT. Clinically, this cut-off is intended to be applied as a complementary risk stratification tool alongside established morphological criteria (such as Milan or UCSF), rather than as a standalone replacement. In the derivation cohort (Shulan Hangzhou Hospital), the model demonstrated strong discriminatory ability, with a Harrell’s concordance index (C-index) of 0.802. The bootstrap validation (1,000 resamples) indicated minimal model optimism (mean≈0.01). The optimism-corrected C-index was 0.792 (apparent: 0.802), demonstrating low overfitting and good model stability (
Supplementary Fig. 9B). Time-dependent receiver operating characteristic analysis yielded area under the curve (AUC) values of 0.807 (95% confidence interval [CI] 0.709–0.905) at 1 year, 0.841 (95% CI 0.758–0.924) at 3 years, and 0.817 (95% CI 0.721–0.913) at 5 years (
Fig. 8B, right). Patients with a nomogram score >100 exhibited significantly poorer outcomes, with a 5-year overall survival (OS) rate of 36% and an HCC recurrence rate of 59%. Conversely, those scoring ≤100 demonstrated markedly improved survival (84% 5-year OS) and lower recurrence (17% HCC recurrence) (
Fig. 8B, left and middle). Thus, a nomogram score ≤100 identifies a subgroup of patients with favorable biological behavior who may be considered appropriate LT candidates when evaluated in conjunction with existing clinical selection frameworks. In the internal validation cohort (Shulan Hangzhou Hospital), the model achieved a C-index of 0.745. Time-dependent AUC values were 0.821 (95% CI 0.675–0.967), 0.758 (95% CI 0.594–0.922), and 0.766 (95% CI 0.602–0.930) at 1, 3, and 5 years, respectively (
Fig. 8C, right). Consistent with the derivation cohort, high-score patients (>100) showed reduced 5-year OS (36%) and elevated recurrence (60%), whereas low-score patients (≤100) achieved 76% 5-year OS with 21% recurrences (
Fig. 8C, left and middle). In the external validation cohort I (First Affiliated Hospital, Zhejiang University), validation in an independent new center cohort yielded a C-index of 0.757. AUC values were 0.802 (95% CI 0.631–0.973), 0.759 (95% CI 0.605–0.913), and 0.745 (95% CI 0.574–0.916) at 1, 3, and 5 years (
Fig. 8D, right). The prognostic stratification remained robust: high-score patients (>100) had 38% 5-year OS and 63% recurrence, compared to 73% OS and 22% recurrence in low-score patients (≤100) (
Fig. 8D, left and middle). More critically, to stringently test the model’s generalizability, we additionally collected and introduced a completely independent external validation cohort from Huashan Hospital & Institute of Organ Transplantation of Fudan University. The model demonstrated strong discriminatory ability, with a C-index of 0.861. Time-dependent AUC values of 0.864 (95% CI 0.715–1.013) at 1 year and 0.815 (95% CI 0.655–0.975) at 3 years (
Fig. 8E, right). As the cohort at Huashan Hospital & Institute of Organ Transplantation of Fudan University exclusively included patients from 2021 onward, the follow-up time was insufficient. This precluded the validation of the model’s accuracy in predicting recurrence within a 5-year timeframe. Patients with a nomogram score >100 exhibited significantly poorer outcomes, with a 3-year OS rate of 60% and HCC recurrence rate of 45%. Conversely, those scoring ≤100 demonstrated markedly improved survival (86% 3-year OS) and lower recurrence (16% HCC recurrence) (
Fig. 8E, left and middle).
Our model outperformed previous criteria including Milan [
4] and UCSF [
5] (AUC=0.642 and 0.653, respectively). Validation cohort results demonstrated sustained stability in predictive accuracy. This nomogram enables clinicians to generate a total point value (nomogram score) for each patient based on individual characteristics, facilitating pre-LT assessment. Calibration curves showed that the solid red line (nomogram predictions) was close to the ideal line in the internal, external I and external II validation cohorts (
Fig. 8F;
Supplementary Fig. 9D–
9K). This nomogram score demonstrates significant clinical utility in stratifying postoperative patients, enabling identification of those with markedly improved 5-year OS and reduced HCC recurrence rates. Our model integrated two cfDNA CNAs fragments and two key clinical variables (PIVKA-II and AFP), and its prediction performance was better than the model with only clinical variables (
Supplementary Fig. 11A–
11E).
DISCUSSION
The critical unmet need for LT treatment for HCC patients lies in an accurate and cost-effective strategy to identify which patients harbor occult metastatic potential. Here, we resolve this challenge by establishing a non-invasive framework for pre-LT recurrence risk assessment, based on a robust biological foundation. The identified predictive signatures serve as robust surrogates for systemic chromosomal instability. Their potential functional contributions to tumor recurrence warrant further mechanistic investigations. By integrating low-coverage sequencing of cfDNA with multi-omics dissection of tumor microenvironment, we achieve three pivotal advances: Proving the clinical utility of cfDNA-based chromosomal instability signatures as the harbingers of recurrence, detectable prior to LT. Uncovering the cellular and microenvironmental drivers of metastasis through single-cell and spatial transcriptomics; Developing and validating the ZJU Criteria that synergize cfDNACNA profiles with clinical variables.
Liquid biopsy has emerged as a non-invasive, potentially cost-effective screening method [
36]. Due to the low fraction of the cancer genomes in the blood samples, WGS of cfDNA is most effective in detecting broad CNAs. To what extent do cfDNA-derived CNA markers reflect causative somatic variations depending on the cancer types under investigation [
37-
39]. In this work, we focused on cancer recurrence in HCC patients post-LT. We adopted a strategy to combine the strengths of cross-sectional cohort study design and longitudinal multi-omics study design [
40]. HCC is known to exhibit high levels of chromosomal instability and more than 40% of the cases harbor high loads of broad CNAs [
41]. By comparing the broad CNA profiles of the recurrent and non-recurrent groups from cfDNAs, we identified specific gain/loss patterns of chromosome fragments associated with metastatic recurrence risks. Some of these CNA patterns have been reported in the context of HCC malignancy, e.g., frequent amplification of chr. 1q and deletion of chr. 16q [
42,
43]. Multiple recent studies also demonstrated correlations between HCC occurrence and cfDNA fragmentomes [
44] and/or methylation patterns [
45].
Unlike prior correlative studies, our multi-omics approach definitively traces recurrence-associated cfDNA alterations to distinct malignant subclones within primary tumors. These subclones, characterized by specific CNA patterns, reside within a pro-metastatic niche orchestrated by PDGFRB+ CAFs and specialized endothelia. Critically, ligandreceptor analyses revealed a feed-forward loop wherein malignant cells secrete VEGF to recruit endothelia, while endothelia and CAFs reciprocally activate NOTCH signaling in tumor cells via JAG1/DLL4. This niche not only fuels subclone aggressiveness but also facilitates vascular intravasation, explaining why these cells dominate cfDNA and seed future metastases. Even though snRNA-seq/spatial datasets are not directly involved in model training, they provide important mechanistic insights that pre-LT cfDNA profiles capture the genomic footprint of metastasis-competent subclones embedded in a high-risk microenvironment.
Building on these insights, the present study highlighted the role of cfDNA-derived CNA signatures in predicting the recurrence of patients with hepatitis B virus (HBV)-related HCC after LT. The ZJU criteria are intended to complement existing morphological selection criteria (such as the Milan/UCSF criteria). In clinical decision-making, the nomogram score serves as an additional biological dimension of risk assessment: patients who meet conventional morphological criteria but exceed the ZJU cut-off (>100) may warrant heightened caution, intensified surveillance, or alternative therapeutic consideration, whereas selected patients exceeding morphological thresholds but falling within the ZJU cut-off (≤100) may still be considered for LT in carefully evaluated settings (e.g., living-donor transplantation). It is crucial to adhere to a strict postoperative follow-up and provide any necessary supportive treatments. Potential recipients who exceed the morphological criteria but meet the ZJU criteria should still be carefully considered for LT, and living donor LT may be an option. It is important to note that the study’s sample size was limited, and the cohort was exclusively derived from a Chinese population with HBV as the predominant etiology. Consequently, the generalizability of the findings may be constrained. Specifically, while chromosomal instability is a common hallmark across HCC etiologies, the molecular pathogenesis, immune contexture, and tumor-microenvironment interactions in HCC driven by other etiologies (such as HCV infection or MASLD) may differ from those observed in HBV-related disease. Therefore, the external application of the ZJU criteria, particularly to non-HBV populations or other regions, should be approached with caution. Prospective validation in etiologically diverse cohorts will be required to determine whether the predictive performance and optimal cut-off of the ZJU Criteria are conserved across HCC subtypes. We recommend that these criteria be validated in broader, independent cohorts and utilized strictly as a supplementary reference, integrated with local guidelines and comprehensive clinical assessment. To date, there is a scarcity of preoperative models to predict recurrent HCC after total hepatectomy and LT; a few others predict the outcome of hepatectomy in liver cancer patients [
46-
50]. Current recurrence prediction relies on anatomical metrics or invasive tissue analyses, neither of which dynamically reflects metastatic potential. The ZJU Criteria overcome these limitations by leveraging accessible liquid biopsies to quantify genomically defined risk. This translates to actionable clinical benefits, including both precise patient selection that avoids futile transplants and expanding transplant access. The validated, non-invasive tool presented by this study enables clinicians to optimize transplant candidacy, a critical step toward personalized, precision medicine in HCC.
FOOTNOTES
-
Authors’ contributions
SY, YW, SLing, ZW, RW, JJ, BZ, CJ, DL, HJ, HS, YZX, JL, SLin, and XX designed the study; YW, JL, JW, ZYang, XZ, ZL, JC, JW, HX, LZ, LC, GC, DC, SZ, JGWang, YZheng, KW, DL, QS, and XW acquired the clinical samples and data; SY, QH, YZhuang analyzed the sequencing data; SY, YW, YZhuang, CJ, and HJ performed statistical analyses; JPWang and FL performed functional analyses, SY, SLin, YW, HS, and XX generated figures and wrote the manuscript; XX, SLin, SZ, and YZX supervised the study.
-
Acknowledgements
This study was supported by grants from the National Key R&D Program of China (2021YFA1100500, XX, SLin & YZX; 2023YFA1800600, SLin; 2023YFC2505900, ZXW), the National Natural Science Foundation of China (82073071 and 32270885, SLin; 92159202, XX; 82241225, ZXW), the China Postdoctoral Science Foundation (2022M722762, YW). We thank Dr. Jianping Jin (Zhejiang University) and Dr. Guoliang Qing (Wuhan University) for critical comments. We thank Cosmos Wisdom Biotech (Hangzhou) for technical inputs on bioinformatics.
-
Conflicts of Interest
The authors have no conflicts to disclose.
SUPPLEMENTARY MATERIAL
Supplementary material is available at Clinical and Molecular Hepatology website (
http://www.e-cmh.org).
Supplementary Figure 1.
The prevalence ratio of CNAs between the recurrent and non-recurrent groups. Each vertical bar represents the ratio of the proportion of patients with CNAs in the recurrent group over the proportion in the non-recurrent group. The centromere position is indicated in red. Each bar covers a 1 MB genomic region.
Supplementary Figure 2.
Quality control and cell clusters of single-nuclei transcriptomes. (A) Violin plots showing the number of genes expressed in each cell of single-nuclei transcriptomes. (B) Violin plots showing the total number of read counts in each cell of single-nuclei transcriptomes. (C) Violin plots showing the proportion of mitochondrial genes in each cell of single-nuclei transcriptomes. (D) Scatter plot showing the correlation of the number of expressed genes to the total number of read counts in each cell. (E) Scatter plot showing the correlation between the proportion of mitochondrial genes and the total number of read counts in each cell. (F) Heatmap showing the differential expression of selected gene sets among cell types. (G) UMAP plot of single nuclei transcriptome of all samples, origins of patients shown by colors. (H) UMAP plot of single nuclei transcriptome of each patient. (I) UMAP plot of malignant single nuclei transcriptome of each patient. (J) UMAP plot of malignant single nuclei transcriptome of all samples, origins of patients shown by colors. (K) Histogram showing the cell proportions of each HCC patient. Hepatic parenchymal cells and non-hepatic parenchymal cells are shown in left and right histograms, respectively. (L) Histogram showing the malignant subtypes’ proportions of each HCC patient. (M) The UMAP visualization showing malignant single nuclei transcriptome of all samples; origins of the Hoshida S1–S3 subtypes shown by colors.
Supplementary Figure 4.
Unsupervised clustering of immune cells in HCC. (A) UMAP projections of immune single nuclei transcriptome of all samples; origins of patients shown by colors. (B) Heatmap showing the differential expression of canonical marker genes among immune subtypes. (C) UMAP projections of immune single nuclei transcriptome of all samples; origins of patient groups shown by colors. (D) UMAP plot of immune single nuclei transcriptome of all samples; origins of patient groups shown by colors. (E) Boxplot indicating the fractions of immune subtypes in the paired tumor (left) and non-tumor (normal) tissues (right) of each patient.
Supplementary Figure 5.
Unsupervised clustering of endothelial cells in HCC. (A) UMAP plot of endothelial single nuclei transcriptome of all samples, origins of patients shown by colors. (B) UMAP plot of endothelial single nuclei transcriptome of all samples, origins of LVEC (red) and LSEC (blue). (C) UMAP plot of endothelial single nuclei transcriptome of all samples, origins of different patients shown by colors. (D) Violin plot indicating the differential expression of the marker genes between the LVEC and LSEC of single-nuclei transcriptomes. (E) Pie chart showing the distribution of endothelial subtypes of paired tumor (bottom) and non-tumor (normal) (top) tissues of each patient.
Supplementary Figure 6.
Interactions between cell clusters. Cell-Cell interactions were inferred by the numbers of common ligand-receptor pairs using CellPhoneDB on snRNA-seq data from primary tumor tissue of each patient. Non-recurrent patients: P0160 and P0161; recurrent patients: P0159, P0158, P0150, and P0131.
Supplementary Figure 8.
Spatial transcriptomic atlas of human HCC tissues revealed a pro-metastatic niche. (A) Unbiased clustering of spatial transcriptomic spots and definition of cell types of each cluster in primary tumor tissue of HCC patient P0146. (B–D) Dot plots showing the average expression of selected marker genes that annotate the clusters of spatial transcriptomes in indicated clusters of HCC patients P0146 (B) P0068 (C) and P0159 (D). The size and color of each dot indicate the proportion of expressed cells and expression of intensity in each cluster, respectively. (E–G) Malignant spots expressing NOTCH2 and endothelial spots expressing DLL4 were juxtaposed (E, G). Malignant spots expressing NOTCH2 and CAFs spots expressing JAG1 were juxtaposed (F).
Supplementary Figure 9.
Nomogram model. (A) Workflow of the nomogram model. (B) The histogram showing distribution of optimism in bootstrap resamples. A total of 1,000 bootstrap resamples were used to estimate model optimism. The red dashed line indicates the mean optimism (approximately 0.01), suggesting minimal optimism bias. Apparent C-index: 0.802, Optimism-corrected C-index: 0.792, Estimated Optimism: 0.01. The minimal optimism of 0.01 indicates a very slight degree of overfitting. (C) The forest plot represents univariable cox regression analysis of key cfDNA-derived CNA fragments and clinical variables. (D–K) Calibration curves of 1-year, 3-year and 5-year recurrent-free survival in derivation cohort (D, E, F). Calibration curves of 3-year and 5-year recurrent-free survival in internal validation cohort (G, H). Calibration curves of 3-year and 5-year recurrent-free survival in external validation cohort I (I, J). Calibration curve of 3-year recurrent-free survival in external validation cohort II (K).
Supplementary Figure 10.
Expression of genes located in fragments 1 and 2. (A) The diagram illustrates the genomic span and relative positions of genes within fragment 1 or fragment 2. Gene structures are shown according to UCSC Genome Browser RefSeq annotations, with gene bodies spanning from transcription start sites to transcription end sites. (B, C) Spatial transcriptomics analysis revealed that genes upregulated in fragment 1 were expressed at higher levels in the recurrence group compared to the non-recurrent group (B), which was further quantified and confirmed (C). The red point indicates the mean. (D) single-cell transcriptomics showing a subset of genes from chr. 7 fragment 1 (SNX8, TTYH3, AMZ1, GNA12 and NUDT1) were highly expressed in Malignant 3. Conversely, ITFG1 and PHKB from chr. 16 fragment 2 were downregulated in the Malignant 3 subpopulation. (E, F) VEGFA and NOTCH2 expression in spatial transcriptomics (E) and single-cell transcriptomics (F).
Supplementary Figure 11.
Comparison of models: only two clinical variables. (A) Comparison of predictive performance between the two models. (B–E) ROC curve and AUC of the model with only two clinical variables. Derivation cohort (B), internal validation cohort (C), external validation cohort I (D), and external validation cohort II (E). ROC, receiver operating characteristics; AUC, the area under ROC curve.
Figure 1.The study design and patient demographics. (A) Schematic representation of the study. (B) Heatmap showing key clinical features before LT of recruited patients. Age and survival time shown in years and months, respectively; largest tumor measured by diameter, cm; AFP, ng/mL; PIVKA-II, mAU/mL. AFP, alpha-fetoprotein; ALT, alanine aminotransferase; AST, aspartate aminotransferase; AUC, area under the curve; cfDNA, cell-free DNA; CNA, copy number alteration; MELD, model for end-stage liver disease; snRNA-seq, single-nuclei RNA sequencing; WES, whole exome sequencing; WGS, whole genome sequencing.
Figure 2.cfDNA revealed genomic characteristics of metastatic recurrence. (A) Chromosome pairs showing the proportion of patients with copy number alterations (CNAs) in the recurrent group (top) and the non-recurrent group (bottom). Red and blue segments indicate chromosomal regions with copy number gain or loss, respectively. Color intensity reflects the proportion of patients affected within each group. Purple-highlighted regions denote informative CNA fragments, defined as 1-Mb genomic segments in which the ratio of the proportion of patients harboring CNAs in the recurrent group to that in the non-recurrent group exceeds 3. R, recurrent group; N, non-recurrent group. (B, C) Comparison of tumor purities (B) and copy number heterogeneity scores (C) in pre-transplantation cfDNA between recurrent and non-recurrent HCC patients. cfDNA, cell-free DNA; HCC, hepatocellular carcinoma. **P<0.01, *P<0.05, unpaired Wilcoxon test.
Figure 3.Mutation and CNA profiles revealed the sub-clonal origin of metastatic tissue in the primary tumor. (A) Square Venn diagram showing the overlap of somatic mutations between the primary and metastatic tissues, as well as their presence in pre-operative cfDNA. Each grid represents a mutational event. Blue and brown boxes enclosed events found in the primary and metastatic tissues, respectively; the intersection represents common events shared by the primary and metastatic tissues. Red shades indicate the presence of an event detected by cfDNAs, measured by variant allele frequency. Data from patient P0158. (B) Phylogenetic tree of P0158 tumors. Each node represents a subclone. Blue and brown represent subclones specific to primary and metastatic tumors. Important genes in each subclone are shown. The length of lines indicates the number of mutations. (C) The representative somatic CNA landscapes in primary and metastatic HCC tissues compared to nontumoral matching tissue by WES (top: from the primary tumor tissue of recurrent HCC patient P0158; bottom: from the metastatic tumor tissue of recurrent HCC patient P0158, respectively). (D) Square Venn diagram showing the overlap of CNAs between the primary and metastatic tissues, as well as their presence in pre-operative cfDNAs. Each grid represents a CNA event. Red shades indicate the presence of an event detected by cfDNAs, measured by the absolute value of the log-ratio of local tumor to normal read depth (|log2R|). Data from patient P0158. (E) CNAs detected in primary and metastatic tumors were relevant (Pearson, P<0.0001, r=0.5668). Each dot indicated the log2R value of separate cytoband segments. cfDNA, cell-free DNA; CNA, copy number alteration; HCC, hepatocellular carcinoma; WES, whole exome sequencing.
Figure 4.Tumor cell clusters revealed by snRNA-seq. (A) UMAP plot of single-nuclei transcriptomes of all samples, clustered to 9 main cell types (left). Origin of patient groups shown by colors (right). CAFs, cancer-associated fibroblasts. (B) UMAP plot of malignant single-nuclei transcriptomes, clustered to 7 subtypes (left). Origin of patient groups shown by colors (right). (C) Heatmap showing the differential expression of marker genes among malignant cell subtypes. (D) KEGG (Kyoto Encyclopedia of Genes and Genomes) analysis in each malignant subtype. GeneRatio, percentage of specific pathway genes presented in differentially expressed genes; P.adjust, adjusted P-value. (E) Proliferation and EMT scores among different malignant subtypes. CAF, cancer-associated fibroblast; EMT, epithelial-mesenchymal transition; snRNA-seq, single-nuclei RNA sequencing; UMAP, Uniform Manifold Approximation and Projection.
Figure 5.Identification of subclones in the primary tumor that gave rise to recurrence-associated CNAs in cfDNA. (A) Heatmap illustrating inferred CNAs across the genome, where each row represents an individual malignant cell, and each column corresponds to genomic regions ordered by chromosome (chr. 1–22). Copy number gains shown in red and losses in blue. Single cells were grouped by patient origin, as indicated on the left, with recurrent and non-recurrent cases annotated accordingly. At the bottom, whole-exome sequencing-derived CNA profiles from the corresponding patients were displayed, with matching color codes indicating patient identity. (B) The UMAP plot of malignant cells clusters showing correlation between snRNA-seq inferred and cfDNA detected CNA events. The red and blue dots indicated significance (P<0.01, Pearson) and non-significance (P≥0.01), respectively. (C) The pie chart shows the contribution of each malignant population to cfDNA-detected CNAs. cfDNA, cell-free DNA; CNA, copy number alteration; snRNA-seq, single-nuclei RNA sequencing; UMAP, Uniform Manifold Approximation and Projection.
Figure 6.Endothelial cells and CAFs constitute a pro-metastatic niche. (A, B) UMAP plot of immune (A) and endothelial (B) subtypes, clustered to 10 and 6 main groups, respectively. (C) Interactions between cell clusters inferred by the numbers of common ligand-receptor pairs using CellPhoneDB on snRNA-seq data from primary tumor tissues of non-recurrent (left) and recurrent (right) groups. (D) Bubble chart showing NOTCH signaling between malignant and endothelial subtypes, and between malignant cells and CAFs, in single-nuclei transcriptomes based on selected ligand and receptor pairs labelled on the left. The color and size of each dot indicate the level and the significance of the interaction, respectively. M3, malignant 3. (E) The spatial distribution of annotated spots in primary tumor tissue (left), P0159. The neighborhood score of each malignant spot with respect to the surrounding CAF spots (middle). The proximity score of each malignant spot with respect to the closest CAFs spot (middle). Comparison of CAFs Nbr and Prox scores between the recurrence and non-recurrence groups (right). (F) The spatial distribution of annotated spots in primary tumor tissue (left), P0068. The neighborhood score of each malignant spot with respect to the surrounding endothelial spots (middle). The proximity score of each malignant spot with respect to the surrounding endothelial spots (middle). Comparison of endothelial Nbr and Prox scores between the recurrence and non-recurrent groups (right). CAF, cancer-associated fibroblast; Nbr, neighborhood; Prox, proximity; snRNA-seq, single-nuclei RNA sequencing; UMAP, Uniform Manifold Approximation and Projection. ***P<0.001, unpaired Wilcoxon test.
Figure 7.VEGFA and NOTCH signaling promote migration of HCC cells. (A) Transwell migration assay of HuH-7 cells. a, negative control (NC). b, Treatment with the NOTCH pathway inhibitor DAPT. c, Stimulation with recombinant VEGFA protein. d, Co-treatment with VEGFA and DAPT. e, Co-culture with cancer-associated fibroblasts (CAFs). f, Co-culture with CAFs in the presence of DAPT. Scale bar, 400 μm. (B) Relative mRNA expression levels of NOTCH1, NOTCH2, and NOTCH3 in HuH-7 cells under monoculture and CAFs co-culture conditions, as determined by RT-qPCR (top). The quantitative results of the transwell assays (bottom). (C) Transwell migration assay of SNU-449 cells. a, Vehicle control. b, Treatment with the NOTCH pathway inhibitor DAPT. c, Stimulation with recombinant VEGFA protein. d, Co-treatment with VEGFA and DAPT. e, Co-culture with CAFs. f, Co-culture with CAFs in the presence of DAPT. Scale bar, 400 μm. (D) Relative mRNA expression levels of NOTCH1, NOTCH2, and NOTCH3 in SNU-449 cells under monoculture and CAFs co-culture conditions, as determined by RT-qPCR (top). The quantitative results of the transwell assays (bottom). 0.1% human serum albumin and/or 0.1% DMSO were used as vehicles in control groups. HCC, hepatocellular carcinoma; RT-qPCR, real-time quantitative polymerase chain reaction. *P<0.05; **P<0.01; ***P<0.001, P-values of unpaired t-test.
Figure 8.Nomogram for predicting pre-transplant recurrence-free survival probability. (A) The forest plot representing multivariate Cox regression results of key cfDNA-derived CNA fragments and clinical variables (left). Hazard ratios and their 95% confidence interval with P-values shown. When hazard ratio >1, the recurrent risk was higher in the group with the fragment CNA (or with higher clinical variables) than the group without the CNA (or with lower clinical variables). The converse was true when the hazard ratio was <1. Fragment 1, gain in chr. 7: 2000001–3000000; fragment 2, loss in chr. 16: 47000001–48000000, based on human genome assembly GRCh38. Characteristics in the nomogram predict probability of recurrence-free survival in patients with pre-transplant cfDNA CNAs and clinical data (right). To use the nomogram, locate the patient’s specific values on each variable axis. Draw vertical lines upward to determine points for each variable and sum up these points and locate the total on the total points axis. Draw a line downward to determine the probability of recurrence-free survival (1-year, 3-year, and 5-year). R, recurrent group; N, non-recurrent group. PIVKA-II: high, >40 mAU/mL; low, ≤40 mAU/mL. AFP: high, >60 ng/mL; low, ≤60 ng/mL. (B–E) Kaplan–Meier curves for overall survival or HCC recurrence based on nomogram scores in derivation cohort (B, left and middle), internal validation cohort (C, left and middle), external validation cohort I (D, left and middle) and external validation cohort II (E, left and middle). P-values of log-rank test. ROC curve and AUC of the nomogram, derivation cohort (B, right), internal validation cohort (C, right), external validation cohort I (D, right) and external validation cohort II (E, right). ROC, receiver operating characteristics; AUC, the area under ROC curve. (F) Calibration curve of the nomogram. Calibration curves of 1-year recurrence-free survival in internal validation cohort (left), external validation cohort I (middle) and external validation cohort II (right). The grey line represents the ideal reference where the predicted probabilities perfectly match the observed survival rates. Blue dots indicate the nomogram performance, calculated by bootstrapping (resampling: 1,000). The proximity of the solid red line (nomogram predictions) to this ideal line reflects model accuracy in predicting recurrence-free survival. cfDNA, cell-free DNA; CNA, copy number alteration; HCC, hepatocellular carcinoma. *P<0.05; **P<0.01; ***P<0.001, P-values of wald test.
Abbreviations
cancer-associated fibroblast
liver sinusoidal endothelial cell
liver vascular endothelial cell
single-nuclei RNA sequencing
Uniform Manifold Approximation and Projection
REFERENCES
- 1. Vogel A, Meyer T, Sapisochin G, Salem R, Saborowski A. Hepatocellular carcinoma. Lancet 2022;400:1345-1362.
- 2. Terrault NA, Francoz C, Berenguer M, Charlton M, Heimbach J. Liver transplantation 2023: status report, current and future challenges. Clin Gastroenterol Hepatol 2023;21:2150-2166.
- 3. Lucey MR, Furuya KN, Foley DP. Liver transplantation. N Engl J Med 2023;389:1888-1900.
- 4. Mazzaferro V, Regalia E, Doci R, Andreola S, Pulvirenti A, Bozzetti F, et al. Liver transplantation for the treatment of small hepatocellular carcinomas in patients with cirrhosis. N Engl J Med 1996;334:693-699.
- 5. Yao FY, Ferrell L, Bass NM, Watson JJ, Bacchetti P, Venook A, et al. Liver transplantation for hepatocellular carcinoma: expansion of the tumor size limits does not adversely impact survival. Hepatology 2001;33:1394-1403.
- 6. Xu X, Lu D, Ling Q, Wei X, Wu J, Zhou L, et al. Liver transplantation for hepatocellular carcinoma beyond the Milan criteria. Gut 2016;65:1035-1041.
- 7. Sapisochin G, Bruix J. Liver transplantation for hepatocellular carcinoma: outcomes and novel surgical approaches. Nat Rev Gastroenterol Hepatol 2017;14:203-217.
- 8. Yao FY. Expanded criteria for liver transplantation in patients with hepatocellular carcinoma. Hepatol Res 2007;37 Suppl 2:S267-274.
- 9. Sapisochin G, Goldaracena N, Laurence JM, Dib M, Barbas A, Ghanekar A, et al. The extended Toronto criteria for liver transplantation in patients with hepatocellular carcinoma: A prospective validation study. Hepatology 2016;64:2077-2088.
- 10. Mazzaferro V, Llovet JM, Miceli R, Bhoori S, Schiavo M, Mariani L, et al. Predicting survival after liver transplantation in patients with hepatocellular carcinoma beyond the Milan criteria: a retrospective, exploratory analysis. Lancet Oncol 2009;10:35-43.
- 11. Toso C, Asthana S, Bigam DL, Shapiro AM, Kneteman NM. Reassessing selection criteria prior to liver transplantation for hepatocellular carcinoma utilizing the Scientific Registry of Transplant Recipients database. Hepatology 2009;49:832-838.
- 12. Ohira M, Tanimine N, Kobayashi T, Ohdan H. Essential updates 2018/2019: liver transplantation. Ann Gastroenterol Surg 2020;4:195-207.
- 13. Chagas AL, Felga GEG, Diniz MA, Silva RF, Mattos AA, Silva R, et al. Hepatocellular carcinoma recurrence after liver transplantation in a Brazilian multicenter study: clinical profile and prognostic factors of survival. Eur J Gastroenterol Hepatol 2019;31:1148-1156.
- 14. Kaido T, Ogawa K, Mori A, Fujimoto Y, Ito T, Tomiyama K, et al. Usefulness of the Kyoto criteria as expanded selection criteria for liver transplantation for hepatocellular carcinoma. Surgery 2013;154:1053-1060.
- 15. Zhan QF, Ling SB, Deng YN, Shan QN, Ye QW, Xu SJ, et al. Hangzhou criteria as downstaging criteria in hepatocellular carcinoma before liver transplantation: a multicenter study from China. Hepatobiliary Pancreat Dis Int 2020;19:349-357.
- 16. Welch DR, Hurst DR. Defining the hallmarks of metastasis. Cancer Res 2019;79:3011-3027.
- 17. Priestley P, Baber J, Lolkema MP, Steeghs N, de Bruijn E, Shale C, et al. Pan-cancer whole-genome analyses of metastatic solid tumours. Nature 2019;575:210-216.
- 18. Ling S, Yu J, Zhan Q, Gao M, Liu P, Wu Y, et al. Multi-omic analysis reveals a CAF-stemness-governed classification in HCC liver transplant recipients beyond the Milan criteria. Nat Commun 2025;16:4392.
- 19. Elazezy M, Joosse SA. Techniques of using circulating tumor DNA as a liquid biopsy component in cancer management. Comput Struct Biotechnol J 2018;16:370-378.
- 20. Torrecilla S, Sia D, Harrington AN, Zhang Z, Cabellos L, Cornella H, et al. Trunk mutational events present minimal intraand inter-tumoral heterogeneity in hepatocellular carcinoma. J Hepatol 2017;67:1222-1231.
- 21. Roessler S, Lin G, Forgues M, Budhu A, Hoover S, Simpson RM, et al. Integrative genomic and transcriptomic characterization of matched primary and metastatic liver and colorectal carcinoma. Int J Biol Sci 2015;11:88-98.
- 22. Grendár M, Martínek P, Loderer D, Ondič O. CNHplus: the chromosomal copy number heterogeneity which respects biological constraints. bioRxiv 2022;2022.2009.2030.510279.
- 23. van Dijk E, van den Bosch T, Lenos KJ, El Makrini K, Nijman LE, van Essen HFB, et al. Chromosomal copy number heterogeneity predicts survival rates across cancers. Nat Commun 2021;12:3188.
- 24. Martínez-Jiménez F, Movasati A, Brunner SR, Nguyen L, Priestley P, Cuppen E, et al. Pan-cancer whole-genome comparison of primary and metastatic solid tumours. Nature 2023;618:333-341.
- 25. Ouyang L, Lee J, Park CK, Mao M, Shi Y, Gong Z, et al. Whole-genome sequencing of matched primary and metastatic hepatocellular carcinomas. BMC Med Genomics 2014;7:2.
- 26. Roth A, Khattra J, Yap D, Wan A, Laks E, Biele J, et al. Py-Clone: statistical inference of clonal population structure in cancer. Nat Methods 2014;11:396-398.
- 27. Hoshida Y, Nijman SM, Kobayashi M, Chan JA, Brunet JP, Chiang DY, et al. Integrative transcriptome analysis reveals common molecular subclasses of human hepatocellular carcinoma. Cancer Res 2009;69:7385-7392.
- 28. Patel AP, Tirosh I, Trombetta JJ, Shalek AK, Gillespie SM, Wakimoto H, et al. Single-cell RNA-seq highlights intratumoral heterogeneity in primary glioblastoma. Science 2014;344:1396-1401.
- 29. Ma L, Hernandez MO, Zhao Y, Mehta M, Tran B, Kelly M, et al. Tumor cell biodiversity drives microenvironmental reprogramming in liver cancer. Cancer Cell 2019;36:418-430.e416.
- 30. Ma L, Heinrich S, Wang L, Keggenhoff FL, Khatib S, Forgues M, et al. Multiregional single-cell dissection of tumor and immune cells reveals stable lock-and-key features in liver cancer. Nat Commun 2022;13:7533.
- 31. Hebbard LW, Garlatti M, Young LJ, Cardiff RD, Oshima RG, Ranscht B. T-cadherin supports angiogenesis and adiponectin association with the vasculature in a mouse mammary tumor model. Cancer Res 2008;68:1407-1416.
- 32. Garcia-Alonso L, Lorenzi V, Mazzeo CI, Alves-Lopes JP, Roberts K, Sancho-Serra C, et al. Single-cell roadmap of human gonadal development. Nature 2022;607:540-547.
- 33. Barkley D, Moncada R, Pour M, Liberman DA, Dryg I, Werba G, et al. Cancer cell states recur across tumor types and form specific interactions with the tumor microenvironment. Nat Genet 2022;54:1192-1201.
- 34. Sahai E, Astsaturov I, Cukierman E, DeNardo DG, Egeblad M, Evans RM, et al. A framework for advancing our understanding of cancer-associated fibroblasts. Nat Rev Cancer 2020;20:174-186.
- 35. Chhabra Y, Weeraratna AT. Fibroblasts in cancer: unity in heterogeneity. Cell 2023;186:1580-1609.
- 36. Cohen SA, Liu MC, Aleshin A. Practical recommendations for using ctDNA in clinical decision making. Nature 2023;619:259-268.
- 37. Mattox AK, Douville C, Wang Y, Popoli M, Ptak J, Silliman N, et al. The origin of highly elevated cell-free DNA in healthy individuals and patients with pancreatic, colorectal, lung, or ovarian cancer. Cancer Discov 2023;13:2166-2179.
- 38. Beroukhim R, Mermel CH, Porter D, Wei G, Raychaudhuri S, Donovan J, et al. The landscape of somatic copy-number alteration across human cancers. Nature 2010;463:899-905.
- 39. Zender L, Spector MS, Xue W, Flemming P, Cordon-Cardo C, Silke J, et al. Identification and validation of oncogenes in liver cancer using an integrative oncogenomic approach. Cell 2006;125:1253-1267.
- 40. Lin S, Yin YA, Jiang X, Sahni N, Yi S. Multi-OMICs and genome editing perspectives on liver cancer signaling networks. Biomed Res Int 2016;2016:6186281.
- 41. Bassaganyas L, Pinyol R, Esteban-Fabró R, Torrens L, Torrecilla S, Willoughby CE, et al. Copy-number alteration burden differentially impacts immune profiles and molecular features of hepatocellular carcinoma. Clin Cancer Res 2020;26:6350-6361.
- 42. Wong N, Lai P, Lee SW, Fan S, Pang E, Liew CT, et al. Assessment of genetic changes in hepatocellular carcinoma by comparative genomic hybridization analysis: relationship to disease stage, tumor size, and cirrhosis. Am J Pathol 1999;154:37-43.
- 43. Kusano N, Shiraishi K, Kubo K, Oga A, Okita K, Sasaki K. Genetic aberrations detected by comparative genomic hybridization in hepatocellular carcinomas: their relationship to clinicopathological features. Hepatology 1999;29:1858-1862.
- 44. Foda ZH, Annapragada AV, Boyapati K, Bruhm DC, Vulpescu NA, Medina JE, et al. Detecting liver cancer using cell-free DNA fragmentomes. Cancer Discov 2023;13:616-631.
- 45. Fu S, Debes JD, Boonstra A. DNA methylation markers in the detection of hepatocellular carcinoma. Eur J Cancer 2023;191:112960.
- 46. Berardi G, Morise Z, Sposito C, Igarashi K, Panetta V, Simonelli I, et al. Development of a nomogram to predict outcome after liver resection for hepatocellular carcinoma in Child-Pugh B cirrhosis. J Hepatol 2020;72:75-84.
- 47. Liu Y, Wei X, Zhang X, Pang C, Xia M, Du Y. CT radiomics combined with clinical variables for predicting the overall survival of hepatocellular carcinoma patients after hepatectomy. Transl Oncol 2022;26:101536.
- 48. Liu Q, Li J, Liu F, Yang W, Ding J, Chen W, et al. A radiomics nomogram for the prediction of overall survival in patients with hepatocellular carcinoma after hepatectomy. Cancer Imaging 2020;20:82.
- 49. Ji GW, Zhu FP, Xu Q, Wang K, Wu MY, Tang WW, et al. Radiomic Features at contrast-enhanced CT predict recurrence in early stage hepatocellular carcinoma: a multi-institutional study. Radiology 2020;294:568-579.
- 50. Zhu Y, Mao Y, Chen J, Qiu Y, Guan Y, Wang Z, et al. Radiomics-based model for predicting early recurrence of intrahepatic mass-forming cholangiocarcinoma after curative tumor resection. Sci Rep 2021;11:18347.
, Jianhua Li4,*, Yan Wang1,*, Jiapei Wang5,*, Qiwei Huang6, Fanming Liu1, Yiqi Zhuang1, Yingyu Zheng1, Rui Wang7, Zhe Yang8, Xiaoping Zheng9, Kai Wang2, Zhikun Liu2, Jun Chen3, Jianguo Wang2, Haiyang Xie5, Lin Zhou5, Leiming Chen5, Guoqiang Cao5, Dandan Chen5, Junfang Ji1, Bin Zhao1, Chao Jiang1, Di Lu2, Xuyong Wei10, Hangjin Jiang11, Qiaonan Shan5, Hengbo Shi12, Yong-Zhen Xu6, Shusen Zheng5,8
