From January 2013 to December 2016, a total of 3,417 consecutive patients diagnosed with histologically proven primary HCC who received tumor resection in the Department of Liver Surgery and Transplantation in Zhongshan Hospital, Fudan University, were recruited. 46 patients with sarcomatoid HCC were identified from the pathology database, and their pathological information was retrieved for reconfirmation. Five patients were excluded from the study because the sarcomatoid tumor components and conventional HCC components were not in the same paraffin blocks. For the remaining 41 patients, 1 case was excluded due to palliative surgery, and 9 cases were eliminated because of prior interventions. Finally, 31 cases with both sarcomatoid tumor components and conventional HCC components were identified and included in this study (Supplementary Fig. 1, Supplementary Table 1), and 10 cases were involved in integrated multiomics study. The pathological diagnosis of sarcomatoid HCC was confirmed through histological examination (using hematoxylin and eosin [H&E] staining) and immunohistochemical staining (including Hep, Par-1, arginase-1, cytokeratin [CK], and Vimentin). Hepatocellular markers were typically negative in the malignant spindle-cell component, whereas CK and Vimentin were positive in most cases [11]. The histological tumor differentiation was determined according to the system proposed by Edmondson and Steiner [15]. Liver function was assessed by the Child–Pugh scoring system. Tumor stage was determined using the 2017 International Union Against Cancer tumor-node-metastasis (TNM) classification system [16]. Before surgical operation and tissue sample collection, we obtained oral and written informed consent from each participant, which included information on the use of clinical characteristics and tissue samples for scientific research. For this research, the Research Ethics Committee of Zhongshan Hospital reviewed and approved the study and granted ethical approval for the use of human subjects. We also obtained oral informed consent for inclusion in the study from the participants at the time of follow-up. The Institutional Review Board (IRB) number is IRB00008978.

All patients with sarcomatoid HCC were monitored after surgery until December 30, 2019 as previously described [17,18]. We diagnosed tumor recurrence on the basis of computed tomography scans, magnetic resonance imaging, digital subtraction angiography, and elevated serum alpha-fetoprotein (AFP) level, with or without histological confirmation [10]. We defined recurrence-free survival (RFS) as the interval between the surgery and any diagnosis of recurrence (intrahepatic or extrahepatic) [19]. We defined overall survival (OS) as the time from the date of surgery until death or the end of follow-up. The surviving patients were censored at the time of the end of follow-up.

Additional materials and methods can be found in the Supplementary Materials and Methods.

We calculated propensity scores according to sex, age, and TNM stage to match the patients with sarcomatoid HCC at a 1:4 ratio to patients with non-sarcomatoid HCC as a control group from our previous next-generation sequencing study of HCC (n=182) [20]. As a result, a total of 124 matched patients with non-sarcomatoid HCC from the prior study were included as controls (Supplementary Table 1). The median follow-up was 40.7 months and 73.7 months for the sarcomatoid and non-sarcomatoid groups, respectively. The 1-year, 3-year, and 5-year OS rates of the patients with non-sarcomatoid HCC were significantly higher than those of the patients with sarcomatoid HCC (87.1% versus 61.3%, 62.1% versus 28.0%, and 43.9% versus 16.8%, respectively; Supplementary Fig. 2A). Similarly, the patients with sarcomatoid HCC had poorer prognosis at 1, 3, and 5 years, with lower RFS rates than the patients with non-sarcomatoid HCC (40.0% versus 79.7%, 15.9% versus 34.3%, and 0.0% versus 26.0%, respectively; Supplementary Fig. 2B). Univariate and multivariate analyses revealed that the tumor subtype (sarcomatoid HCC vs. non-sarcomatoid HCC) was an independent prognostic factor for both OS (P=0.002; hazard ratio [HR]=2.27) and RFS (P=0.002; HR=2.24; Supplementary Fig. 2).

We performed WES of 28 paired sarcomatoid tumor components and conventional HCC components and 10 matched normal liver samples from 10 patients with sarcomatoid HCC. The average sequencing depth was 360.3-fold for the sarcomatoid tumor components, 364.3-fold for the conventional HCC components, and 133.5-fold for normal liver samples (Supplementary Table 2). We mapped the sequence reads to the human reference genome and identified a total of 3,696 somatic non-synonymous mutations in the sarcomatoid tumor components (42–1,160 per tumor) and 2,302 somatic non-synonymous mutations in the conventional HCC components (39–1,222 per tumor; Fig. 1A, Supplementary Table 3). The mean number of somatic non-synonymous mutations per patient in the sarcomatoid tumor components was 220.1±333.9, which was comparable to that in the conventional HCC components (215.7±361.3). Sanger sequencing to validate 320 randomly selected non-synonymous mutations showed a high true-discovery rate (91.6%).

We compared the most frequently mutated cancer-related genes identified in the sarcomatoid HCC cohort (n=10) to those identified in other non-sarcomatoid HCC cohorts, including our previous cohort [10], the Gao cohort [21], and a TCGA cohort [4]. We found that ARID2 mutation occurred in 70% (7/10) of the patients with sarcomatoid HCC but only 1–5% of the patients with non-sarcomatoid HCC in the other three cohorts (P<0.001 for each comparison; Fig. 1B}.

We identified an average of 878.4 MB gain and 534.3 MB loss in the sarcomatoid tumor components, which was comparable to the total length of copy-number variations in the conventional HCC components (668.9 MB gain and 415.2 MB loss; Fig. 1C). GISTIC analysis identified 5 recurrently amplified segments and 12 recurrently lost segments in the sarcomatoid tumor components, and 4 recurrently amplified segments and 11 recurrently lost segments in the conventional HCC components (Fig. 1D, Supplementary Table 4). Furthermore, 8p23.2 was recurrently deleted in the sarcomatoid tumor components but not in the conventional HCC components. qRT-PCR results demonstrated that the expression of EGR3, GATA4, SOX7, and LZTS1, which are common tumor-suppressor genes located on 8p23.2 [22-25], was downregulated in the sarcomatoid tumor components compared with the conventional HCC components (Supplementary Fig. 3A). Immunohistochemical staining results also confirmed that EGR3 and SOX7 expression was lower in the sarcomatoid tumor components compared with the conventional HCC components (Supplementary Fig. 3B). These results suggested a possible role of 8p23.2 deletion in sarcomatoid HCC development and progression.

To infer the evolutionary trajectory of sarcomatoid HCCs, we analyzed the subclonal architectures using calculated mutation clusters based on all identified somatic mutations (SNVs+indels). Figure 2 shows the results rendered using MapScap [26], with different colored ovals indicating clones and subclones and diameters representing the corresponding cancer cell fractions (CCFs). There were 4–8 clonal clusters estimated per patient (Fig. 2), suggesting that multiple clones were present in the samples. The phylogenetic trees showed that all of the sarcomatoid tumor components shared a proportion of mutations with their matched conventional HCC components, confirming the existence of common ancestors from which both tumor components were derived in each patient. We also observed branched tumor evolution during sarcomatoid HCC development and progression; however, we did not observe any difference in the number of clones between the sarcomatoid tumor components and the conventional HCC components in each patient.

We next analyzed the distribution of mutations affecting known HCC driver genes across the clonal phylogenies to determine the timing of acquisition and clone specificity of the driver mutations. TP53 mutations occurred in seven of the 10 patients and were always exhibited in a trunk clone (black clone, 6/7, except for P13), supporting a relatively ancestral role for TP53 mutation. ARID2 mutations appeared in trunk clones in only two patients, whereas they appeared in trunk or branch subclones in four patients. These results confirmed that TP53 mutations commonly occurred at tumor initiation during sarcomatoid HCC development, whereas ARID2 mutations often occurred later as a prerequisite for rapid growth and progression.

We further analyzed the mutational spectrum according to the genetic phylogenies of the sarcomatoid HCCs (i.e., trunk mutations, conventional HCC component mutations, sarcomatoid tumor component mutations). T>A transversion was enriched in trunk mutations, whereas C>T transition was enriched in mutations occurring in either the conventional HCC component or the sarcomatoid tumor component (Supplementary Fig. 4A).

We next analyzed mutational signatures in all 28 paired sarcomatoid tumor components and conventional HCC components using DeconstructSigs, which accurately reconstructs mutational profiles based on a predefined mutational spectrum of 30 COSMIC signatures. One prominent feature revealed by the mutational signature analysis was that the mutational signatures were similar in the sarcomatoid tumor components and the conventional HCC components (Supplementary Fig. 4B), indicating that both types follow similar mutational processes. The other significant feature was that the heterogeneity of mutational signatures across patients was considerably greater than the heterogeneity across different evolutionary stages within any given patient (Supplementary Fig. 4B). That finding suggests that a given sarcomatoid HCC accesses only a subset of the mutational processes that are potentially available to it, but these mutational processes contribute genomic variation on an ongoing basis. Nonetheless, there were some shifts in the relative contributions of the mutational processes over time. For example, signature 22, which corresponds to T>A transversions and is correlated with aristolochic acid exposure [10], was strongly enriched among trunk mutations, especially in patients P13, P14, and P15. Signature 22 is usually associated with a high mutational burden [10], which might account for the relatively high mutational burden in those three patients. Similarly, alcohol-related signature 16 was also enriched among trunk mutations, suggesting that alcohol-related etiology mainly contributed to sarcomatoid HCC formation but not progression. By contrast, signature 1 and signature 3, which are associated with age and homologous recombination deficiency, respectively, were enriched in conventional HCC components and further enriched in sarcomatoid tumor components, suggesting that those mutational processes contributed more to the stage of sarcomatoid HCC progression.

To comprehensively understand the transcriptional characteristics of sarcomatoid HCCs, we selected one tissue specimen (P03) from among 10 patients with sarcomatoid HCC and applied ST sequencing using the 10×Genomics Visium platform. After clustering, four spatially sequential sections were obtained (Fig. 3A, Supplementary Fig. 5A). Based on marker gene expression and histological characteristics, we identified the four sections: peritumor liver cluster, conventional HCC cluster, sarcomatoid tumor cluster, and margin cluster (Fig. 3A, Supplementary Fig. 5A, 5B). Moreover, we compared the H&E staining images with the ST data regarding the marker gene expression and verified that the transcriptional features were consistent with the histological information (Fig. 3B). Gene Ontology (GO) results showed that different clusters diversified pathway activities within certain tissue components. The activities of cell cycle and DNA repair pathways, which were reported to promote HCC occurrence and progression in previous studies, were higher in conventional HCC components than in peritumor liver components [9,10]. Furthermore, some other pathways, such as cell junction, cell adhesion, and cell migration, were enriched in sarcomatoid tumor components. Importantly, we found that EMT and hypoxic signaling exhibited high enrichment in both conventional HCC and sarcomatoid tumor components, and this enrichment was highest in the sarcomatoid tumor components (Fig. 3C).

Next, we collected 35 samples (including 12 conventional HCC components, 14 sarcomatoid tumor components, and 9 matched peritumor liver tissues) from the remaining 9 patients with sarcomatoid HCC and performed bulk RNA-seq to validate and supplement the transcriptional results obtained by ST sequencing. The results showed that 103 genes were upregulated both in conventional HCC components (compared with peritumor liver samples) and in sarcomatoid tumor components (compared with conventional HCC components), while 76 genes were similarly downregulated (Fig. 3D, Supplementary Fig. 5C). GO analysis also demonstrated that the activities of the cell cycle, DNA repair, response to hypoxia, and regulation of EMT pathways were enriched in conventional HCC components compared with peritumor liver tissues. Furthermore, when compared with conventional HCC components, sarcomatoid tumor components displayed further enrichment of cell migration, cell adhesion, regulation of EMT, and response to hypoxia (Fig. 3E). Gene Set Enrichment Analysis (GSEA) also indicated that the gene sets of EMT and hypoxic signaling were enriched in conventional HCC components, compared with peritumor liver tissues, and further enriched in sarcomatoid tumor components (Fig. 3F). In addition, we observed that compared with peritumor liver tissues, conventional HCC components had downregulated expression of E-cadherin, and upregulated expression of Vimentin, indicating an EMT phenotype, and these changes in EMT marker expression were even more pronounced in the sarcomatoid tumor components (Fig. 3G).

To verify the EMT phenotype of sarcomatoid HCCs revealed by transcriptome analyses, we investigated the expression of the EMT markers E-cadherin, Vimentin, and N-cadherin by immunohistochemical staining in 31 sarcomatoid HCCs with both sarcomatoid tumor components and conventional HCC components. Representative cases of immunohistochemical staining are shown in Figure 4. Compared with peritumor liver tissues, both the conventional HCC components and the sarcomatoid tumor components exhibited a typical EMT phenotype characterized by downregulated E-cadherin and upregulated Vimentin and N-cadherin; moreover, these changes in EMT marker expression were more pronounced in the sarcomatoid tumor components than in the conventional HCC components (Fig. 4). In addition, immunohistochemistry results confirmed a hypoxic microenvironment in these sarcomatoid tumor components, which exhibited dense staining of HIF-1α and CAIX (Fig. 4).

A total of 25 ARID2 somatic mutations (10 different mutations) were identified in 7 conventional HCC components and 10 sarcomatoid tumor components from 7 patients (Fig. 5A). We used Sanger sequencing to validate these results, except for c.1201-1202del in ST16 because of insufficient DNA (Supplementary Fig. 6). The 10 different ARID2 somatic mutations included 4 frame-shift indels, 2 in-frame indels, 3 nonsense SNVs, and 1 missense SNV. Polyphen-2 analysis revealed that the missense SNV was predicted to adversely affect the function of the ARID2 protein.

We next evaluated ARID2 expression by immunohistochemistry in 10 sarcomatoid HCCs and 124 non-sarcomatoid HCCs from matched patients (Fig. 5B). The results showed that ARID2 expression was downregulated in the tumors overall compared with that in adjacent peritumor liver samples (Fig. 5C). Furthermore, ARID2 expression was more downregulated in sarcomatoid HCCs than in non-sarcomatoid HCCs, even those carrying ARID2 somatic mutations (Fig. 5D, Supplementary Fig. 7).

We correlated ARID2 expression with clinical characteristics and outcomes in 124 patients with non-sarcomatoid HCCs. Low ARID2 expression was correlated with increased tumor size and vascular invasion and poor tumor differentiation (Supplementary Table 5). Kaplan–Meier survival analysis revealed significantly shorter OS and RFS among patients with low ARID2 expression (Fig. 5E). Univariate and multivariate analyses confirmed that low ARID2 expression was an independent prognostic factor for OS and RFS (Fig. 5E).

Our results suggested a possible tumor-suppressor role for ARID2 in HCC. To test that hypothesis, we analyzed the genotype and expression of ARID2 in six HCC cell lines. All six HCC cell lines were validated as wild-type (WT) by Sanger sequencing, using the same filter criteria used for WES. Western blots confirmed that ARID2 protein levels in the high metastatic potential HCC cells (MHCC97L, MHCC97H, and HCCLM3) were lower than that in the low metastatic potential HCC cells (HepG2, PLC/PRF/5, and Hep3B) (Fig. 6A) [2,17].

Next, we knocked down ARID2 in HepG2 cells and conducted biofunctional investigations to determine the phenotypic effects. We found that the knockdown of ARID2 resulted in increased HCC cell proliferation and colony formation activity and led to enhanced invasive ability (Fig. 6B–6D). In vivo HCC mouse models showed that ARID2 knockdown accelerated tumor growth and pulmonary metastasis (Fig. 6E).

We next generated lentiviral constructs to re-express WT or mutant (MT) ARID2 in HCCLM3 cells. We found that overexpression of WT ARID2 substantially suppressed HCC cell proliferation, colony formation, and invasion ability. By contrast, all MT ARID2 variants failed to cause these effects in whole or in part (Fig. 6B–6D). In agreement with the in vitro studies, analysis of an in vivo HCC mouse model showed that WT ARID2 significantly suppressed tumor growth and pulmonary metastasis, whereas MT ARID2 yielded a larger tumor volume and increased pulmonary metastasis compared with WT ARID2 (Fig. 6E). These results support the notion that ARID2 is a tumor-suppressor gene in HCC and that certain somatic mutations abolish its function and its tumor-inhibitory effect.

Considering the high frequency of ARID2 mutations in sarcomatoid HCCs, which exhibited an EMT phenotype and hypoxic microenvironment, especially in sarcomatoid tumor components, we speculated that HCC cells carrying ARID2 mutation may undergo EMT under hypoxic conditions, contributing to sarcomatoid HCC formation and development. We observed that the HepG2 cells with ARID2 knockdown took on a spindle-like fibroblastic morphology, whereas the HCCLM3 cells overexpressing WT ARID2 presented a cobblestone-like appearance of epithelium (Supplementary Fig. 8). Next, we screened for epithelial and mesenchymal markers in HCC cells. The results showed that knockdown of ARID2 in HepG2 cells resulted in downregulation of E-cadherin and upregulation of Vimentin and N-cadherin (Fig. 7C, 7D). In HCCLM3 cells, reexpression of WT ARID2 had an inverse effect, including upregulation of E-cadherin and downregulation of Vimentin and N-cadherin, whereas expression of MT ARID2 failed to cause these effects in whole or in part (Fig. 7A, 7B). Furthermore, these HCC cells displayed consistent snail expression with phenotypes typical of cells undergoing EMT, but not consistent expression of other transcriptional factors such as slug, twist, E47, ZEB1, or ZEB2 (data not shown). We validated the in vitro results with immunohistochemical staining of tumor sections from a mouse model (Fig. 7G). These results suggested that ARID2 inactivation leads to EMT of HCC cells. More importantly, we found that in HCCLM3 and HepG2-shARID2 cells—two cell lines with ARID2 deficiency—hypoxia further induced phenotypes typical of cells undergoing EMT, including further downregulation of E-cadherin and further upregulation of Vimentin and N-cadherin (Fig. 7C–7F). These results demonstrated that hypoxia facilitated EMT of HCC cells carrying inactivated ARID2, which may contribute to sarcomatoid HCC formation and development.