Abstract Alpha-fetoprotein (AFP) is a classic biomarker for hepatocellular carcinoma (HCC). AFP-positive HCC (AFP^+ HCC) has been intensively investigated; however, the genomic, transcriptomic and microenvironmental characteristics of AFP-negative HCC (AFP^− HCC) remain to be deciphered. Here we show that tumors display mild differences in genetic alterations between AFP^− HCC and AFP^+ HCC patients, while AFP^− HCC exhibits hyperactive arachidonic acid metabolism. Furthermore, the transcription activity of androgen receptor (AR) is significantly increased in AFP^− HCC and plays a positive regulatory role in arachidonic acid metabolism and its metabolite 11,12-epoxyeicosatrienoic acid (11,12-EET). The tumor-derived 11,12-EET exhibits high affinity for EGFR that promotes the migration and tube formation of endothelial cells in vitro. Combination of lenvatinib and bicalutamide (an AR antagonist) enhances the therapeutic efficacy for AFP^− HCC. Overall, we uncover the AR-mediated hyperactive arachidonic acid metabolism in AFP^− HCC, and reveal AR-11,12-EET-EGFR axis-induced angiogenesis, providing a promising strategy of combined AR antagonist with lenvatinib for AFP^− HCC treatment. Subject terms: Tumour angiogenesis, Cancer metabolism, Cellular signalling networks, Tumour biomarkers __________________________________________________________________ Alpha-fetoprotein is a known biomarker for hepatocellular carcinoma. Here, the authors analyse differences between AFP positive and negative patients to identify alterations in androgen receptor signalling. Introduction Liver cancer is the third leading cause of cancer-related deaths, with an overall low 5-year survival rate of 20.8%^[66]1,[67]2. Hepatocellular carcinoma (HCC) accounts for 80–90% of liver cancer worldwide^[68]3. Alpha-fetoprotein (AFP) is the most widely used biomarker for HCC in clinical practice, primarily aiding in HCC screening, diagnosis, prognosis prediction, and treatment efficacy monitoring^[69]4. However, our knowledge on genetic features, molecular characteristics and tumor ecosystem profiling of AFP-related HCCs is very limited. Extensive evidence has established that AFP exerts dual oncogenic effects by enhancing tumor cell proliferation and suppressing apoptosis in HCC^[70]5. In addition, AFP also plays pivotal roles in regulating immune responses, for instance, it can induce natural killer (NK) cell apoptosis, promote Treg differentiation and inhibit dendritic cell (DC) maturation, ultimately repressing CD8^+ T cells and resulting in an immunosuppressive microenvironment^[71]6. Recently, comprehensive multi-omics approaches integrating whole-exome sequencing (WES), bulk RNA-sequencing (RNA-seq) and single-cell RNA-sequencing (scRNA-seq) have been adopted to reveal the multi-dimensional characteristics of tumor immune ecosystem in AFP-positive HCC (AFP^+ HCC, serum AFP level ≥ 20 ng/mL at diagnosis). The findings showed that tumor-associated macrophages (TAMs) suppressed CD8^+ T cell function through the SPP1-CD44 axis in the tumor microenvironment (TME) of AFP^+ HCC, thus providing a promising target for achieving a more favorable efficacy in AFP^+ HCC treatment^[72]7. Therefore, AFP^+ HCC has been widely explored in previous studies^[73]6–[74]10. However, approximately 30–40% of HCC showed negative AFP serum level (AFP^− HCC, serum AFP level < 20 ng/mL at diagnosis)^[75]11,[76]12. Of note, several recent global clinical trials have indicated that AFP^− HCC patients showed moderate treatment response to the first-line drug lenvatinib or its analogues^[77]13–[78]15. Nevertheless, due to the few studies on AFP^− HCC, the underlying mechanism remains largely unknown without available interventions. Therefore, we aimed to comprehensively decipher the genomic, transcriptomic and microenvironmental characteristics of AFP^− HCC that helped guide the precise treatment strategy for this subtype of HCC. In the present study, we performed an integrated omics analyses containing WES on 314 samples, scRNA-seq on 25 samples, bulk RNA-seq on 434 samples, single-nucleus ATAC-sequencing (snATAC-seq) on 4 samples, methylation profiling by Reduced Representation Bisulfite Sequencing (RRBS) on 26 samples and lipidomics on 34 samples to systematically compare the genetic, transcriptional and microenvironmental differences between AFP^+ HCC and AFP^- HCC. While no significant genetic differences were observed between the two subtypes, AFP^− HCC exhibited pronounced activation of arachidonic acid metabolism mediated by the androgen receptor (AR), which induced substantial secretion of the downstream metabolite 11,12-epoxyeicosatrienoic acid (11,12-EET) that directly activated epidermal growth factor receptor (EGFR) on endothelial cells (ECs), thereby leading to aberrant angiogenesis. Thus, we have revealed an angiogenesis regulatory pathway controlled by AR-11,12-EET-EGFR axis in AFP^− HCC, which cannot be effectively inhibited by lenvatinib. This finding provided the scientific rationale for combination treatment of lenvatinib and inhibition of AR-mediated angiogenesis for AFP^− HCC. Results Profiling of cell diversity in the AFP^− HCC and AFP^+ HCC To comprehensively elucidate the distinct features between AFP^− HCC and AFP^+ HCC, we conducted integrated omics analyses including whole-exome sequencing on 314 samples (157 tumor and 157 non-tumor adjacent tissues) from 157 patients, scRNA-seq on 25 samples (17 tumor and 8 non-tumor adjacent tissues) from 17 patients, bulk RNA-seq on 434 samples (434 tumor tissues) from 434 patients, snATAC-seq on 4 samples (4 tumor tissues) from 4 patients, methylation sequencing on 26 samples (26 tumor tissues) from 26 patients and lipidomics on 34 samples (14 tumor tissues and 20 blood plasma specimens) from 34 patients (Fig. [79]1a and Supplementary Fig. [80]1a). Clinical information of these patients was provided in Supplementary Data [81]1, with no significant differences in HBsAg, antiviral treatment, liver cirrhosis status, tumor number; modestly significant bias in gender, age and tumor size (p < 0.05); and significantly lower microvascular invasion (MVI), Barcelona Clinic Liver Cancer (BCLC) stages and Edmondson-Steiner (E-S) grades in AFP^− HCC versus AFP^+ HCC (p < 0.001). Fig. 1. Profiling of genetic alteration and cell diversity in HCC. [82]Fig. 1 [83]Open in a new tab a Schematic overview of the research. The numbers of patients and multiple experimental validation methods are provided. The lipidomics datasets (n = 34) comprise two subsets: tumor tissue lipidomics (n = 14) and blood plasma lipidomics (n = 20). b Genetic profiles and associated clinical features of all 157 HCC patients with WES data. The top panel shows mutation numbers followed by clinicopathological features. The middle panel exhibits significantly mutated genes calculated through MutSigCV. The bottom panel shows the CNV profiles of top chromosomal lesions with the most significant q values. Different alteration types and clinical features are signified with different color codes. Significance of these variables between AFP^− HCC and AFP^+ HCC was calculated by chi-square test or Fisher’s exact test, which was provided in Supplementary Data [84]1. c The UMAP visualization of 47 cell types from scRNA-seq data. d Scatter plot showing positive correlation between serum AFP level and transcriptomic AFP expression in scRNA-seq (left panel) and bulk RNA-seq (right panel) data of FAH-SYSU cohort. Transcriptomic AFP expression in scRNA-seq data was calculated as the average expression across tumor cells per sample. R indicates the correlation coefficient calculated by Spearman correlation test. The number of dots indicates the number of patients. e Representative images of IHC staining in FFPE tissues (left panel) and protein expression of AFP in AFP^− HCC and AFP^+ HCC (right panel). The number of dots indicates the number of patients. Scale bars, 50 μm. Data were shown as mean ± SD. **p < 0.01 by two-sided t-test. f Scatter plot showing positive correlation between proteinic AFP expression and available transcriptomic AFP expression. R indicates the correlation coefficient calculated by Spearman correlation test. The number of dots indicates the number of patients. Source data are provided as a Source Data file. To decipher the genomic profiles of AFP^− HCC, we performed WES on 157 tumors, including 44 AFP^− HCC and 113 AFP^+ HCC, and their paired non-tumor adjacent tissues (NATs). Among them, 18,441 non-synonymous somatic mutations (range 0–636, average 117.46 per sample) and 2776 indels (range 0–420, average 17.68 per sample) across all samples were characterized. Consistent with previous findings^[85]16–[86]18, nine recurrently mutated genes, including TP53, CTNNB1, AXIN1, ARID1A, and TSC2 were identified (Fig. [87]1b), which showed no significance between AFP^– HCC and AFP^+ HCC (Supplementary Data [88]2). By identifying copy number variation (CNV) events and calculating GISTIC scores, we observed mild differences in amplification and deletion events between AFP^− HCC and AFP^+ HCC, with significant events occupying 4% and 3%, respectively (Supplementary Fig. [89]1b, c). The frequently amplificated and deleted events were listed in Supplementary Data [90]3. In addition, no significant differences of immune editing score, clonal expansion of recurrently mutated genes and COSMIC mutation signatures were observed between the two subtypes (Supplementary Fig. [91]1d–g). Therefore, these results indicated that AFP^– HCC and AFP^+ HCC patients displayed similar genomic alterations. Next, we aimed to decipher the transcriptomic differences between the two subtypes by performing scRNA-seq on 6 AFP^− HCC and 11 AFP^+ HCC tumor samples, with 8 paired NATs. The detailed clinical information of these patients was provided in Supplementary Data [92]4. The baseline of patients with scRNA-seq data was matched between AFP status groups (AFP^− HCC vs. AFP^+ HCC). Notably, we also analyzed the paired WES data from samples with matched scRNA-seq profiles (excluding one case due to tissue unavailability). No significant differences were observed between the two HCC subtypes in terms of gene mutations, CNVs, immune editing scores, clonal expansion of recurrently mutated genes, or COSMIC mutation signatures (Supplementary Fig. [93]2a–e and Supplementary Data [94]5). In scRNA-seq data, total 210,545 single cells were collected and were classified into 47 cell types based on gene expression panorama (Fig. [95]1c, Supplementary Fig. [96]3a–d, and Supplementary Data [97]6). Subsequently, CNVs were inferred to validate the definition of tumor cells through inferCNV (Supplementary Fig. [98]3e). We then further determined that the transcriptome level of AFP in tumors was positively correlated with its serum levels in both FAH-SYSU and TCGA-LIHC cohorts (Fig. [99]1d and Supplementary Fig. [100]3f), and exhibited high expression in AFP^+ HCC tumor samples, with no detectable expression in tissues obtained from healthy liver, NATs of AFP^+ HCC, AFP^– HCC tumor and NATs of AFP^– HCC (Supplementary Fig. [101]3g). Similar results were observed in staining of FFPE specimen (Fig. [102]1e, f), suggesting that transcriptomic expression of AFP could represent the serum AFP level and define the two HCC subtypes. AFP^– HCC tumors presented hyperactive arachidonic acid metabolism Next, via differential gene expression analysis based on scRNA-seq data, significantly altered genes (550 upregulated and 876 downregulated genes) in tumor cells from AFP^− HCC were identified compared to AFP^+ HCC. The metabolism-related genes, such as CYP2A6, FGGY, and CYP2C8, were significantly upregulated in AFP^− HCC, while proliferation-related genes such as IGF2, EPCAM, and ARID3A, were significantly upregulated in AFP^+ HCC (Fig. [103]2a). Consistently, Gene Set Enrichment Analysis (GSEA) suggested that the AFP^– HCC tumors enriched metabolism and cell adhesion-related pathways, such as xenobiotic metabolic process, constitutive androstane receptor pathway, arachidonic acid metabolism and extracellular matrix (ECM) receptor interaction, a critical molecular feature of early-stage HCC^[104]19, whereas the AFP^+ HCC tumors upregulated proliferation and carcinogenesis-associated pathways, such as Myc targets, regulation of cell division and non-canonical WNT signaling pathways (Fig. [105]2b). These results were also supported by previous reports^[106]8. To gain additional insights into the molecular differences of the two HCC subtypes, we defined the 434 HCC patients by Hoshida’s classification (Hoshida’s S1, S2, and S3) based on bulk RNA-seq data using the Nearest Template Prediction (NTP) algorithm^[107]20,[108]21. S1 reflected abnormal activation of the WNT and TGF-β signaling pathways; S2 was characterized by proliferation features and activation of MYC and AKT signaling, and exhibited the highest AFP expression among the three subtypes; S3 showed well differentiation of tumor cells, with the lowest AFP expression among the three Hoshida’s subtypes. The results showed that AFP^− HCC presented similar features to the S3 subtype, whereas AFP^+ HCC exhibited Hoshida’s S1/2 subtypes characteristics (Supplementary Fig. [109]4a). Moreover, the differentially expressed genes (DEGs) and GSEA results from the bulk RNA-seq data also verified the hyperactive metabolic features in AFP^− HCC and the proliferative properties in AFP^+ HCC (Supplementary Fig. [110]4b, c). Fig. 2. Transcriptomic differences of tumor cells between AFP^– HCC and AFP^+ HCC. [111]Fig. 2 [112]Open in a new tab a Differentially expressed genes (DEGs) calculated from scRNA-seq data between AFP^− HCC and AFP^+ HCC. b Pathways enriched in tumor cells in AFP^– HCC and AFP^+ HCC from scRNA-seq data by GSEA. The p-values, calculated by permutation test, of all listed pathways were less than 0.05. c The volcano plot from scRNA-seq data manifested differentially metabolic activity of tumor cells in AFP^– HCC when compared to AFP^+ HCC. The color and dot size correspond to the log[2]FC in metabolic activity between AFP^− HCC and AFP^+ HCC. d Correlation analysis between metabolic pathways and transcriptomic AFP expression in scRNA-seq data by Spearman correlation test. e Metabolic pathway diagram of arachidonic acid metabolism and gene signature scores of three branches in arachidonic acid metabolism. Box plot depicts the median and interquartile range, and the lower and upper hinges denote the 25–75% interquartile range (IQR), with whiskers extending up to a maximum of 1.5 times IQR. f Scatter plot showing negative correlation between transcriptomic AFP expression and CYP gene signature in scRNA-seq data by Spearman correlation test. g The heatmap of genes involved in arachidonic acid metabolism in AFP^– HCC and AFP^+ HCC, colored by average expression value. h Relative expression of CYP2C8/CYP1B1 detected by RT-qPCR in tumor tissues from 15 AFP^− HCC and 15 AFP^+ HCC patients. Data were shown as mean ± SD. i The volcano plot shows metabolites detected by targeted lipidomics. The abbreviations ARA and DTA stand for arachidonic acid and docosatetraenoic acid, respectively. j The expression levels of AFP were detected by RT-qPCR and western blot in HCC cell lines. k ELISA analysis revealed elevated levels of 11,12-EET in the CM derived from AFP^low cell lines compared to AFP^high cell lines. Data were shown as mean ± SD. *p < 0.05, **p < 0.01, ***p < 0.001 by Wilcoxon rank sum test in (e, g, i), two-sided t-test in (h) and one-way ANOVA with Tukey’s multiple comparisons test (j, k). Source data are provided as a Source Data file. To further delineate the metabolic features in AFP^− HCC, we employed the scMetabolism algorithm^[113]22 to tumor cells from scRNA-seq data. This analysis revealed a significant upregulation of the arachidonic acid metabolism pathway in AFP^– HCC (Fig. [114]2c and Supplementary Data [115]7), which exhibited a strong negative correlation with AFP expression (Fig. [116]2d). These findings were further confirmed by the GSEA results from scRNA-seq and bulk RNA-seq data, respectively (Fig. [117]2b and Supplementary Fig. [118]4c). Compared to all other cell types, tumor cells displayed the highest scores for arachidonic acid metabolism and the strongest difference between AFP^− HCC and AFP^+ HCC (Supplementary Fig. [119]4d). Based on the well-established main branches of arachidonic acid metabolism, comprising the cyclooxygenase (COX), lipoxygenase (LOX), and cytochrome P450 (CYP) pathways^[120]23, we conducted differential expression profiling and correlation analysis. These analyses demonstrated that CYP pathway, which generates bioactive epoxyeicosatrienoic acids (EETs) and hydroxyeicosatetraenoic acid (HETEs) to regulate cell proliferation, survival, invasion, metastasis, and angiogenesis^[121]23, was the significantly elevated signature in AFP^− HCC (Fig. [122]2e). Notably, the CYP pathway was the only branch exhibiting a strong negative correlation with AFP expression (Fig. [123]2f and Supplementary Fig. [124]4e, f). This observation indicated that tumor cells in AFP^− HCC were more likely to reshape their metabolic patterns through the CYP pathway. Elevation of 11,12-EET in AFP^– HCC tumor cells versus AFP^+ HCC In agreement with the elevated activity of CYP pathway in AFP^− HCC, the CYP family genes, including CYP2C8 and CYP1B1, were upregulated in AFP^− HCC patients (Fig. [125]2g), and were confirmed by RT-qPCR assays in tumor samples (Fig. [126]2h). We then further sought to determine the key metabolite participating in the arachidonic acid metabolism in AFP^− HCC. Through targeted lipidomics, we identified five significantly upregulated lipid-related metabolites in AFP^– HCC compared to that in AFP^+ HCC (Fig. [127]2i). We decided to study 11,12-EET because: (1) the top two metabolites (ARA and DTA) served as the common substrates for COXs, LOXs, and CYPs as documented in prior studies^[128]23–[129]26, whereas only the CYP pathway showed significant activation in AFP^− HCC; (2) 11,12-EET was the specific and direct metabolic product of the CYP pathway; (3) the upregulated CYP2C8 and CYP1B1 in AFP^− HCC were the key metabolic enzymes responsible for 11,12-EET generation^[130]23,[131]27; (4) elevation of 11,12-EET inversely correlated with serum AFP levels (Supplementary Fig. [132]4g); (5) moreover, 11,12-EET was reported to exhibit diverse biological functions^[133]23 that might play an important role in tumor biology. To validate the expression relation between AFP and CYPs/11,12-EET, AFP expression level in five HCC cell lines was further measured, where PLC5, Hep3B and Huh7 were defined as AFP^high HCC cell lines; MHCC97H and HCCLM3 belonged to AFP^low HCC cell lines (Fig. [134]2j). In line with DEGs and lipidomics analysis, AFP^low cell lines displayed elevated CYP2C8 and CYP1B1 expression (Supplementary Fig. [135]4h, i) and higher 11,12-EET levels (Fig. [136]2k) in comparison to the AFP^high cell lines, solidifying 11,12-EET as the pivotal metabolite in AFP^– HCC. Upregulated AR transcriptional activity in AFP^– HCC tumor cells Transcription factors (TF) are key molecules that control gene expression, and their activities determine the cell functions and response to intricacies of environmental disturbances^[137]28. Thus, to elucidate the potential regulatory mechanisms for the upregulation of arachidonic acid metabolism in AFP^− HCC, we assessed the TF activity across tumor cells in scRNA-seq data using the single-cell regulatory network inference and clustering algorithm (SCENIC)^[138]29. The SCENIC results showed that the TF activities of MSC, KLF9, EBF1, IRF6, and AR were significantly upregulated in AFP^− HCC tumor cells (Fig. [139]3a and Supplementary Data [140]8). Notably, AR transcriptional activity showed the most substantial negative correlation with AFP expression and positive association with CYP-related gene expression (Fig. [141]3b and Supplementary Data [142]9). In agreement with this finding, AR was further identified as the second most significant TF in AFP^– HCC based on the bulk RNA-seq data (Supplementary Fig. [143]5a, b). Fig. 3. Upregulation of AR transcription activity in AFP^− HCC tumor cells. [144]Fig. 3 [145]Open in a new tab a SCENIC analysis in scRNA-seq data shows TF activities in AFP^− HCC and AFP^+ HCC tumor cells. b Spearman correlation analysis between activity of TFs listed in (a), and AFP expression (colored in blue) as well as CYP genes signature (colored in red) in tumor cells. c Detection of AR expression using IHC in AFP^– HCC (n = 10) and AFP^+ HCC (n = 10). Scale bars, 50 μm. d Correlation between AFP and AR expression by Spearman correlation test (n = 20). e Relative expression of AR detected by RT-qPCR (AFP^− HCC vs. AFP^+ HCC = 15:15). f AR expression in AFP^– HCC compared to AFP^+ HCC, as determined by bulk RNA-seq data from FAH-SYSU (AFP^– HCC vs. AFP^+ HCC = 161:273) and TCGA-LIHC cohort (AFP^– HCC vs. AFP^+ HCC = 142:127). g Concentrations of testosterone and 5α-DHT detected by lipidomics using blood plasma from 10 AFP^−/AFP^+ HCC. h Relative expression of AR detected by RT-qPCR and western blot in HCC cells. i Relative concentration of 11,12-EET detected by ELISA assays in CM from HCC cells. j Relative expression of CYP2C8 and CYP1B1 with or without 5α-DHT stimulation. k Validation of the AR knockdown (shAR) in MHCC97H and HCCLM3 using western blot. l 11,12-EET concentrations detected by ELISA assays in CM from MHCC97H and HCCLM3 cells (shNC) compared to their shAR counterparts. m Detection of 11,12-EET levels in AR-overexpressing Hepa1-6 cells (oeAr) versus control cells (oeNC). Validation of AR binding with CYP2C8 and CYP1B1 promoters by ChIP-PCR (n) and ChIP-qPCR (o) in MHCC97H and HCCLM3. p Normalized snATAC-seq tracks of AFP, AR, CYP2C8 and CYP1B1. Three biological replicates were employed (h–j, l, m, o). Data were shown as mean ± SD (c, g–j, l, m, o). *p < 0.05, **p < 0.01, ***p < 0.001 and ns stands for no significance by Wilcoxon rank sum test (f, g), two-sided t-test in (c, e, i, j, m) and one-way ANOVA with Tukey’s multiple comparisons test (h, l, o). Source data are provided as a Source Data file. To further examine the TF activity of AR in AFP^− HCC, we analyzed and intersected the predicted target genes of AR from scRNA-seq and bulk RNA-seq data, as shown in Supplementary Data [146]10 and [147]11, respectively. We observed that these potential AR-target genes exhibited lower methylation levels in AFP^− HCC, as revealed by RRBS and TCGA-LIHC 450 K methylation sequencing, compared to those in AFP^+ HCC (Supplementary Fig. [148]5c, d). The reduced methylation of these target genes indicated the function of AR acting as a TF in AFP^− HCC. Moreover, AR nuclear expression was observed significantly higher in AFP^− HCC tumors than that in AFP^+ HCC tumors, and was significantly negatively correlated with AFP expression (Fig. [149]3c, d). Taken together, these findings suggested that AR might be a core TF of AFP^− HCC tumor cells. As a male hormone receptor and major oncogenic driver in prostate cancer^[150]30, AR was also highly expressed in other cancer types, including breast cancer and HCC (Supplementary Fig. [151]6a). However, the role of AR in HCC is controversial. Some studies have reported the negative prognostic effect of AR in HCC patients^[152]31, while others have demonstrated the anti-tumor functions of AR through inhibiting p38 and NF-κB pathways in HCC cells^[153]32. Therefore, we sought to determine the regulatory functions of AR in AFP^− HCC. We observed a tightly positive correlation of AR with CYP pathway genes in HCC (Supplementary Fig. [154]6b). Interestingly, in addition to the upregulation of TF activity, the transcriptomic levels of AR were also upregulated in AFP^− HCC tumors (Fig. [155]3e, f). Generally, HCC is a male-dominant cancer, and AR is highly expressed in male HCC patients^[156]30,[157]33. Given the higher expression and activities of AR in AFP^− HCC, we then further quantitatively analyzed plasma concentrations of testosterone and 5α-dihydrotestosterone (5α-DHT), the primary endogenous ligands of AR in HCC patients^[158]34. Notably, these androgen levels exhibited significant elevation in AFP^− HCC compared to AFP^+ HCC (Fig. [159]3g), further suggesting that AR might be significantly activated in AFP^− HCC patients. AR activates the CYP pathway to promote 11,12-EET production in AFP^− HCC tumor cells To explore whether AR participated in the activation of the CYP pathway, we identified PLC5, Hep3B, Huh7 and Hepa1-6 as AFP^highAR^low cell lines, while MHCC97H, HCCLM3 and RIL-175 were labeled as AFP^lowAR^high cell lines by RT-qPCR and western blot assays (Fig. [160]3h and Supplementary Fig. [161]6c, d). Next, AFP^lowAR^high, but not the AFP^highAR^low cell lines, boosted the production of 11,12-EET and the expression of CYP2C8 and CYP1B1 in the presence of 5α-DHT stimulation (Fig. [162]3i, j). In consistent with this finding, depletion of AR markedly repressed 11,12-EET production in AFP^lowAR^high cells (Fig. [163]3k, l); whereas overexpression of AR significantly promoted the production of 11,12-EET in AFP^highAR^low cells (Fig. [164]3m and Supplementary Fig. [165]6e). These data together indicated the critical role of AR in regulation of the CYP pathway activity in AFP^− HCC. Next, we aimed to elucidate whether AR could directly regulate the expression of CYP2C8 and CYP1B1. By using JASPAR database^[166]35, we predicted AR binding sites on the CYP2C8 and CYP1B1 promoters and generated the primers. ChIP assays were then employed, and the data confirmed that AR could bind with CYP2C8 and CYP1B1 promoters (Fig. [167]3n). Besides, ChIP-qPCR results showed that AR depletion could significantly attenuate its TF activity on CYP2C8 and CYP1B1 (Fig. [168]3o). Furthermore, AFP^− HCC tumors exhibited higher chromatin accessibility signals of AR, CYP2C8, and CYP1B1, whereas displayed lower accessibility signals of AFP in comparison to AFP^+ HCC (Fig. [169]3p and Supplementary Data [170]12). Taken together, these observations indicated that AR positively regulated the CYP pathway activity in AFP^− HCC and promoted the production of 11,12-EET by enhancing the transcription of CYP2C8 and CYP1B1. Enrichment of tip-like endothelial cells in AFP^− HCC microenvironment Accumulating evidence has suggested that metabolic alterations in tumor cells modulated the composition and function of surrounding cells, thereby reshaping the TME^[171]36,[172]37. To study the tumor microenvironment of AFP^– HCC, we first compared the cellular composition between AFP^− HCC and AFP^+ HCC by calculating the percentage of each cell subtype relative to the total cell population in each sample using the scRNA-seq data. The analysis revealed a significant enrichment of the endothelial cell subtype EC_PLPP1 in AFP^– HCC (Fig. [173]4a), which was further supported by the results analyzed by the Milo algorithm^[174]38, a statistical framework that assesses differential abundance by assigning cells to partially overlapping neighborhoods on a k-nearest neighbor graph (Supplementary Fig. [175]7a, b). Furthermore, the results of cellular composition were also validated by xCell deconvolution based on bulk RNA-seq data (Fig. [176]4b) and IHC staining of CD31, a specific biomarker of endothelial cell^[177]39, in HCC tumor samples (Fig. [178]4c). Notably, the CD31 level exhibited a negative correlation with AFP expression as analyzed from the IHC data (Supplementary Fig. [179]7c). Altogether, these results implied that the angiogenesis ability might be a distinct signature of AFP^− HCC compared to AFP^+ HCC. Fig. 4. Enrichment of endothelial cells in AFP^– HCC. [180]Fig. 4 [181]Open in a new tab a Cell proportion analysis of each cell cluster between AFP^− HCC and AFP^+ HCC patients, as shown by boxplots and pie charts. P values are displayed only for cell types with significant differences between the two groups (AFP^− HCC vs. AFP^+ HCC = 6:11). b Endothelial cell proportion deconvoluted by xCell using bulk RNA-seq from FAH-SYSU cohort (AFP^− HCC vs. AFP^+ HCC = 161:273) and TCGA-LIHC cohort (AFP^− HCC vs. AFP^+ HCC = 142:127), which showed higher infiltration of endothelial cells in AFP^− HCC. c Representative images of IHC staining in FFPE tissues (left panel) and microvascular density quantification (right panel) in AFP^− HCC and AFP^+ HCC. The number of dots indicates the number of patients (AFP^− HCC vs. AFP^+ HCC = 10:10). Data were shown as mean ± SD. Scale bars, 50 μm. d The UMAP visualization of four subtypes of endothelial cells (upper panel) and corresponding tissue distribution (lower panel). e Marker gene expression for endothelial cells. f Pathway enrichment analysis for endothelial cells using GO-BP gene sets. g Functional scoring of endothelial cell phenotypes based on tip-like and stalk-like gene signatures. *p < 0.05, **p < 0.01, ***p < 0.001, and ns stands for not significant by Wilcoxon rank sum test in (a, b), two-sided t-test in (c) and Wilcoxon rank sum test with FDR correction (g). Source data are provided as a Source Data file. We thus further explored the distribution and function heterogeneity of endothelial cells between AFP^– HCC and AFP^+ HCC using the scRNA-seq data (Fig. [182]4d and Supplementary Fig. [183]7d). The EC_FTL with high expression of FTH1 and FTL, was associated with metabolic pathways such as cell respiration, oxidative phosphorylation, and iron ion transport^[184]40 (Fig. [185]4e, f). EC_CLEC4G, with high expression of canonical liver sinusoidal endothelial cell (LSEC) markers like CLEC4G, CD14 and CLEC4M, showed preferential enrichment in NATs of AFP^+ HCC and presented active pathways related to wound healing, regulation of immune effector process, and receptor-mediated endocytosis (Fig. [186]4e, f and Supplementary Fig. [187]7d). EC_PLPP1 with expression of IGFBP3, CXCR4 and PLPP1 displayed preferential enrichment in AFP^– HCC tumor tissues, and upregulated pathways related to ameboidal-type cell migration, leukocyte migration, and endothelial development (Fig. [188]4e, f and Supplementary Fig. [189]7d), likely to be the tumor-associated endothelial cell. EC_ACKR1 with high expression of ACKR1 and VWF was featured by the enriched pathways related to regulation of cell-cell adhesion, positive regulation of T cell activation, and T-helper cell differentiation (Fig. [190]4e, f), likely to be the venous-related endothelial cell^[191]41. Endothelial cells commonly exhibited two functional phenotypes in tumor angiogenesis. The tip-like endothelial cells were characterized by cell budding, migration, and formation of filopodia, while the stalk-like endothelial cell presented proliferation, adhesion, and tight connection abilities to ensure the stability of new sprouts^[192]42,[193]43. Here, we observed that EC_PLPP1, enriched in AFP^− HCC, exhibited a high tip-like signature, and significantly upregulated tip-like endothelial cell-related pathways (Fig. [194]4g and Supplementary Fig. [195]7e), indicating the enhancing angiogenesis ability in AFP^− HCC. Upregulated AR transcriptional activity promotes angiogenesis in AFP^− HCC Next, we aimed to examine whether AR orchestrated the enrichment of tip-like endothelial cells in AFP^− HCC. To this end, we performed correlation analysis and found that the expression of nuclear AR was significantly positively correlated with CD31 expression (Fig. [196]5a). Furthermore, spatial transcriptomics analysis using a public dataset revealed that AR^+ tumor spots were in closer proximity to PECAM1^+ spots (namely endothelial cell spot) compared to AR^– tumor spots (Supplementary Fig. [197]8a, b). These results implied that AR might play a crucial role in regulating angiogenesis within the AFP^− HCC TME. Fig. 5. Activation of EGFR with AR-induced 11,12-EET from AFP^− HCC tumor cells promoted aberrant angiogenesis. [198]Fig. 5 [199]Open in a new tab a Scatter plot showing positive correlation between proteinic nuclear AR expression and proteinic CD31 expression. R indicates the correlation coefficient calculated by Spearman correlation test. The number of dots indicates the number of patients. b Sensorgrams for detecting the binding affinity of 11,12-EET and EGFR using SPR assay. c Representative images (left panel) and quantitative analysis (right panel) demonstrating the pro-angiogenic effects of 11,12-EET on HUVEC migration and tube formation compared to non-11,12-EET treatment group (NC). A total of three biological replicates were employed. Data were shown as mean ± SD. d Protein expression of VEGFR2, phosphorylated VEGFR2 (p-VEGFR2), EGFR2 and phosphorylated EGFR2 (p-EGFR2) in HUVECs when co-cultured with 11,12-EET compared to non-11,12-EET treatment group (NC). Representative images (left panel) and corresponding statistical results (right panel) of migration (e) and tube formation (f) of HUVECs when co-cultured with the CM from shNC and shAR cell lines, including MHCC97H and HCCLM3, respectively. A total of three biological replicates were employed. Data were shown as mean ± SD. g Representative images (left panel) and quantitative analysis (right panel) of migration and tube formation of HUVECs when co-cultured with CM from MHCC97H. The MHCC97H cells were pre-treated with 10 nM 5α-DHT for 48 h, followed by separate treatments with either 6 μM lenvatinib, 10 μM gefitinib, 20 μM gefitinib or vehicle control for 24 h. A total of three biological replicates were employed. Data were shown as mean ± SD. Scale bars, 100 μm. *p < 0.05, **p < 0.01, ***p < 0.001 and ns stands for no significance by two-sided t-test in (c) and one-way ANOVA with Tukey’s multiple comparisons test (e–g). Source data are provided as a Source Data file. Due to the strong association between AR transcriptional activation and 11,12-EET production in AFP^− HCC, we interrogated whether AR-mediated 11,12-EET could promote sprouting angiogenesis by interacting with tip-like endothelial cells. To this end, we first predicted the binding affinity of 11,12-EET to the endothelial-specific pro-angiogenic targets by Autodock^[200]44,[201]45, and found that 11,12-EET exhibited the strongest binding affinity to EGFR among the candidate receptors (Supplementary Fig. [202]8c). Moreover, the data from surface plasmon resonance (SPR) assay confirmed that 11,12-EET could directly bind to EGFR (Fig. [203]5b). As a result, EGFR expression and the downstream signaling pathway were also significantly upregulated in EC_PLPP1 of AFP^− HCC (Supplementary Fig. [204]8d, e). We then performed in vitro assays to examine the interaction between AR-mediated 11,12-EET production and endothelial cells. The 11,12-EET treatment was shown to promote vascular migration and tube formation of human umbilical vein endothelial cells (HUVECs) (Fig. [205]5c). Of note, the phosphorylation level of EGFR, but not VEGFR2, was upregulated in the presence of 11,12-EET stimulation (Fig. [206]5d). Bulk RNA-seq data derived from HUVECs treated with 11,12-EET also confirmed the activation of EGFR-related signaling pathway, while VEGFR-related pathway remained unaffected (Supplementary Fig. [207]8f, g). In addition, we cultured HUVECs with the conditioned medium (CM) from AFP^lowAR^high cell lines (MHCC97H and HCCLM3). The results demonstrated that vascular cell migration and tube formation were robustly inhibited in the AR-knockdown group (Fig. [208]5e, f). Notably, we further revealed that gefitinib, an EGFR inhibitor, showed a significantly inhibitory effect on HUVEC migration and tube formation than lenvatinib in a dose-dependent manner (Fig. [209]5g), demonstrating that lenvatinib showed mild inhibitory effect on tumor angiogenesis in the presence of 11,12-EET stimulation in AFP^− HCC. Overall, AR-11,12-EET-EGFR axis played a critical role in promoting angiogenesis in AFP^− HCC. AR modulates tumorigenesis and vascularization through 11,12-EET-EGFR axis in vivo To demonstrate the eliciting effects of AR on tumorigenesis and vascularization in vivo, we established the subcutaneous HCC models using two human cell lines (MHCC97H, defined as AFP^lowAR^high; Huh7, defined as AFP^highAR^low) in immunodeficient NCG mice, and two murine cell lines (RIL-175, defined as Afp^lowAr^high; Hepa1-6, defined as Afp^highAr^low) in immunocompetent C57BL/6 mice, respectively. Lentiviral transfection-mediated gene expression was performed to generate AR-knockdown counterparts for AFP^lowAR^high cell lines and AR-overexpression counterparts for AFP^highAR^low cell lines (Fig. [210]3k and Supplementary Figs. [211]6e, [212]9a). In NCG mice, AR knockdown significantly suppressed tumor burden, manifested as reduced invasive growth, decreased tumor volume and weight (Fig. [213]6a–e), accompanied with lower expression of the proliferative marker Ki67 (Supplementary Fig. [214]9b). In contrast, AR overexpression significantly fueled the tumor growth (Fig. [215]6f–i and Supplementary Fig. [216]9c). We further examined the expression profile of AR-11,12-EET-EGFR axis. The results showed that knockdown of AR significantly inhibited the expression levels of CYP2C8 and CYP1B1 (Supplementary Fig. [217]9d) as well as the production level of 11,12-EET (Fig. [218]6j). The expression of downstream molecule EGFR on endothelial cells (Fig. [219]6k) and the microvascular density (Supplementary Fig. [220]9e, f) were also reduced upon AR knockdown, while overexpression of AR generated significant contrast results (Fig. [221]6j, [222]l and Supplementary Fig. [223]9g–i). Fig. 6. Modulation of AR in tumorigenesis and vascularization in vivo. [224]Fig. 6 [225]Open in a new tab a Experimental design using NCG mice (n = 6 per group). Representative images of tumors (b), tumor growth curve (c), tumor volume measurements (d) and tumor weight measurements (e) in MHCC97H subcutaneous HCC mouse models (shNC vs. shAR = 6:6). Representative images of tumors (f), tumor growth curve (g), tumor volume measurements (h) and tumor weight measurements (i) in Huh7 subcutaneous HCC mouse models (oeNC vs. oeAR = 6:6). j Relative concentrations of 11,12-EET measured by ELISA assays in MHCC97H and Huh7 subcutaneous HCC tumors, and their corresponding shAR and oeAR counterparts (n = 6 per group), respectively. Flow cytometry analysis of CD45^-CD31^+EGFR^+ endothelial cell (EC) fractions in MHCC97H (k) and Huh7 (l) subcutaneous HCC tumors (n = 6 per group). m Experimental design using C57BL/6 mice (n = 6 per group). Representative images of tumors (n), tumor growth curve (o), tumor volume measurements (p) and tumor weight measurements (q) in RIL-175 subcutaneous HCC mouse models (shNC vs. shAr = 6:6). Representative images of tumors (r), tumor growth curve (s), tumor volume measurements (t) and tumor weight measurements (u) from Hepa1-6 subcutaneous HCC mouse models (oeNC vs. oeAr = 6:6). v Relative concentration of 11,12-EET detected by ELISA kit in RIL-175 and Hepa1-6 subcutaneous HCC tumors, and their corresponding shAr and oeAr counterpart (n = 6 per group), respectively. Flow cytometry analysis of EGFR^+ ECs fractions in RIL-175 (w) and Hepa1-6 (x) subcutaneous HCC tumors (n = 6 per group). Data were shown as mean ± SD. *p < 0.05, **p < 0.01, ***p < 0.001 by two-way ANOVA in (c, g, o, s) and two-sided t-test in (d, e, h–l, p–q, t–x). Source data are provided as a Source Data file. In C57BL/6 mice, similar to the results observed in NCG mice models using human HCC cell lines, knockdown of Ar significantly repressed the HCC growth and tumor angiogenesis while overexpression of Ar promoted HCC progression and vascularization within the same species (Fig. [226]6m–x and Supplementary Fig. [227]10a–h) Altogether, these data revealed the important functions of AR in regulation of tumorigenesis and vascularization in AFP^− HCC. To further consolidate the regulatory link between AR and 11,12-EET-EGFR axis, we established an orthotopic HCC mouse model using MHCC97H cells, comprising four experimental groups: NC (control), shAR (AR knockdown), oeCYP2C8 (CYP2C8 overexpression), and shAR plus oeCYP2C8 (AR knockdown with CYP2C8 rescue) (Supplementary Fig. [228]11a, b). The data demonstrated that AR knockdown significantly suppressed tumor growth (Supplementary Fig. [229]11c, d) accompanied with reduced Ki67 (Supplementary Fig. [230]11e), CYP2C8/CYP1B1 expression (Supplementary Fig. [231]11f) and 11,12-EET production level (Supplementary Fig. [232]11g) as well as decreased endothelial EGFR expression level (Supplementary Fig. [233]11h) and microvascular density (Supplementary Fig. [234]11i, j), while CYP2C8 overexpression partially rescued the shAR-eliciting anti-tumor and anti-angiogenesis effects (Supplementary Fig. [235]11). The in vivo experimental data provided bona fide evidence supporting the regulatory role of AR in control of 11,12-EET-EGFR axis that promoted angiogenesis in AFP^− HCC. Combining AR inhibitor bicalutamide with lenvatinib suppresses tumor progression and angiogenesis in AFP^− HCC in vivo To investigate the therapeutic potential of targeting the AR-11,12-EET-EGFR for AFP^− HCC, we established an orthotopic HCC mouse model by injecting MHCC97H cells into the liver lobes of NCG mice (Fig. [236]7a). The control and treatment groups were set as follows: (1) shNC groups including: (1) vehicle control (Ctrl); (2) gefitinib monotherapy (GEF); (3) bicalutamide monotherapy (BIC); (4) lenvatinib monotherapy (LEN); (5) bicalutamide plus gefitinib combination (BIC + GEF); (6) bicalutamide plus lenvatinib combination (BIC + LEN). (2) shAR groups including: (1) vehicle control (Ctrl); 2) gefitinib monotherapy (GEF); (3) bicalutamide monotherapy (BIC); (4) lenvatinib monotherapy (LEN); 5) bicalutamide plus gefitinib combination (BIC + GEF); (6) bicalutamide plus lenvatinib combination (BIC + LEN). Fig. 7. Combined inhibition of AR-mediated anti-angiogenesis with lenvatinib repressed tumor progression in AFP^− HCC in vivo. [237]Fig. 7 [238]Open in a new tab a Schematic of the experimental design using NCG mice (n = 5 per group). Representative gross images of MHCC97H orthotopic tumors (b) and statistical analysis of tumor volume (c). The “+” indicates treatment with gefitinib (GEF), bicalutamide (BIC) or lenvatinib (LEN), while the “–” indicates treatment with vehicle control (n = 5 per group). Data were shown as mean ± SD. d, e Representative images of IHC staining (left panel) and microvascular density quantification (right panel) in MHCC97H orthotopic tumors across different treatment groups (n = 5 per group). Data were shown as mean ± SD. f A graphical summary. In AFP^− HCC patients, TF activity of AR was elevated and promoted the expression levels of CYP2C8 and CYP1B1, which enhanced arachidonic acid metabolism and the production of the downstream metabolite 11,12-EET. The 11,12-EET metabolite activated EGFR on endothelial cells, thereby promoting angiogenesis. Targeting this pathway, a combination of AR antagonists with lenvatinib might represent a promising therapeutic strategy for AFP^− HCC. Scale bars, 50 μm. *p < 0.05, **p < 0.01, ***p < 0.001, and ns stands for no significance by one-way ANOVA with Tukey’s multiple comparisons test (c, e). Source data are provided as a Source Data file. In the shNC group, GEF monotherapy showed negligible effects on tumor burden compared to controls, whereas both LEN and BIC monotherapies elicited moderate tumor suppression (Fig. [239]7b, c and Supplementary Fig. [240]12a, b) accompanied by reduced angiogenesis (Fig. [241]7d, e). Notably, the combination of BIC + GEF moderately reduced the tumor growth, while the combination of BIC + LEN demonstrated the most remarkable suppression of tumor burden (Fig. [242]7b, c and Supplementary Fig. [243]12a, b). In the shAR group, the vehicle controls (AR knockdown solo) exhibited significantly lower baseline tumor burden and microvascular density compared to shNC controls (Fig. [244]7b–e), further validating the role of AR in regulation of tumor progression and vascularization. Among the monotherapy groups, only LEN treatment significantly repressed tumor burden compared to the shAR alone. Of note, both BIC + GEF and BIC + LEN significantly damped down the tumor burden, with the BIC + LEN showing more pronounced suppressive effects in tumorigenesis and vascularization (Fig. [245]7b–e). Furthermore, to solidify the synergistic effects of AR antagonist and lenvatinib, we also established a Hepa1-6 orthotopic HCC mouse model (oeNC vs. oeAR). The results showed that lenvatinib monotherapy moderately inhibited tumor growth but this inhibitory effect was significantly attenuated upon Ar overexpression (Supplementary Fig. [246]12c, d). Besides, combination of lenvatinib with bicalutamide significantly suppressed tumor burden, reduced Ki67 level and decreased microvascular density in the oeAr group (Supplementary Fig. [247]12c–f), further supporting the therapeutic potential of this combined treatment approach. Taken together, we uncovered the hyperactive metabolic activity of arachidonic acid metabolism mediated by AR in AFP^− HCC, and shed insights into the mechanism of AR-induced aberrant angiogenesis, providing a promising therapeutic strategy of combining AR antagonists (e.g., bicalutamide) and lenvatinib for AFP^− HCC (Fig. [248]7f). Discussion In the present study, we identified arachidonic acid metabolism as a specifically activated metabolic pathway in AFP^− HCC, with AR being its key regulatory factor by integrated omics analysis and multiple experiments. Mechanistically, AR could promote the transcription of CYP2C8 and CYP1B1 by binding to their promoter regions, thereby enhancing the arachidonic acid metabolism that produced 11,12-EET, which activated EGFR on endothelial cells and eventually fueled tumor angiogenesis. Building on the potent roles of AR in metabolic reprogramming^[249]46, including its regulation on lipid pathways of prostate cancer^[250]47, and cholesterol metabolism^[251]48 as well as glycolysis^[252]49 in HCC, here, we unveil the impact of AR-mediated arachidonic acid metabolism on the reprogramming of AFP^− HCC TME, providing AR as the target for AFP^− HCC treatment. To our knowledge, the role of AR in HCC remains highly controversial. AR is originally regarded as an oncogenic driver of HCC, maintaining the stemness of HCC cells and promoting cell cycle process^[253]30,[254]50,[255]51. However, parts of clinical trials have not shown significant beneficial effects of anti-androgenic therapy in HCC patients^[256]52, and some studies showed that hepatic AR suppresses HCC metastasis^[257]32. These findings implied that hepatic AR might have dual but opposing roles, resulting in a dilemma in personalized treatment of HCC, and further fueling our interest in potential AR-driven TME reprogramming of AFP^− HCC tumors. Furthermore, our findings demonstrated that the AR-CYP2C8/CYP1B1 axis promoted the generation of the downstream metabolite 11,12-EET, which could activate the EGFR on endothelial cells, thus inducing a tip-like phenotype in endothelial cells and ultimately leading to aberrant angiogenesis. Currently, anti-angiogenic therapy targeting VEGFRs such as lenvatinib is strongly recommended as the first-line systemic treatment for the advanced HCC and is widely applied in the clinic^[258]53. However, several recent global clinical trials have indicated that AFP^− HCC patients showed moderate treatment response to the first-line drug lenvatinib or its analogues^[259]13–[260]15. Coincidentally, another study reported that HCC could be categorized into three subtypes (C1, C2, and C3) based on metabolism-related characteristic genes^[261]54. Among them, the C1 subtype exhibited higher metabolic activity and lower AFP expression, conferring high activity of angiogenesis-related characteristics^[262]54, which exhibited resistant to cabozantinib, a second-line anti-angiogenic drug targeting VEGFR2^[263]54. In this study, we employed various animal models and demonstrated that AR^high HCC mice possessed higher tumor burden and vascular density than those in AR^low HCC mice, while combination therapy with bicalutamide and lenvatinib significantly inhibited tumor burden and vascular density in AFP^− HCC. Interestingly, both in vitro and in vivo findings suggested that AFP^− HCC might promote angiogenesis through AR-induced metabolic reprogramming and subsequent generation of 11,12-EET that binding to EGFR on endothelial cells, which might lead to the insensitivity of AFP^− HCC to canonical anti-angiogenic drugs. In conclusion, our findings reveal the critical role of AR in regulation of arachidonic acid metabolism in AFP^− HCC and provide insights into the mechanism of AR-CYPs-11,12-EET-EGFR axis-induced tumor angiogenesis. These findings not only disclose the link between tumor metabolism and abnormal vascular growth but also highlight the promising strategy of combining AR antagonists to enhance the efficacy of lenvatinib therapy in AFP^− HCC. Methods Study approval The study was approved by the Institutional Research Ethics Committee of First Affiliated Hospital of Sun Yat-sen University (FAH-SYSU) under Ethical Approval ID [2023]380. Informed consent, including to publish clinical information potentially identifying individuals, was obtained from all patients. All animal experiments were performed in accordance with protocols approved by the Institutional Animal Care and Use Committee, Sun Yat-sen University (IACUC, SYSU). Approval numbers are SYSU-IACUC-2024-000262 and SYSU-IACUC-2024-000264. Patients and samples A total of 448 treatment-naïve patients who underwent curative liver resection at FAH-SYSU were included, with tumor samples and NAT samples. This study has received approval from the Ethics Committee of FAH-SYSU and adhered to relevant ethical standards. Clinical information of patients was summarized in Supplementary Data [264]1 and Supplementary Data [265]4. Two experienced pathologists confirmed the diagnoses through histological examinations. The sequencing types utilized in this study and their corresponding sample information were provided in Supplementary Fig. [266]1A. Exome and transcriptome capture, library preparation and sequencing For exome capture and library preparation, total DNA was extracted from snap-frozen tissues using the QIAGEN DNeasy Blood & Tissue Kit (Qiagen, Hilden, Germany). Qualified DNAs from tissue samples were fragmented into 200-300 bp fragments using Covaris technology with adapters ligated to both ends. Next, the extracted DNA was amplified by ligation-mediated PCR (LM-PCR), then purified and hybridized to NimbleGen or Agilent human exome array for enrichment. Non-hybridized fragments were washed away. For transcriptome capture and library preparation, total RNA was extracted using the Trizol (Invitrogen, USA). Poly(A) mRNA was isolated from total RNA using oligo-dT-coupled beads. The mRNA was fragmented into short fragments by adding a fragmentation buffer. Using these short fragments as templates, the first-strand cDNA was synthesized with random hexamer primers. The second-strand cDNA was then synthesized using buffer, dNTPs, RNase H, and DNA polymerase I. The short fragments were purified using the QIAQuick PCR extraction Kit (Qiagen, Hilden, Germany), dissolved in EB buffer for end reparation and adding poly(A). Next, the short fragments were ligated with sequencing adapters. Finally, exome and transcriptome libraries were constructed and loaded onto the Illumina NovaSeq 6000 sequencing platform, generating paired-end reads of 150 base pairs. Mutation calling and annotation FASTQ files were filtered for adapter sequences and low-quality reads using TrimGalore (v.0.6.1, [267]https://github.com/FelixKrueger/TrimGalore). FastQC was then used to evaluate the quality of each sequencing read. Qualified reads were mapped to the human reference genome GRCh38 in paired-end mode using BWA-mem2^[268]55 (v.2.2.1). Duplicate reads were marked using Sambamba^[269]56 (v.0.8.2). Following GATK best practices, Mutect2^[270]57 was then employed to detect confident somatic mutations. ANNOVAR^[271]58 was used to annotate the final somatic mutations. Finally, MutSigCV^[272]59 (v.1.3.5) was employed to identify significantly mutated genes with all SNVs and INDELs used as input. Genes with q-values < 0.05 were defined as significantly mutated. Tumor mutation burden (TMB), defined as the number of somatic mutations per megabase (Mb) in the coding region, was determined and quantified using our WES data. To avoid overestimating TMB, mutations with population frequencies higher than 0.05 in the Genome Aggregation Database and driver mutations curated in COSMIC database were excluded. Detection of CNVs CNVkit^[273]60 (v.0.9.10) was employed to call CNVs, with tumor copy number estimates derived from BAM files using batch processing and segmentation methods. GISTIC^[274]61 (v.2.0.23) was applied to process the results files from the previous step, identifying focal amplifications and deletions in cytoband level with amplification threshold set to 0.3, deletion threshold set to −0.3. The G-score was calculated based on the amplitude of the CNVs and its frequency of occurrence in the sample, reflecting the frequency of CNVs in the segments. To compare CNV differences between AFP^− HCC and AFP^+ HCC patients, we first identified frequently amplified and deleted events across all patients (q value < 0.1). Next, we conducted a comparative analysis of frequently amplified and deleted events across cytobands in AFP^− HCC and AFP^+ HCC samples. For each cytoband, we calculated the amplification or deletion frequency of CNVs in both groups and used chi-square test or Fisher’s exact test with FDR adjustment to evaluate the differences. Statistical significance was defined as FDR of less than 0.05. Microsatellite instability estimation MSIsensor^[275]62 was used to detect frameshift mutations in microsatellite regions and distinguish them as somatic mutations. In brief, a chi-square test or Fisher’s exact test was employed to compare the length distribution of microsatellite sequences repeated at each locus between matched tumor and NAT samples. If the MSIsensor score of a sample was greater than 3.5, it was considered microsatellite instability (MSI); otherwise, it was classified as microsatellite stability (MSS). Cancer cell fraction analysis PyClone^[276]63 (v.0.13.1) was used to assign somatic mutations to clonal clusters and estimate their cellular prevalence. The cancer cell fraction (CCF) of each somatic mutation was estimated based on the SNV frequency and copy number profile. For each sample, the corresponding tumor purity was used to adjust the estimated CCF. Immune editing Polysolver^[277]64 (v.1.0) was employed to determine the HLA typing for each patient using the corresponding NAT sample. The non-synonymous mutations that occurred in each sample were obtained from the processed VCF file. Corresponding amino acid sequences were acquired from UniProt database, with the amino acid sequences sliding around the mutation site (9–11 residues) being selected as input. These mutated sequences, along with the HLA typing information for each sample, were collectively input into NetMHCpan^[278]65 (v.4.0). Peptide segments with predicted binding affinities less than 500 nM were considered as immunogenic and used for subsequent analysis. The same process was applied to TCGA-LIHC cohort to construct a background condition. Finally, the immune editing score was calculated according to the literature published previously^[279]66,[280]67. Mutation signature identification SigProfilerExtractor^[281]68 (v.1.1.21) was used to de novo decompose the mutation spectrum in the processed VCF files for all samples. The decomposed signatures were compared to known signatures from the Catalogue of Somatic Mutations in Cancer (COSMIC) database (v.3) using cosine similarity (>0.8) to determine their corresponding types. Wilcoxon rank sum test was applied to assess activity differences in contributions to each signature between AFP^− HCC and AFP^+ HCC. RNA sequencing data analysis By filtering low-quality reads and removing sequencing adapters using TrimGalore (v.0.6.1, [282]https://github.com/FelixKrueger/TrimGalore), the FASTQ files of all samples were mapped to the human reference genome GRCh38 using STAR^[283]69 (v.2.7.9) with default parameter. The gene-level read counts and expression values in fragment per kilobase per million mapped fragments (FPKM) units were obtained using the RSEM^[284]70 (v.1.2.28) pipeline. Based on the quantified gene read count, DESeq2^[285]71 was employed to assess the differentially expressed genes between AFP^− HCC and AFP^+ HCC, with a significance threshold set as adjusted p-value < 0.05 and |log[2]-transformed fold changes (log[2]FC)| > 1. GSEA^[286]72 between AFP^− HCC and AFP^+ HCC was conducted using pathway databases (Hallmark, KEGG, and REACTOME) from the Molecular Signatures Database (MSigDB, [287]https://www.gsea-msigdb.org/gsea/msigdb/index.jsp). In short, all detected genes were ranked based on the log[2]FC between AFP^− HCC and AFP^+ HCC for GSEA using the pre-ranked mode with default parameter settings. The xCell^[288]73 tool was employed to estimate the proportion of endothelial cells in treatment-naïve patients from FAH-SYSU cohort and TCGA-LIHC cohort data. Classification of HCC To delineate the transcriptional profiles of AFP^− HCC and AFP^+ HCC, we employed the NTP algorithm^[289]21 to predict the molecular subtypes of HCC based on the previously published Hoshida’s classification^[290]20,[291]21. NTP was utilized with default parameters. Differences in Hoshida’s classification were performed by chi-square test, based on methods described in previous studies^[292]74–[293]76. Namely, we conducted a chi-square test and calculated the ratio of observed to expected sample numbers for each subtype (Eqs. [294]1 and [295]2). [MATH: X2=fofe2fe :MATH] 1 [MATH: Ro/e=fo f e :MATH] 2 In the equation above, [MATH: X2 :MATH] represents the chi-square value. f[0] indicates the observed sample counts, while [MATH: fe :MATH] denotes the expected sample counts, which can be derived from chi-square test. [MATH: Ro/e :MATH] signifies the ratio of observed to expected sample counts for each subtype in different groups. If [MATH: Ro/e :MATH]  > 1, we assumed that a subtype showed enrichment preference in a specific group. Methylation profiling by RRBS Methylation profiling was performed using RRBS. Briefly, a total of 50 ng genomic DNA was digested with the MspI restriction enzyme for a sufficient duration (1–3 h) to generate DNA fragments, and then ligated with a methylated adapter containing complementary sticky end. The ligated products were bisulfite-converted with sodium bisulfite using the MethylCode^TM Bisulfite Conversion Kit (ThermoFisher, MECOV50), PCR-amplified and purified using magnetic beads according to manufacturer’s instruction. Purified library was quantified using Qubit, and then diluted to a specified concentration suitable for high-throughput sequencing. DNA library was conducted for sequencing and methylation analysis. Methylation analysis FASTQ files were filtered for adapter sequences and low-quality reads using TrimGalore (v.0.6.1, [296]https://github.com/FelixKrueger/TrimGalore). FastQC was then used to evaluate the quality of each sequencing reads. Subsequently, the qualified reads were aligned to GRCh38 reference assembly and the methylation state was determined using Bismark^[297]77 (v.0.24.1). In order to investigate whether there were differences in the methylation levels of specific genes between AFP^- HCC and AFP^+ HCC, we calculated the methylation levels within the promoter regions, defined as the 0-1500 bp upstream of the transcription start site (TSS) of the target genes. Generation of scRNA-seq data Fresh tumor samples were collected within 30 min after resection and preserved in MACS Tissue Storage Solution (130-100-008, Miltenyi Biotec). Subsequently, the samples were digested into a single-cell suspension following these steps: (I) Physical slicing of the tumor tissue using a blade, followed by digestion in RPMI 1640 containing 5% FBS using Tumor Dissociation Kit (130-095-929, Miltenyi Biotec) and DNaseI (DN25-100 mg, Sigma Aldrich) for 20 min at 37 °C; (II) The digested mixture was filtered through 70 μm and 30 μm MACS strainers (130-098-462; 130-098-458, Miltenyi Biotec). Cell suspension was centrifuged at 400 x g for 6 min, treated with 1× Red Blood Cell Lysis Buffer (00430054, Invitrogen) for 5 min at 4 °C. Cells were washed twice with PBS, and AOPI staining was performed to assess the concentration and viability of the single-cell suspension; (III) The concentration of the cell suspension was adjusted to ~700–1300 cells/μL, followed by reverse transcription and cDNA amplification according to the standard protocol (Single Cell 5’ Reagent Kits User Guide) to prepare the library construction, and sequenced on the Illumina NovaSeq 6000. scRNA-seq raw data processing The raw sequencing data were processed using CellRanger (v.6.1.2, 10x Genomics). The human genome GRCh38 was used to generate expression matrices. To eliminate contamination from ambient RNA present in droplets, we applied SoupX^[298]78 (v.1.5.2) to each sample. Cells with detectable expression of at least three genes were retained, while cells with fewer than 200 detected genes were excluded. Additionally, cells with gene counts exceeding 6000 and cells with a mitochondrial gene percentage greater than 15% were removed to ensure high-quality cells. Normalization of the data was performed based on the raw counts using the Seurat^[299]79 (v.4.1.0) R pipeline. The “vst” method in the Seurat FindVariableFeatures function identified the top 2000 highly variable genes for principal component analysis. A shared nearest neighbor graph was constructed using the top 50 principal components, followed by cell clustering using the Louvain algorithm. The resolution parameter was set to 0.7. Harmony^[300]80 (v.1.0.3) pipeline was applied to remove batch effects across all samples, encompassing both malignant and non-malignant cells, using default parameters. Visualization of the results was achieved using Uniform Manifold Approximation and Projection (UMAP). Typical doublets were selected based on co-expression of lineage-specific marker genes (e.g., T/B cells, NK/myeloid cells) and subsequently removed. Single cell CNV calling To confirm the identification of tumor cells, we employed inferCNV (v.1.14.2, [301]https://github.com/broadinstitute/inferCNV/wiki) to calculate CNV score for each cell. The gene expression matrix of tumor cells was extracted from the Seurat object, and CNVs were computed using cancer-associated fibroblasts and endothelial cells as references. The parameters used were “denoise” = TRUE, “hidden Markov