Eleven metabolism‑related genes composed of Stard5 predict prognosis and contribute to EMT phenotype in HCC

The mRNA expression profiles of HCC patients and the corresponding clinical profiles, including age, gender, grade, stage, alcohol consumption, Hepatitis B, Hepatitis C and survival time, were downloaded from the TCGA-LIHC database (https://​gdc.​nci.​nih.​gov/​) and were detailed in Table. 1. Validation dataset GSE14520 was downloaded from Gene Expression Omnibus (GEO) database, and survival information for the samples was shown in Table. 2.

A total of 145 epithelium (EPI) genes and 170 mesenchyme (MES) genes were obtained from PMID: 25214461 [13]. Based on the above gene sets, the samples’ EPI enrichment score and MES enrichment score were calculated by the R package GSVA (v1.34.0), and the EMT enrichment score was subtracted from the two. Surv_cutpoint of R package survminer (v0.4.8) was used to find the most appropriate node to differentiate EMT-H from EMT-L groups. Survival analysis of the two groups was performed by the R package surv (v3.2–7). Differences in clinical characteristics between the EMT subgroups were detected using Kruskal wallis test.

Enrichment scores for metabolic pathways were calculated for samples based on metabolic pathway-related genes provided by PMID: 33917859 [14]. The differences of metabolic enrichment scores between EMT groups were analyzed using Kruskal wallis test.

A subtype analysis related to metabolic reprogramming of HCC was performed based on genes of metabolic pathways, which had significant differences between the EMT-H and EMT-L groups. Unsupervised cluster analysis was applied to all samples in the TCGA-LIHC dataset through the R package ConsensusClusterPlus (v1.50.0) with the algorithm K-means. The clusters were then analyzed for survival using R package survival and survminer.

Differentially expressed genes (DEGs) of the clusters were acquired by the R package limma [15]. A threshold of |\({\mathrm{log}}_{2}foldchange\)|> 1 and adjusted p < 0.05, were considered for DEGs. Overlapping DEGs from the three clusters were used for subsequent analysis. The Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis and Gene Ontology term (GO) analysis, which consists of biological processes (BP), cellular component (CC), and molecular function (MF), were performed using DEGs shared by the three clusters [16].

Univariate Cox regression analysis was applied to the significant DEGs, using p < 0.01 as the threshold, in combination with the overall survival data. DEGs were then further filtered by LASSO-Cox regression analysis, and risk score models were constructed, a process that resorted to the R package glmnet (v4.0–2). Lambda screening was used for cross-validation. The model corresponding to lambda.min was used to collect the gene expression matrix. The risk score for each sample was calculated using the following equation: \({\mathbf{R}\mathbf{S}\mathbf{c}\mathbf{o}\mathbf{r}\mathbf{e}}_{{\varvec{i}}}= \sum_{{\varvec{j}}=1}^{{\varvec{n}}}{\mathbf{exp}}_{{\varvec{ji}}}\times {{\varvec{\upbeta}}}_{{\varvec{j}}}\). The median risk score was used to classify high-and low-risk groups. A p-value of Kaplan–Meier survival analysis < 0.05 was considered to indicate a significant difference between the two groups. The area under the curve (AUC) values for the model were calculated using the survival data and demonstrated by time-dependent receiver operating characteristic (ROC) curves, with AUC values greater than 0.6 indicating good predictive power of the prediction model.

To explore the response of patients to Erlotinib, Shikonin, Metformin, Bortezomib, Metformin, and Lapatinib, the predictive value of IC50 was obtained using the R package pRRophetic (v 0.5) analysis. The difference in IC50 between the high-and low-risk groups was tested using the Wilcoxon test. R package CIBERSORT (v1.03) was used to analyze the proportion of immune cells in all patients.

To validate the expression levels of genes in the prognostic model, we collected 80 tumors and adjacent normal tissues from patients with HCC. All patients participating in this study signed an informed consent form. This project was approved by the Human Research Ethics Committee of Zhongshan Hospital, Fudan University (Y2021-242). Tumors and adjacent normal tissues were then collected from patients who underwent surgical resection of the liver. All tissues were obtained immediately after surgical resection and frozen at − 80 °C. Huh7 cells, derived from the Chinese Academy of Sciences, were cultured in DMEM medium (D5796, sigma) containing 10% fetal bovine serum (16140071, Gibco) at 37 °C and 5% CO₂.

Total RNA was extracted from the tumors and adjacent normal tissues and reverse transcribed to cDNA using the Kit (EZBioscience, MN, USA). Then, quantitative PCR amplification was operated by a CFX384 real-time PCR machine (Bio-Rad, USA) using SYBR Green (Vazyme, China). Gene abundances were normalized to GAPDH. The primer sequences were shown in Table 3.

HCC and adjacent normal tissues were deparaffinized, rehydrated, blocked for endogenous peroxidase activity, antigen repair, and blocking, before being incubated overnight at 4 °C with primary antibodies against Stard5(ab178688, Abcam), N-cadherin, vimentin, E-cadherin and zo-1(9782 T, CST). The sections were then incubated with horseradish peroxidase-conjugated secondary antibodies for 1 h at room temperature and stained with 3, 3-diaminobenzidine tetrahydrochloride (DAB). Finally, cells were observed under a microscope.

The cDNA or shRNA (Genepharma, China) targeting Stard5 were recombined into lentiviral vectors to overexpress or knockdown Stard5, then transfected into 293 T cells. The mature infectious lentivirus was collected after 72 h. Stable Stard5-overexpressing and Stard5-knockdown Huh7 cell lines were constructed and verified by western blot.

Cells were lysed to extract total protein and heated to 100 °C for 20 min. Protein was added onto 8–12% SDS-PAGE electrophoreses and transferred to the PVDF membrane, then the blocked PVDF membrane was incubated with 1:1000 diluted Stard5 (ab178688, Abcam), β-Actin (3700 T, CST), N-cadherin, vimentin, E-cadherin and zo-1 (9782 T, CST) antibody at 4 °C overnight. After washing with TBST, the PVDF membrane was incubated with a 1:10000 diluted secondary antibody for 1.5 h at room temperature. Finally, a chemiluminescence analysis was performed.

The cells were inoculated in six-well plates at 1 × 10⁶ cells per well to form a dense monolayer after 12 h. Lines were drawn with a 200 μL tip to the cell layer to form straight cell wounds. After washing with PBS, the cells were incubated with serum-free medium at 37 °C for 48 h. The wound width was recorded at 0 and 48 h.

80 μl BD Matrigel mixture (diluted 1:10 with DMEM) was pre-coated in a transwell chamber (3513, Corning) at 37 °C overnight. Cells were diluted with serum-free DMEM and 4 × 10⁴ cells were added to the upper chamber. Then, 500 μl of DMEM containing 30% fetal bovine serum was added to the bottom chamber. After incubation at 37 °C for 48 h, the chambers were fixed in 4% paraformaldehyde for 2 h. Cells in the upper chamber were removed, then stained with crystal violet, washed with PBS and photographed under a microscope.

The study flowchart was shown in Fig. 1. First, we downloaded the expression data and clinical data of LIHC from the UCSC Xena database, removing the cases with missing survival information, and finally included 367 cancer samples. Expression matrix, which contains 16515 protein-coding genes, was subsequently used in the analysis. Based on the EMT score of each sample, we obtained 348 EMT-L samples and 19 EMT-H samples. Kaplan–Meier survival analysis showed that the overall survival (OS) of the EMT-L group was significantly longer than that of the EMT-H group (Fig. 2a). We then compared the clinical characteristics between the EMT subgroups. No significant differences in clinical characteristics were observed, possibly due to the small sample size of EMT-H samples (Additional file 1: Fig. S1a). Next, we obtained the genes of LIPID, NUCLEOTIDE, Carbohydrate, TCA, ENERGY, VITAMIN, and AMINOACID metabolic pathways and calculated the pathway enrichment scores for each sample. We found that LIPID, ENERGY, Carbohydrate, VITAMIN, and AMINOACID levels were significantly different between the EMT subgroups (Fig. 2b). It is implied that metabolic reprogramming occurred when tumors underwent EMT. The shared signature genes in the different metabolic pathways were also shown, with LIPID and VITAMIN sharing the most signature genes [26] (Additional file 1: Fig. S1b). Then, unsupervised cluster analysis based on the genes contained in these five metabolic pathways was performed (Fig. 2c). The most appropriate number of cluster was three (Fig. 2d, e). The Kaplan–Meier curves showed significant differences in survival among the three subtypes (Fig. 2f). Heat maps showed that enrichment scores for three critical metabolic pathways were also different among the three subtypes (Fig. 2h). We further investigated the differences in age, gender, grade, stage, alcohol consumption, Hepatitis B, Hepatitis C, and EMT subgroups among the three clusters, and found that the grade, stage, and EMT groups differed significantly (Fig. 2i).

We retrieved the DEGs among clusters, including 1739 DEGs between cluster1 and cluster2, 4779 DGEs between cluster1 and cluster3, 2519 DEGs between cluster2 and cluster3, and finally obtained 638 DEGs after intersection of the three clusters (Additional file 2: Fig. S2a). Functional enrichment analysis was then performed on the 638 DEGs. The pathways enriched by KEGG and BP were mainly focused on metabolic pathways related to catabolic or biosynthetic process (Fig. 3a, b). The pathways enriched for CC mainly focused on cell adhesion, such as collagen-containing extracellular matrix, apical part of cell, etc. (Fig. 3c). The pathways enriched for MF mainly involved metabolism-related enzyme activity (Fig. 3d).

We identified 150 genes that were significantly associated with overall survival (P < 0.01) by univariate Cox regression analysis from 638 DEGs (Additional file 3: Fig S3a), and top six genes showed the Kaplan–Meier curve (Additional file 3: Fig. S3b). Eleven genes were then selected as active covariates to evaluate the patients’ risk score using the Lasso Cox regression algorithm (Fig. 4a–c). The risk score was calculated using the following formula: risk score \(=STARD5\times (-0.0699)+FTCD\times (-0.0335)+SCN4A\times (-0.0281)+ADH4\times (-0.0134)+CFHR3\times (-0.0050)+CYP2C9\times (-0.0043)+CCL14\times (-0.0016)+GADD45G\times (-0.0013)+SOX11\times 0.0151+SCIN\times 0.0231+SLC2A1\times 0.0723\). The above equation shows that high levels of STARD5, FTCD, SCN4A, ADH4, CFHR3, CYP2C9, CCL14, and GADD45G were prognosis-protective factors associated with low risk. However, high expression of SOX11, SCIN, and SLC2A1 was associated with high risk. We divided the sample into high (n = 183) and low-risk groups (n = 184) using the median risk score as the cut-off point, and the expression of 11 genes was shown in the heat map (Fig. 4d). Kaplan–Meier curve showed that the high-risk group had a lower survival rate than low-risk group (Fig. 4e, g, h) (p < 0.0001). The prognostic model area under the curve (AUC) values of the time-dependent ROC curve were 0.77, 0.71, and 0.69 for 1 year, 3 year, and 5 year survival, respectively, which demonstrated that the multi-gene signature had better prognostic performance in predicting patient outcomes (Fig. 4f). Next, we performed risk stratification of patients according to their clinical characteristics. The results showed significant differences in risk scores between grade, stage, and EMT subgroups (Fig. 4i), which supported the accuracy of our risk model. Additionally, compared to the low-risk group, the number of plasma cells, Tregs, and macrophages M0 was significantly higher, while T cells CD4 + memory resting, NK cells activated, monocytes, mast cells resting were significantly lower (Fig. 4j). We also analyzed response to the drugs Erlotinib, Shikonin, Metformin, Bortezomib, Metformin, and Lapatinib, and found that the high-risk group was more likely to show resistance to the drugs (Fig. 4k). Hence, our risk model might be useful in predicting the response to immunotherapy and chemotherapy in patients with HCC.

We downloaded the GSE14520 dataset as a validation set, from which 221 samples with survival data were extracted. Risk scores were calculated according to the model of each patient. The expression of some genes in the prognostic model was shown in the heat map (Fig. 5a). Kaplan–Meier survival analysis showed that the high-risk group had poorer OS rate than low-risk group (p = 0.0035) (Fig. 5b, d, e). The prognostic model AUC values of the time-dependent ROC curve were 0.66, 0.68, and 0.67 for 1 year, 3 ear, and 5 year survival, respectively, indicating good predictive power of the prediction model (Fig. 5c). In addition, we also used the LIRI-JP database from ICGC data portal as a validation set, which also validated the stability of the model (Additional file 4: Fig. S4a–e).

To validate whether the expression of the 11 genes in HCC tissues was consistent with our model, we determined their expression in tumors and adjacent normal tissues of 28 HCC patients by qRT-PCR. The results suggested that the mRNA levels of STARD5, FTCD, SCN4A, ADH4, CFHR3, CYP2C9, CCL14, and GADD45G were downregulated (Fig. 6a–h), and SOX11, SCIN, and SLC2A1 were overexpressed in HCC tissues (Fig. 6i–k). Furthermore, the representative protein expression of nine genes in HCC and normal liver tissues was retrieved from the Human Protein Atlas (https://​www.​proteinatlas.​org), of which two genes were not found. Consistent with the qRT-PCR results, the protein level of STARD5, FTCD, ADH4, CYP2C9, CCL14, and GADD45G were low in HCC tissues (Additional file 5: Fig. S5a–f). SOX11, SCIN, and SLC2A1 were overexpressed in HCC tissues (Additional file 5: Fig. S5g–i).

Currently few studies are available on Stard5 in cancer, with one paper showing that hypermethylation of the STARD5 in clear cell renal cell carcinoma is significantly associated with poor prognosis [17]. Our prognostic model indicated that Stard5 deficiency in HCC was associated with poor prognosis, and we observed that STARD5 mRNA was markedly decreased in 82% (23/28) of HCC tumors compared to the corresponding adjacent normal liver tissue (Fig. 6a). Western blot analysis of six randomly selected pairs of HCC samples showed that Stard5 expression was significantly reduced in tumors (Fig. 7a). Furthermore, in 80 HCC patients, Stard5 expression was scored as moderately positive in 25%, while 36.25% were strongly positive, compared to 28.75% moderately positive and 52.5% strongly positive in adjacent normal liver tissue (Fig. 7b, c). Next, patients were divided into low (negative and weakly positive) and high (moderately and strongly positive) expression groups based on Stard5 expression in the tumor tissue. Kaplan–Meier survival analysis showed that low Stard5 expression was associated with significantly poorer TTR (Fig. 7d) and shorter OS (Fig. 7e) than high Stard5 expression. Together, these data suggest that reduced Stard5 expression in tumor tissues may be an important indicator of poor prognosis in HCC.

Stard5 was derived from our EMT-related prognostic model, so we first examined the impact of Stard5 on EMT. Western blot analysis showed that knockdown of Stard5 increased the expression of mesenchymal markers (N-cadherin and Vimentin) and decreased the expression of epithelial markers (E-cadherin and zo-1) in Huh 7 cells (Fig. 8a). EMT is an important step before cell invasion and migration, and we subsequently explored the role of stard5 in the invasion and migration of Huh 7 cells. In the wound-healing and transwell assays, the migration and invasion abilities decreased in shSTARD5 cell line (Fig. 8b–d). Furthermore, overexpression of Stard5 in Huh 7 cells decreased the levels of Vimentin and N-cadherin and increased those of E-cadherin and zo-1 (Fig. 9a). Upregulation of Stard5 inhibited cell invasion and migration (Fig. 9b–d). We then detected N-cadherin, Vimentin, E-cadherin, zo-1 as well as Stard5 protein in 15 human HCC samples. Stard5 expression was scored and classified into stard5 low- and high- expression groups according to the median. The expression of mesenchymal markers (N-cadherin and Vimentin) was negatively associated with Stard5 expression, while epithelial markers (E-cadherin and zo-1) were positively associated with Stard5 expression (Fig. 10a). These data suggest that Stard5 deficiency induces EMT, resulting in HCC metastasis.