Identification of fatty acids synthesis and metabolism-related gene signature and prediction of prognostic model in hepatocellular carcinoma

A total of 664 datasets of HCC samples were supplied by two public databases. RNA sequencing data underwent variance-stabilizing transformation (VST) using the DESeq2 package in R. These datasets included 424 samples were from TCGA-LIHC cohort (https://​portal.​gdc.​cancer.​gov/​) [21] and 240 samples were from ICGC-LIRI-JP cohort (https://​dcc.​icgc.​org/​) [22]. Importantly, the TPM matrix normalized by counts in TCGA and ICGC was used for subsequent analysis. And the “log2” was used to reduce the TPM matrix variability to make the TPM matrix closer to the normal distribution. The clinical information of samples such as age, gender, race, treatment type, stage, and status were also obtained from TCGA and ICGC (Additional file 2: Table S1 and Additional file 3: Table S2). Additionally, the single cell RNA sequencing dataset of GSE125449 was from the GEO [23]. We established a gene set by obtaining FASM-related genes from the molecular signatures database (MSigDB) database (https://​www.​gsea-msigdb.​org/​gsea/​msigdb/​index.​jsp) (Additional file 4: Table S3) [24].

To identify differential expression of FASM-related genes between precancerous and cancerous tissues, we utilized the DESeq2 package in R with thresholds for significance set at a false discovery rate < 0.05 and | Log2 fold change |> 1 [25]. Moreover, we identified FASM-associated genes that significantly influenced overall survival (OS) in HCC through univariate Cox regression analysis using the Survival package. The maftools package were employed to describe somatic mutations of these genes in HCC patients [26].

For unsupervised clustering, we employed the ConsensusClusterPlus package [27], selecting an optimal number of subtypes based on proportion of ambiguous clustering (PAC). Principal component analysis (PCA) and t-Distributed Stochastic Neighborhood Embedding (tSNE) methods were carried out to compare the expression levels among different FASM subtypes. In order to assess survival outcomes within clusters derived from the TCGA-LIHC and ICGC-LIRI-JP cohorts, survminer and survival packages were used to draw Kaplan–Meier survival curves and conduct log-rank tests.

Based on 26 FASM associated genes, the reliability factors of LASSO regression analysis were executed by using multivariate Cox regression method [28]. Patients with LIHC were distinguished into high- or low-risk groups according to the polygenic risk score obtained by the prognostic features. In addition, area under curve (AUC) score of the receiver operating characteristic (ROC) curve was used for evaluating the prediction ability of prognostic signatures. The R packages survival, rms and timeROC were performed to establish and verify the prognostic model of FASM associated genes.

We employed single-sample gene set enrichment analysis (ssGSEA) to measure the relative abundance of 28 immune cell types associated with immune response [29, 30]. The ssGSEA algorithm enabled us to express the relative abundance of each immune cell as an enrichment score normalized to a ranging from 0 to 1. This approach was utilized to evaluate the LIHC cohort, and explore the heterogeneity of TME among different FASM subtypes in LIHC. To evaluate the immune cell status in LIHC patients, we utilized CIBERSORT, a deconvolution method that utilizes gene expression profiles [31, 32]. Additionally, we investigated the distribution of 22 immune cell types using CIBERSORT and examined their infiltration levels in high- and low-risk populations. Furthermore, we explored the correlation between risk scores and immunologic function. For simulating tumor immune escape mechanisms and predicting potential response to tumor immunotherapy, we adopted TIDE algorithm [33].

In order to explore the differences in biological processes among different FASM patterns, Gene set variation analysis (GSVA) was carried out using the GSVA package [34]. Moreover, The Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways related to different patterns were identified using clusterProfiler R package with a false discovery rate (FDR) cutoff below 0.01 [35].

The transcriptional mRNAsi index for each LIHC sample was computed using One-class logistic regression (OCLR), which is based on pluripotent stem cell samples and strongly correlated with stem cell features. This index can be applied for cancer stemness predictions [36]. Both prognostic value of mRNAsi as well as the Spearman correlation between FASM subtypes and mRNAsi indices across all 363 patients with LIHC were analyzed.

STRING database (https://​cn.​string-db.​org/​), a functional protein-related network, assembles all known and predicted proteins [37]. The PPI network interactions file with a medium confidence score (> 0.4) was available. Furthermore, cytoscape software (version 3.8.2), a common open source, was used to beautify and analysis interaction network [38]. To explore hub genes, the cytoHubba-MCC was used [39].

Normalized data were integrated using the "Findinintegrationanchors" function in the Seurat software package [40]. Then, the data were reduced dimensionality and performed PCA. “FindNeighbors” and “FindCluster” were performed to analysis LIHC cell clusters. The tSNE was utilized to visualize the cell clusters. The cell cluster markers were acquired by screening the literature and retrieving the CellMarker 2.0 database (http://​bio-bigdata.​hrbmu.​edu.​cn/​CellMarker/​) to annotate cell type [41]. Based on scRNA-seq data, the R package CellChat can infer, visualize, and analyze of intercellular communication and describe the interactions among ligands, receptors and secreted factors [42]. To investigate the possible communication between CSCs and other cells, the ligand-receptor and secretion interactions between cell types were analyzed by using Cellchat.

L02, HepG2, HCCLM3 cell lines were supplied by American Type Culture Collection or Cell Bank, Shanghai Institutes for Biological Sciences, Chinese Academy of Sciences. These cells were authenticated using Short Tandem Repeat (STR) analysis. L02 cells were cultured in RPMI 1640 medium (10% FBS and 1% streptomycin/penicillin) at 37 °C under 5% CO₂. HepG2, HCCLM3 cells were cultured in DMEM media at same growth conditions. The colony formation ability of HCC cell lines was detected by plate cloning experiment. 500 cells were colonized into 6-well plates with 20% FBS medium. After culturing for 7 days under the condition of 37 °C and 5% CO₂, colonies were terminated in 4% paraformaldehyde (PFA) and dyed with crystal violet for 2 min. Then the colonies were captured by microscopy. The cell counting kit-8 (CCK-8) was utilized to detect cell proliferation. 10000 cells were inoculated in 96-well plates with condition medium 48 h. The medium absorbance was examined using microplate reader. HCC cells were plated into 6-well plates with low adhesion (500 cells per well). Then we added 2 ml DMEM/F-12 medium containing 1% B-27, 1% N-2, 20 ng/mL epidermal growth factor, and 10 ng/ml basic fibroblast growth factor into 6-well plates. Replace the medium every 3 days. The stem cell spheres were captured by microscopy.

Tumor tissues from 5 patients with HCC in Chongqing University Cancer Hospital were collected. And the clinical characteristics of 5 samples in Supplementary Materials (Additional file 5: Table S4). 4% PFA was used to fix HCC tissue for 24 h, the paraffin embedding was processed according to standard procedures. The embedded tissues were cut to a thickness of 8 μm and placed on a slide. After deparaffinizing, tissues were performed antigen repair in citrate buffer in microwave for 15 min at least. Following manufacturer’s instructions, the treated sections with anti-ACACA antibody were incubated overnight at 4 °C, washed thrice with phosphate-buffered saline (PBS) for 5 min each, and reacted with the appropriate concentration of secondary antibodies for 1 h at 25 ℃. Immunohistochemical microscopic images were obtained utilizing optical microscope.

To prepare the cells for imaging, cell and climbing placed in 24-well plates and washed twice with PBS before being fixed in 4% PFA for a duration of 15 min. Subsequently, the climbing with cells underwent three washed with PBS and were permeabilized using a solution containing 0.3% Triton-X100/PBS at 25 ℃ for 15 min, followed by another round of triple washing with PBS. For blocking purposes, the cells were treated with a solution consisting of 5% bovine serum albumin (BSA) at 25 ℃ for 1 h after which they were incubated overnight at 4 °C in a mixture containing anti-ACACA antibody. Following washed with PBS to remove any excess primary antibody, secondary antibody was applied to the cells and allowed to incubate for an hour at room temperature. After another round of triple washing with PBS, phalloidin-iFluor staining was carried out on the cells for half an hour and DAPI staining at room temperature for ten minutes.

The total RNA from cells was extracted using TRIzol. The transcriptional level of ACACA was detected via qRT-PCR with following program: 95 ℃ for 30 s, 40 cycles of 95 °C for 5 s, and 60 °C for 30 s. The reaction system consisted of 5 μL SYBR Green (Foregene, Chengdu, China), 1 μL each primer, 1 μL cDNA, and 2 μL diethylpyrocarbonate (DEPC) water.

ACACA forward primer: 5′ - AGGAGCTGTCTATTCGGGGT - 3′,

reverse primer:5′ - ATGTCTGGCTTGCACCTAGTA - 3′;

and GAPDH forward primer: 5′ - GGTATGACAACGAATTTGGC - 3′,

reverse primer: 5′ - GAGCACAGGGTACTTTATTG - 3′.

Unpaired student’s t test was applied to analysis two sets of experiments, one-way ANOVA was utilized for 3 or more groups. Data was analyzed using GraphPad Prism software 8.0 or R version 4.3.0. Each experiment was repeated three times, unless otherwise specified in the figure legends.

The summary clinical characteristics of 664 HCC samples was presented in Table 1. Comprehensive analysis of FASM genes was performed based on multi-group data of TCGA-LIHC cohort (Additional file 4: Table S3). Differential gene analysis from transcriptome data revealed that 73 of 203 FASM genes were upregulated or downregulated in HCC (Fig. 1A; Additional file 1: Fig S1A). Univariate Cox regression results suggested that 26 out of 73 FASM differential expression genes were associated with prognosis in HCC (Fig. 1B). The expression of 26 FASM genes displayed particularly significant heterogeneity between normal and HCC tissues (Fig. 1C). The mutational landscape for the 26 FASM genes was displayed in a waterfall plot. Eight out of the 26 FASM genes owned a high mutation frequency and were closely relevant to progression or recrudesce in HCC (Fig. 1D). Moreover, the kyoto encyclopedia of genes and genomes (KEGG) result showed that 26 FASM genes were concentrated in the arachidonic acid metabolism, fatty acid metabolism and PPAR signaling pathway (Fig. 1E). Thus, these results suggest that the changes of FASM genes expression level may regulate the development and progression of HCC.

According to the RNA-seq data of 26 FASM genes and clinical data from TCGA-LIHC, unsupervised clustering was performed by using ConsensusClusterPlus package to classify LIHC patients into two different clusters (Clusters1 and Clusters2) (Fig. 2A). The differences between the two FASM models were firstly evaluated by PCA algorithm and tSNE algorithm. It was observed that there were obvious differences in transcriptional profile among the FASM clusters (Fig. 2B; Additional file 1: Fig S2A). We then assessed the clinical prognostic value of FASM patterns in HCC patients via a survival analysis. Patients corresponding to two clusters exhibited a significant variation in survival from TCGA dataset (Fig. 2C). Based on the data of TCGA-LIHC, the expression of 26 FASM genes was prominently different between the two clusters (Fig. 2D; Additional file 1: Fig. S2B). To explore the stability and applicability of two FASM patterns in HCC, the unsupervised clustering analysis was repeated by using LIRI-JP cohort from ICGC. The result showed that populations could be well divided into two categories (Fig. 2E). The PCA and tSNE analysis results confirmed the two disparate patterns of FASM in HCC (Fig. 2F; Additional file 1: Fig. S2C). And there was also an obvious difference in survival from ICGC dataset between the patients of two clusters (Fig. 2G). Based on ICGC-LIRI expression profiling data, 26 FASM genes in the two clusters were greatly differentially expressed (Fig. 2H; Additional file 1: Fig. S2D).

To explore the immune characteristics in FASM patterns, we estimated the immune infiltration characteristics of 28 immune cell proportions in FASM patterns and made visual analysis via heatmap (Fig. 3A). In comparison, some kinds of tumor immunosuppression cells were more abundant in Cluster1, including regulatory T cells and macrophages, indicating that the patients of Cluster1 subtype presented an immunosuppressive microenvironment that promoted tumor progression. We then explored enrichment fraction of the immune cells and their immune pathway activity. These two models showed entirely different immune properties. As in TCGA, regulatory T cells and M0 macrophages were mainly concentrated in Cluster1, while activated NK cells and proinflammatory M1 macrophages mostly concentrated in Cluster2 (Additional file 1: Fig. S3A). Consistently, similar infiltrative features of regulatory T cells and M0 macrophages in Cluster1 were also observed in ICGC (Additional file 1: Fig. S3B). In immune pathway activity, Cluster1 possessed higher CCR, Check-point and T cell co-inhibition activity. Whereas, Type I and II IFN Reponses were mainly enriched in Cluster2 (Fig. 3B). At last, we evaluated the sensitivity of FASM patterns to immunotherapy. The results displayed that Cluster1 had enhanced TIDE, indicating that Cluster1 had a higher probability of immune escape and a poor response to immunotherapy (Fig. 3C). Besides, the biological functions of FASM Pattern were further investigated. We assessed the enrichment scores of the Hallmark pathways and KEGG pathways in the FASM cluster subtypes by GSVA. The two subtypes exhibited different pathways, including pathogenic escherichia coli infection, lysine degradation and adipocytokine signaling pathway (Additional file 1: Fig. S3C).

Fatty acid synthase expression occurs predominantly in proliferating fetal cells [43], implying that reactivation of FA synthesis in cancer cells may be a signal to return a stemness maintenance status. By the OCLR algorithm, mRNAsi of each LIHC patients was estimated on basis of the gene expression profiles, and then the correlations between mRNAsi and the FASM subtypes was detected (Fig. 4A). In addition, Kaplan–Meier analysis showed that LIHC patients with high-mRNAsi experienced shorter OS than low-mRNAsi patients (Fig. 4B). By sequencing mRNAsi from low (left) to high (right), we found that FASM Cluster1 was largely enriched in high mRNAsi regions, while FASM Cluster2 presented the lowest mRNAsi, which was demonstrated by the comparative analysis (Fig. 4C), and the distribution of these patients in FASM patterns were displayed in Sankey diagram (Fig. 4D).

At present, surgery and systemic chemotherapy are the main clinical treatment options for HCC patients. Therefore, we calculated half maximal inhibitory concentration (IC50) values for chemotherapy sensitivity evaluations of several chemotherapeutics drugs by pRRophetic algorithm and compared IC50 values among FASM patterns. The results showed that estimated IC50 values of Sorafenib, Cytarabine, Docetaxel, Cisplatin, Doxorubicin, Etoposide, Nilotinib and Paclitaxel were substantially lower in Cluster2, suggesting that Cluster2 might be more sensitive to these drugs. On the contrary, Cluster1 were more tolerant to these drugs (Fig. 5A). The study of chemotherapy sensitivity among FASM patterns provided a certain theoretical basis for clinical medication.

Based on 26 FASM genes, Lasso analysis and multivariable Cox regression analysis were performed. As a result, 11 candidates (EIF6, ELOVL3, LPL, CYP2C9, PRXL2B, ELOVL1, ACACA, CYP7A1, DEGS1, PTGES3, AKR1C3) were marked as independent prognostic genes (Fig. 6A, B). The ROC curves displayed that the AUC of the risk cohort was > 0.7 at 1, 3, 5 years (Fig. 6C; Additional file 1: Fig. S4A). LIHC patients were divided into low- or high-risk groups, and the prognosis of patients in the high-risk group was poor. (Fig. 6D, E). Importantly, Cluster1 exhibited a higher risk score compared to Cluster2 from TCGA and ICGC (Fig. 6F; Additional file 1: Fig. S4B), indicating that FASM subtypes had a good consistency among the risk prognostic model of FASM. To establish a more practical and reliable nomogram, some familiar clinicopathological indicators were introduced (gender, race, age and tumor stage). Univariate Cox and Multivariate Cox analysis revealed that FASM riskScore was a risk factor for the prognosis of HCC patients (Fig. 6G, H). In view of the high correlation between FASM riskScore and prognosis of HCC patients, we integrated conventional clinicopathological parameters with FASM riskScore to establish a new nomogram (Fig. 6I). We first used the ROC curve to assess the AUC values of various indicators to predict the prognosis of HCC patients, and the results showed that the FASM riskScore was greatly better than other clinicopathologic characteristics. And this nomogram further improved the accuracy of predicting the prognosis of HCC patients (Additional file 1: Fig. S4C). Furthermore, the calibration curves further illustrated the good accuracy and robustness of nomogram in evaluating patient survival at 1, 3, and 5 years (Fig. 6J). Interestingly, the FASM high-risk group exhibited the same characteristics of immune infiltration and drug tolerance as Cluster1 (Additional file 1: Fig. S5A–C; Additional file 1: Fig. S6A). Those data indicated that 11 prognostic signatures could serve as reliable predictors of OS in LIHC patients.

To identify the key regulatory genes of FASM, we evaluated the correlations among FASM risk genes and found obvious synergistic effects (Fig. 7A). Whereafter, a PPI network was established using the STRING database, and 3 genes with a degree > 2 by the cytoHubba-MCC plugin were identified (Fig. 7B, C). The expression of CYP7A1, ELOVL1 and ACACA were higher in LIHC tumor tissues than normal tissues (Fig. 7D). In survival analysis, ACACA with high expression group had the worst prognosis (Fig. 7E). Those findings indicated that ACACA was the key regulatory gene of FASM in HCC for further study.

We further investigated the pathological mechanisms of ACACA in HCC progression. The differential expression of ACACA in FASM risk and cluster subtypes were calculated. The expression of ACACA in Cluster1 and FASM high-risk group was notably higher than Cluster2 and FASM low-risk group (Fig. 8A, B). ACACA also showed differential expression in different stages of tumors, especially in stage III and T3 stage (Fig. 8C). ACACA had the most significant positive correlation with M0 macrophages. Meanwhile, ACACA was negatively correlated with CD8 T cells (Fig. 8D). Similarly, LIHC patients with high ACACA expression had lower responsiveness to CTL4 and PD1 (Fig. 8E). Likewi se, ACACA was significantly positively correlated with HCC stemness and metastasis (Fig. 8F, H). Finally, we detected the expression of ACACA in scRNA sequencing of patient with HCC from GSE125449. By removing low-quality cells, normalizat ion, integration, and PCA, we divided the cells into 10 clusters (Additional file 1: Fig. S7A), which were then annotated for analysis. Finally, 7 cell types, including T cells (clusters 0, 8), Cancer stem cells (CSCs) (clusters 1, 7), Fibroblasts (cluster 2), Endothelial cells (cluster 3, 9), Macrophages (cluster 4), B cells (clusters 5) and Malignancies (clusters 6) (Fig. 8I; Additional file 1: Fig. S7B) were identified. The results indicated that the expression level of ACACA in CSCs was much higher than that in malignant cells (Fig. 8J).

The communication properties of CSCs with highly expressing ACACA and other cell types were also investigated. We used Cellchat to test Secreted Signaling, Cell–Cell Contact and ECM-Receptor interactions between different cells (Fig. 8K–M). The results suggested that Secreted Signaling-mediated cellular interactions mainly existed in the MIF signaling pathway (MIF-CD74 + CXCR4) (Additional file 1: Fig. S7C–D). APP-CD74 pathway contributed the most to Cell–Cell Contact (Additional file 1: Fig. S7E–F). In addition, VTN-(ITGAV-ITGB1) pathway was dominant in ECM-Receptor (Additional file 1: Fig. S7G, H). The above results demonstrated that ACACA may motivate the stemness of HCC through the fatty acid pathway, thereby leading to immune escape and tumor metastasis through CD74 and ITGB1 signaling.

To gain a more comprehensive understanding of ACACA's role in cancer, the TCGA and Genotype-Tissue Expression (GTEx) databases were used to estimate the expression level of ACACA in pan-cancer compared with normal tissues. ACACA is highly expressed in most cancers, including CESC, CHOL, COAD, ESCA, LIHC, PRAD, READ, STAD, UCEC. On the contrary, low ACACA expression was found in GBM, KICH, KIRC and THYM (Additional file 1: Fig. S8A). To further understand the effect of ACACA on patient prognosis, univariate Cox regression was employed to analyze the prognosis across 33 cancer types. The forest plot revealed that the downregulation of the ACACA expression was significantly related to the prolongation of OS time in ACC (HR = 2.79[95% CI 1.26–6.21], p = 0.012), BLCA (HR = 1.36[95% CI 1.01–1.82], p = 0.041), KICH (HR = 9.63[95% CI 1.2–77.57], p = 0.033), LIHC(HR = 1.65[95% CI 1.16–2.33], p = 0.005), MESO (HR = 1.84[95% CI 1.14–2.97], p = 0.012) and UVM (HR = 2.92[95% CI 1.14–7.43], p = 0.025) (Additional file 1: Fig. S8B). Then genomic variant analysis of ACACA showed that ACACA variants are ubiquitous in pan-cancer. Meanwhile, the maximum frequency of ACACA mutation (> 10%) were found in SKCM and UCEC patients, but the frequency of ACACA changes in KICH and MESO patients were minimum (Additional file 1: Fig. S8C).

Next, the types and loci of ACACA alteration were also validated (Additional file 1: Fig. S8D). To further comprehend the role of ACACA mutation in pan-cancer, we assessed the association of ACACA expression with tumor mutation burden (TMB) and microsatellite instability (MSI). To be expected, ACACA expression was positively related to TMB in most cancers, including LIHC, but it showed a negative correlation with BRCA and ESCA (Additional file 1: Fig. S8E). When we compared the relationship of ACACA positive expression and MSI in most cancers, the noted result was positively correlated with GBM, KIRC, LIHC, LUAD, LUSC, STAD, UCEC and CESC, and negatively correlated with BRCA, DLBC, HNSC and THCA (Additional file 1: Fig. S8F). All the above, these results suggested that ACACA was associated with tumorigenesis and tumor immunity.

To verify the important role of ACACA in HCC, we used the HCC patient samples and SD rat orthotopic liver cancer models. Immunohistochemical staining displayed that ACACA expression was higher in HCC compared to normal tissues (Fig. 9A, B). Analogously, qRT-PCR assay and immunofluorescence staining revealed the expression level of ACACA was obviously higher in HCC cell lines (HepG2 and HCCLM3) than in liver cell (L02) (Fig. 9C, D). Particularly, the expression of ACACA was significantly up-regulated in induced liver cancer stem cells (M3CSs and G2CSs) compared to their parent cells (Fig. 9E). To better understand the function of ACACA in vitro, we analyzed the oncogenic phenotype of HCCLM3 and HepG2 cells with ACACA knockdown (shACACA). CCK8, Edu and colony formation assays manifested that the depletion of ACACA impaired HCC cell proliferation (Fig. 9F–H). Sphere formation assay and wound healing were carried out to evaluate the stemness traits and metastatic potential of ACACA. The results suggested that the sphere sizes and migration distance were noticeably shortened after shACACA transfections in HCC cells, exhibiting a suppression of stemness properties and metastatic potential in HCC cells upon ACACA depletion (Fig. 9I, J).