Prediction of hepatocellular carcinoma prognosis and immunotherapeutic effects based on tryptophan metabolism-related genes

The available RNA-seq data and follow-up data from patients with HCC were downloaded from The Cancer Genome Atlas Liver Hepatocellular Carcinoma (TCGA-LIHC) dataset, which contains data from 360 HCC tissues and 50 adjacent nontumourous liver tissues. For validation purposes, the gene expression profiles and clinical data from the GSE14520 (including 242 HCC tissues) and GSE76427 (including 115 HCC tissues) cohorts were obtained from the Gene Expression Omnibus (GEO) database. HCCDB18 (including 389 HCC tissues) data were downloaded from the Hepatocellular Carcinoma Database (HCCDB). In this study, we used the TCGA-LIHC dataset as the training set and the GSE14520, GSE76427, and HCCDB18 datasets as independent validation sets.

HCC cell lines Hep-G2, Huh-7, Hep-3B, and SK-Hep-1 and normal liver cells LO2 (provide by the China Center for Type Culture Collection) were maintained in DMEM supplemented with 10% fetal bovine serum (FBS) and 1% penicillin/streptomycin (Gibco, USA), which were maintained at 37℃ in a 5% CO2 incubator. The method of qRT-PCR has been described in our previously study [17]. GAPDH was used as control for TPH1. The information of primers sequences in this study were shown in Additional file 5: Table S1.

We extracted the tryptophan metabolism-related genes involved in the tryptophan metabolism pathway "KEGG TRYPTOPHAN METABOLISM" from The Molecular Signatures Database (MSigDB) database (http://​software.​broadinstitute.​org/​gsea/​msigdb/​index.​jsp) [18].

We conducted a series of steps to preprocess the TCGA data, including removing the samples without survival time, follow-up data and status. The Ensembl data were converted to gene symbols, and the expression values obtained with multiple Gene Symbols were set at the median value. For the GEO dataset, we downloaded the annotation information of the corresponding chip platform, mapped probes to genes according to the annotation information, and removed probes that matched one probe to multiple genes. When multiple probes matched a gene, the median value was taken as the gene expression value.

The ConsensusClusterPlus package was used to establish a consistency matrix and cluster HCC patients into distinct subgroups based on the expression data of genes related to tryptophan metabolism [19]. Then, we screened the differentially expressed genes (DEGs) between the two clusters, among which the genes correlated with prognosis were further analyzed (|logfc|> 1 & p < 0.05). Next, we performed lasso Cox regression to compress the core genes that were used for establishing the risk model. The formula used was RiskScore = Σβi × Expi), where i represents the gene expression level, and β is the Cox regression coefficient of the corresponding gene.

To detect the biological signaling pathway, we performed GSEA in distinct clusters built based on tryptophan metabolism-related genes [20]. Here, we performed GSEA against the background of c2.cp.kegg.v7.0.symbols.gmt. Moreover, to observe the relationship between the risk score of patients and their biological functions, we selected the gene expression profiles of each HCC patient to perform ssGSEA based on the R software package. We calculated the scores of each patient in different functions to obtain the ssGSEA enrichment score.

CIBERSORT is a useful method for characterizing the hematopoietic cell composition of tissues based on RNA transcript profiles [21]. It is usually employed to analyze large-scale gene expression data to develop cellular biomarkers and therapeutic targets [22]. Here, we used the CIBERSORT method to calculate the relative abundance of 22 primary immune cells in distinct subgroups.

Statistical analysis of the data was performed using GraphPad Prism 8 (GraphPad Software Inc., San Diego, CA, USA). Student’s t test of variance was used to calculate the differences between two groups. The survival R package (Version 2.43-3) was adopted to analyze survival rate. p < 0.05 was regarded as statistically significant.

40 Trp metabolism-related genes were involved in the production of Trp metabolites [23]. Therefore, we compared the mRNA expression levels of Trp metabolism-related genes between HCC tissues and liver tissues in the TCGA dataset. The expression levels of most Trp metabolism-related genes, such as OGDHL, ALDH1B1, CYP1A2, GCDH, ACAT1, TDO2, ALDH2, HADH, ALDH9A1, KYNU, KMO, CYP1A1, INMT, MAOA, ECHS1, IDO2, ACMSD, AOX1, EHHADH, MAOB, HAAO, AOC1, AADAT, ACAT2, and CAT, were downregulated in HCC tissues than in paracarcinoma tissues (Fig. 1A). To deeply understand the role of Trp metabolism in HCC development and prognosis, we explored the metabolic phenotypes based on Trp metabolism-related genes. First, we conducted a univariate Cox regression analysis to explore the association of the 40 genes with the overall survival of HCC patients collected from the TCGA-LIHC dataset. The expression levels of 8 of 40 genes were significantly correlated with the prognosis of HCC (p < 0.05) (Fig. 1B). As a potential risk factor for TPH1, its expression in HCC cells and normal liver cell were tested by qRT-PCR. And results showed that the expression level of TPH1 was higher in Hep-G2, Huh-7, and Hep-3B compared to LO2 (Fig. 1C). In addition, we observed that there was a significant correlation among the levels of Trp metabolism-related genes using correlation analysis (Fig. 1D). Then, we conducted consistent clustering analysis based on the expression profiles of 8 prognostic genes of Trp metabolism to classify patients. When the number of clusters was selected as 2, we obtained two distinct metabolic phenotypes (Fig. 1E–G). We further analyzed the prognostic characteristics of the two metabolic phenotypes, and we found that patients in C2 had a better prognosis than that of patients in C1. Consistently, we observed similar prognostic features of the two clusters in the GSE14520 dataset (Fig. 1H). In addition, we calculated the "tryptophan metabolism scores" of each patient using the method of ssGSEA in the TCGA and GSE14520 cohort, and we found that the score of C2 with good prognosis was significantly higher than that of C1 (Fig. 1I). These results demonstrated that the Trp metabolism-related clusters demonstrated remarkable differences in the overall survival of HCC patients.

To further analyze the relationship between different metabolic phenotypes and clinicopathological features in the TCGA-LIHC cohort, we compared the distribution of different clinical features in the two metabolic phenotypes. Patients in C1 tended to have higher T stage, stage, grade, and death probability than those in C2 (Additional file 1: Figure S1). What causes the distinct clinical features between the two metabolic subtypes remained unknown. Therefore, we analyzed differences in genomic alterations between these two metabolic phenotypes. To this end, we obtained information on the molecular features of the TCGA-LIHC dataset from a previous pancancer study [24]. In this published study, the authors classified HCC patients into 5 immune subtypes, of which subtypes 3 and 6 had the best prognosis, and immune subtype 1 had the worst prognosis (Additional file 2: Figure S2A). Here, we found that C1 showed a higher aneuploidy score, more homologous recombination defects, and different fractions than those of C2 (Additional file 2: Figure S2B). Furthermore, we found that immune subtype 3 among the immune molecular subtypes was more prevalent in the patients of the C2 metabolic subtype who had good prognoses, while immune subtype 1 was more prevalent in the patients of the C2 subtype who had poor prognoses (Additional file 2: Figure S2C). Furthermore, we explored the differences in gene mutations between different metabolic phenotypes, and the top 10 genes with significant differences are shown in Additional file 2: Figure S2D. We found that the mutation frequencies of various genes, such as those in TP53, TTN and MUC16, were significantly different between the two metabolic phenotypes.

The immune microenvironment plays a key role in the development of HCC and the response to immune checkpoint blockers (ICBs) [8, 25, 26]. Therefore, we elucidated the immune microenvironment of patients with different metabolic phenotypes. In the TCGA-LIHC cohort, we assessed the extent of immune cell infiltration in patients by calculating the expression levels of genes in immune cells via CIBERSORT. Many types of immune cells, including CD4 + memory-activated T cells, monocytes, and M1 macrophages, were more abundant in C2 than in C1 (Fig. 2A). The abundance of infiltrating immunosuppressive cells, including that of regulator T cells, for example, was higher in C1 than in C2 (Fig. 2A). Moreover, we detected whether there were differences in the response to immunotherapy between different metabolic phenotypes. First, we explored the immune checkpoint genes expression levels, and the results showed that the expression levels of most of the immune checkpoint genes, such as CTLA4, IDO1, TNFSF18, TNFSF4, TNFSF9, and NRP1, were higher in C1 than in C2 (Fig. 2B). Then, we compared the differences in the potential effects of immunotherapy between different metabolic phenotypes by Tracking of Indels by Decomposition (TIDE) software. The higher the TIDE prediction score was, the higher the possibility of immune escape and the lower the possibility of patients benefiting from immune therapy (Fig. 2C). As expected, the TIDE score in C1 was higher than that in C2 in the TCGA cohort, suggesting that C1 patients have a higher possibility of exhibiting immune escape. All the results suggested that the metabolic subtypes might be used for assessing the immune microenvironment and predicting the response to immunotherapy for patients with HCC.

To further explore the difference in the underlying regulatory mechanisms of C1 and C2, we employed GSEA in the TCGA cohort. Results showed that some pathways, including the Notch signaling pathway and the pathway related to pathogenic Escherichia coli infection, were enriched in C1, while more pathways, such as tryptophan metabolism, peroxisome, propanoate metabolism, retinol metabolism, valine, leucine and isoleucine degradation pathways, were enriched in C2 (Additional file 3: Figure S3A). We also analyzed the differences in the enrichment of 10 oncogenic pathways (WNT, TP53, RAS, TGF-beta, PI3K, NRF1, NOTCH, MYC, HIPPO, and cell cycle) in the two metabolic subtypes in a previous study [27]. We observed significant differences in most types of pathways except TGF-beta and MYC pathways (Additional file 3: Figure S3B).

In the above analysis, we identified two distinct metabolic subtypes based on the genes associated with Trp metabolism. Next, we used the limma package to calculate the differentially expressed genes (DEGs) between C1 and C2 (FDR < 0.05 and |log2FC|> 1). Finally, we screened 739 DEGs between the clusters. All the DEGs are shown in the volcano plot (Fig. 3A). The univariate Cox regression model identified 189 genes that were markedly associated with prognosis (Fig. 3B). Among 189 genes, 95 genes were risk genes. Then, lasso Cox analysis was adopted to further compress the number of key genes (Fig. 3C). When lambda = 0.066, the model reaches the optimum (Fig. 3D). Therefore, we chose the 10 identified genes (cyclin-dependent kinases (CDK1), TROAP, G6PD, MMP1, BAIAP2L2, PTTG1, LCAT, CYP2C9, CFHR3, and SLC22A10) for which lambda = 0.066 for further study.

Based on the ten-gene model, the AUCs for 1-, 3-, and 5-year overall survival were 0.72, 0.65, and 0.74, respectively (Fig. 3E). Additionally, we found that patients in the group with low risk scores tended to have significantly longer overall survival than those in the group with high risk scores based on TCGA cohort (Fig. 3E). Furthermore, similar AUCs and prognostic differences were observed in the GSE76427 cohort (Fig. 3F).

To examine the relationship between the risk score and clinical features of HCC, we analyzed the difference in risk score between patients with different TNM grades and stages. As the clinical grade increased, the risk score increased, which indicates that patients with higher clinical grades had higher risk scores (Additional file 4: Figure S4A). Additionally, patients in the high-risk-score groups had a higher T stage, stage, and grade than those in the low-risk group (Additional file 4: Figure S4B). This evidence demonstrated that the risk score could function as a biomarker for prognostic prediction and may provide clues for developing precise treatment strategies.

To further clarify the characters in different scores, we compared the abundances of infiltrating immune cells in the groups (Fig. 4A). We observed significant differences in the abundances of many immune cells, such as memory B cells, resting CD4 + T cells, M0 monocytes, M1 monocytes, M2 monocytes, and resting mast cells, between the risk subgroups. We also performed ESTIMATE to explore the immune cell infiltration levels (Fig. 4B). And we found that the risk score was markedly related with the abundances of CD4 + T cells and macrophages (Fig. 4C). We performed ssGSEA based on TCGA-LIHC cohort, and Oocyte meiosis was mainly enriched in the high-risk score subgroup, while steroid hormone biosynthesis, butanoate metabolism and beta alanine metabolism were mainly enriched in the low-risk score subgroup (Fig. 4D). Additionally, the correlation between these biological functions and the risk score was further assessed (Fig. 4E). The risk score was positively associated with the cell cycle but negatively correlated with beta alanine metabolism, fatty acid metabolism, peroxisome, primary bile acid biosynthesis, propanoate metabolism and tryptophan metabolism(Fig. 4 E). Notably, we found a remarkably negative correlation between the risk score and tryptophan metabolism ssGSEA scores. The risk score was closely associated with tryptophan metabolic pathways and could be used to indirectly assess the immune microenvironment (Fig. 4F).

Immunotherapy for HCC is promising but particularly challenging due to the complex immune microenvironment and lack of reliable biomarkers for immunotherapy [28]. Here, we calculated the expression levels of immune checkpoint genes between risk subgroups. The results showed that the expression levels of various genes, such as TNFRSF14, NRP1, LAIR1, TNFSF4, CD276, CD80, CD44, and CD86, were higher in the high-risk score subgroups (Fig. 5A). Then, we analyzed the differences in the effects of immunotherapy between different risk score subgroups using TIDE software. The high-risk subgroup had a higher TIDE score, suggesting that this group had a higher possibility of immune escape (Fig. 5B). Further research showed that there was a significant correlation between the risk score and the TIDE, interferon gamma (IFN-γ), myeloid-derived suppressor cell (MDSC), and exclusion scores (Fig. 5C).

HCC patients with clinical information were selected for further analysis. Univariate analysis indicated that T stage, M stage, pathological stage and risk model were significantly related with overall survival (Fig. 6A). After the multivariate analysis, only the risk model remained an independent risk factor associated with prognosis (Fig. 6B). Our findings indicated that the risk model had excellent predictive efficiency for prognostic prediction in HCC.

To observe immunotherapy in the different risk score, we assessed the ability of the risk model to explore the response to ICIs in HCC patients. In the IMvigor210 cohort, patients were varied into complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD). SD/PD patients had higher risk scores than those of other types of responders (Fig. 6C). The percentage statistics indicated that the treatment effect was better in the low-risk score group (Fig. 6D). We analyzed the survival difference in all samples in IMvigor210, and we observed that patients with a higher risk score were related with a poorer survival rate (Fig. 6E). In addition, there was a significant survival difference in early stage and advanced stage, respectively (Fig. 6F and G). All the results suggest that the ten-gene model could be used for the assessment of immunotherapy efficacy and for guiding therapeutic strategies for patients with HCC.