Identification of Proteome-Based Immune Subtypes of Early Hepatocellular Carcinoma and Analysis of Potential Metabolic Drivers

Abstract

Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality worldwide, ranking fourth in frequency. The relationship between metabolic reprogramming and immune infiltration has been identified as having a crucial impact on HCC progression. However, a deeper understanding of the interplay between the immune system and metabolism in the HCC microenvironment is required. In this study, we used a proteomic dataset to identify three immune subtypes (IM1-IM3) in HCC, each of which has distinctive clinical, immune, and metabolic characteristics. Among these subtypes, IM3 was found to have the poorest prognosis, with the highest levels of immune infiltration and T-cell exhaustion. Furthermore, IM3 showed elevated glycolysis and reduced bile acid metabolism, which was strongly correlated with CD8 T cell exhaustion and regulatory T cell accumulation. Our study presents the proteomic immune stratification of HCC, revealing the possible link between immune cells and reprogramming of HCC glycolysis and bile acid metabolism, which may be a viable therapeutic strategy to improve HCC immunotherapy.

Graphical Abstract

Highlights

• •Three proteome-based immune subtypes of early-stage HCC.
• •Comprehensive immune landscape of early-stage HCC.
• •The dysregulation of glycolysis and bile acid metabolism in early-stage HCC.
• •Correlation between metabolic dysregulation and immune infiltration.

In Brief

This study used proteomic datasets to identify three immune subtypes in early-stage hepatocellular carcinoma (HCC), each of which has distinctive clinical, immune, and metabolic characteristics. The subtype (IM3) with the worst prognosis showed the highest levels of immune infiltration. IM3 also exhibits increased glycolysis and decreased bile acid metabolism, which is closely related to CD8 T cell exhaustion and regulatory T cell accumulation. These findings are helpful for precision medicine of HCC patients based on molecular characteristics.

Hepatocellular carcinoma (HCC), accounting for more than 80% of primary liver cancers globally, represents the fourth leading cause of cancer-related mortality worldwide (1, 2). Alcoholism, cirrhosis, and hepatitis B virus (HBV) infection are among the major risk factors associated with the development of HCC (3, 4).

Although existing breakthroughs have been made in surgical treatment, such as ablation, resection, and transplantation for HCC interventions (5), patients with HCC still face a poor prognosis with an overall 5-year survival of only approximately 19.9% (6). This continued poor outcome is largely because individuals with HCC often present at advanced stages of the disease. The only means of achieving long-term survival is diagnosing the disease at an early stage when potentially curative treatments are feasible (7).

Advances in omics technologies permit the large-scale analysis of the molecular characteristics for HCC and have defined several clinal patient subtypes. Transcriptomic subtyping has revealed HCC subclasses that differ in the expression of genes related to proliferation, stemness, metabolism, hepatocyte differentiation, and liver function (8, 9, 10, 11, 12, 13). Among these transcriptome-based subtyping, Sia et al. divided HCC into two subtypes according to distinct TME characteristics, where “immune active” subtype demonstrated favorable prognosis and “immune exhausted” subtype had a bad prognosis. Proteomics has proven to be a valuable approach for investigating cancer biological changes and disease classification, as it offers a comprehensive representation of cellular states (14). Based on proteomic data, Jiang et al. stratified the 101 early-stage HCC cohort into molecular subtypes S-I, S-II, and S-III. Among these subtypes, the S-III (characterized by disrupted cholesterol homeostasis and SOAT1 overexpression) was associated with poor outcomes (15), and 90.9% of S-III tumors were observed a high level of infiltration. Additionally, a subset of S-III tumors in Jiang et al.’s subtyping showed enrichment of the T-exhaustion (42.4%), M2-macrophage (39.4%), and regulatory T cell signatures (42.4%).

While these studies have provided great insights into the HCCs, only Jiang et al.'s study focused on early-stage HCC (15), and there is still no immune signature-based classification of early-stage HCC for better immunotherapy. In fact, although immune checkpoint blockade (ICB) has been shown to enhance clinical survival in patients with HCC (16, 17, 18), about 70% to 80% of cases exhibit no response to ICB and the underlying mechanism for therapeutic resistance remains unclear (19). It is necessary to stratify clinical HCC patients into different immune subtypes for targeted treatment design boosting the efficacy of cancer immunotherapy (20).

In this study, we identified three immune subtypes of HCC based on the proteome and investigated the interactions between immune infiltration and metabolic reprogramming in the HCC immune microenvironment. This research has the potential to shed light on the HCC immune microenvironment and HCC personalized treatment.

Experimental Procedures

Experimental Design and Statistical Rationale

For the discovery cohort, we used “Jiang et al. (15) proteomics dataset”. To validate the findings, three other datasets were established, including “Gao et al. (21) proteomics dataset”, “Roessler et al. (22) transcriptomics dataset”, and “TCGA-LIHC and Jiang et al. transcriptomics dataset”. We also designed a protocol to identify and characterize immune subtypes, examined the metabolic heterogeneity and its correlation with immune infiltration, and investigated the role of glycolysis and bile acid metabolism in the IM3 immune microenvironment (23).

Kruskal–Wallis test was used to test whether there are significant differences in biological characteristics among three subtypes, such as immune cell, stromal cell, Treg, exhausted T cell, macrophage M2, cell proliferation, and hypoxic activity. Wilcoxon test was used to test whether there were significant differences in response to immunotherapy (TIDE score) between two groups of immune subtypes. Spearman correlation was used to examine whether two variables are correlated with one another or not. Kaplan–Meier plots (Log-rank test) were used to describe overall survival. Univariate Cox proportional hazards regression models were used to identify variables associated with overall survival. Student’s t test was used to determine the statistical significance of the difference for immune signatures. All statistical tests were two-sided, and statistical significance was defined as p-value < 0.05. To account for multiple testing, Benjamini-Hochberg FDR correction was applied to adjust the p-values. All these statistical analyses were performed in R (version 4.1.1).

HCC Cohorts

We used quantitative proteomics data of 199 (101 early-stage HCC tumors and 98 non-tumors) samples from Jiang et al. (15) for the discovery cohort (“Jiang et al. proteomics data”). These data were downloaded from PRIDE database (24) (PXD006512 and PXD008373). First, the iBAQ intensities of 199 samples were extracted from the MaxQuant result files to represent the final expression of a particular protein across samples, resulting in a 9252 × 199 protein-expression matrix. Then, the expression matrix was subjected to quantile normalization using normalized quantile functions implemented in the R package limma. After a log2 transformation and removal of proteins expressed in less than 25% of the samples, the remaining missing data was then imputed by the global minimum value. Finally, a 7613 × 199 protein-expression matrix was used for the subsequent analysis.

To evaluate the robustness of our immune subtypes derived from “Jiang et al. proteomics dataset”, we also established three other transcriptomics and proteomics datasets of early-stage HCC (Supplemental Table S1).

Set 1: Gao et al. Proteomics Dataset

We downloaded this dataset from Gao et al (21). This dataset contained 68 tumor samples classified as BCLC stage A. The raw protein expression values (TMT intensities) have been subjected to median normalization by column and log2 transformation. K-nearest neighbor imputation was applied to impute the missing values. We normalized the logged expression values for each gene to standard deviations from the median.

Set 2: Roessler et al. Transcriptomics Dataset

We downloaded this dataset from the GEO repository with accession number GSE14520. This dataset consisted of 172 tumor samples classified as BCLC stage A. Raw gene expression data has been normalized using the Robust Multi-array Average method and the log2 transformation (22, 25).

Set 3: TCGA-LIHC and Jiang et al. Transcriptomics Dataset

The TCGA-LIHC dataset consisted of gene expression datasets of 79 tumor samples classified as tumor Stage-I. We downloaded the log2(FPKM+1) transformed data from TCGA data center (https://gdc-hub.s3.us-east-1.amazonaws.com/download/TCGA-LIHC.htseq_fpkm.tsv.gz). The Jiang et al. (15) transcriptomics dataset consisted of 31 early-stage HCC tumor samples with complete survival information. Genes detected log2(FPKM+1) >0 in at least 60% of tumor samples in each dataset were considered. Considering there are only 79 and 31 samples in TCGA-LIHC and Jiang et al. datasets respectively, we integrated them as “TCGA-LIHC and Jiang et al. transcriptomics dataset” for subsequent analysis.

Representative Immune Signatures

First, 181 tumor-related immune signatures (each signature is a gene set, Supplemental Table S2) were collected from previous literature (20, 23, 26, 27, 28). Then, single-sample gene set enrichment analysis (ssGSEA) (29) was used to calculate separate enrichment scores for each pairing of protein expression profile and immune signature as implemented in the R package GSVA (30) v1.44.0. Spearman correlation coefficients of enrichment scores among each immune signature were calculated to create a correlation matrix. Next, WGCNA package (31) in R with the following parameters: “power = 16, TOMType = unsigned, pamStage = FALSE, and minModuleSize = 2” was used to cluster the correlation matrix, generating 15 immune signature modules. Finally, within each module, the signature with the highest kME was considered as the most representative signature of that module, where the kME of a signature was measured by the signed KME function in WGCNA (31).

Identification and Analysis of Immune Subtypes

The 15 representative signatures were used to cluster HCC tumor samples using a model-based clustering method performed with the R package mclust (32) v5.4.9. Maximal Bayesian Information Criterion (BIC) was found to occur with three cluster solution, therefore three immune subtypes were used for the remainder of analyses. The binomial logistic regression through the R package glmnet (v2.0–16) was used to identify which signature explained one subtype versus the other two. A log-rank test was applied to compare the overall survival (OS) or disease-free survival (DFS) among three immune subtypes.

Evaluation of TME Cell Abundance and Patient’s Response to Immunotherapy

ESTIMATE algorithm was used to calculate the immune score, and stroma score (33). The basic idea of ESTIMATE algorithm is to seek a signature that is characteristic of immune cells and stromal cells in general, because the fibroblast/mesenchymal nature of stromal cells can separate their gene expression profile from that of the epithelial tumor cells (33). Tumor Immune Dysfunction and Exclusion (TIDE, http://tide.dfci.harvard.edu/) tool (34) was used to estimate T cell Exclusion, T cell dysfunction, and also TIDE score for the clinical responsiveness of patients to ICB therapy (anti-PD1/CTLA4).

Analysis of Metabolic Pathway Activity

The gene sets of 101 metabolic pathways were downloaded from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database (35). An R package developed by Xiao et al. (36) was used to calculate metabolic pathway scores measuring the activity of each metabolic pathway. In this method, the pathway activity is quantified as the weighted average expression levels of all genes within the pathway in each sample. Here, the pathway activity S_k,j of the k-th pathway in the j-th sample was computed as follows. First, for the i-th gene in the k-th pathway, G_i,j denotes the expression level of the i-th gene of the j-th sample,

$$G_{i,j},i\in 1...M,j\in 1...N$$ (1)

where M is the number of genes, and N is the number of samples. Then, the relative expression level R_ij of the i-th gene in the j-th sample was then denoted as the ratio of G_i,j to its average over all samples.

$$R_{i,j}=\frac{G_{i,j}}{\frac{1}{N}{\sum }_j^N{G}_{i,j}}$$ (2)

Here, R_i,j >1 means that the expression level of gene i is higher in sample j compared to its expression level over all samples. Finally, the pathway activity score S_k,j for the k-th pathway in the j-th sample was then denoted as the weighted average of R_i,j over all genes included in this pathway.

$$S_{k,j}=\frac{{\sum }_{i=1}^{m_k}{w}_i\times R_{i,j}}{{\sum }_{i=1}^{m_k}{w}_i}$$ (3)

Here m_k is the number of genes in the pathway k, w_i is the weighting factor equal to the reciprocal of the number of pathways that include the i-th gene.

Calculation of Hazard Ratio to Measure the Protein/Pathway Risk Score

For each protein or pathway, the median protein expression values or pathway activity scores were used as a cut-off to split early-stage HCC patients into two groups. Univariate Cox proportional hazard regression analyses were then employed to estimate the hazard ratio (HR) of each protein/pathway using the survival v3.3 to 1 package in R. The hazard ratio (HR) of each protein/pathway was displayed in the form of forest plots using the forest plot package in R.

Calculation of Hypoxia Score

The HALLMARK hypoxia scores were calculated using ssGSEA based on HALLMARK_HYPOXIA gene set from the Molecular Signature Database (MSigDB) v7.5.1 (37). The Buffa hypoxia scores (BHS) and the Winter hypoxia scores (WHS) were respectively estimated by Buffa (38) and Winter method (39).

Calculation of the Davies Bouldin Index and Calinski Harabasz Index

The Davies Bouldin Index (40) and the Calinski Harabasz index (41) were calculated using Python sklearn library (42) to evaluate clustering performance.

Results

Immune Subtypes of HCC and Their Associations With Overall Survival

The biological mechanisms underlying the response to immunotherapy in the HCC immune microenvironment are still unclear, which is crucial for the design of immunotherapy strategies (19). In order to further explore the immune microenvironment in HCC and the underlying metabolic drivers, we designed a protocol to classify the 101 early-stage HCC cohorts into different immune subtypes.

First, we undertook an extensive literature search (20, 23, 26, 27, 28) and collected 181 signatures (each signature is a gene set, Supplemental Table S2) that are known to be associated with immune activity in tumor tissue. Based on the protein expression profile, we scored 181 immune signatures across each of the 101 HCC tumor samples using ssGSEA. Then, we used WGCNA to cluster 181 immune signatures into 15 modules based on their enrichment scores. WGCNA has been successfully used to identify representative immune signatures for pan-cancer subtyping (23), for encapsulating the core features and key patterns of the overall gene expression profiles. Finally, using ssGSEA scores of 15 representative immune signatures (Supplemental Table S3), we stratified 101 early-stage HCC samples into three clusters (Fig. 1A). These three immune subtypes (IM1, IM2, and IM3, with 41, 28, and 32 cases, respectively) showed distinct representative immune signatures (Fig. 1C and Supplemental Fig. S1) and were associated with different OS (Log-rank test, p-value = 0.013, Fig. 1B). The IM1 subtype has the best OS, whereas both IM2 and IM3 has worse survival. Following the same protocol as Calza et al. (43), we performed an overlap analysis of our immune subtypes (IMS: IM1, IM2, and IM3) with the molecular subtypes identified by Jiang et al. (JMS: S-I, S-II, and S-III). We observed the highest rate of concordant assignments occurs between IM1 and S-I subtypes (29.7%), IM3 and S-III subtypes (26.7%), and IM2 and S-II subtypes (22.8%). The overall concordance rate (43) between IMS and JMS (15) was 79.2% (which is the sum of the percentage of patients allocated in both S-I and IM1 ("SI-IM1"), both S-II and IM2 ("SII-IM2"), and both S-III and IM3 ("SIII-IM3", Supplemental Tables S4 and S5). As shown in Figure 1C, our immune subtypes (IMS) displayed substantial overlap with Jiang et al. (15)’s proteome-based molecular subtypes (JMS) in group assignment, although only 197 genes appear in both IMS and JMS gene sets, which account for 13.97% of the IMS gene set and 10.93% of the JMS gene set (Supplemental Fig. S2A and Supplemental Tables S4 and S5).

Fig. 1.

The immune subtypes in early-stage HCC cohort.A, scatter plot of Bayesian Information Criterion (BIC) values for mclust models across the number of clusters. Based on the maximal BIC values for up to 10 clusters, we consider K = 3 to be the optimal number of clusters. Each line and symbol represent a different parameterization of the covariance matrix of models, where each letter describes the volume, shape, and orientation of the covariance structure. B, Kaplan–Meier curves of overall survival (OS) for each immune subtype in early-stage HCC cohort (IM1, n = 41; IM2, n = 28; and IM3, n = 32). The p-value was calculated by the two-sided log-rank test. C, heatmap of immune signature scores calculated by ssGSEA. Our immune subtypes and Jiang et al.'s molecular subtypes are annotated on the top of the heatmap. Color scale indicates the z-score of the data by row. E, equal; I, identity matrix; V, variable.

To evaluate the robustness of our immune subtype results derived from Jiang et al proteomics dataset, we compiled three available transcriptomics and proteomics datasets of early-stage HCC samples (“Roessler et al. transcriptomics dataset”, “TCGA-LIHC and Jiang et al. transcriptomics dataset”, and “Gao et al. proteomics dataset”, Supplemental Table S1), and performed immune subtyping using the same protocol as that for IMS. As shown in Supplemental Fig. S3A, all patients in three datasets can be grouped into three immune clusters (IM1, IM2, IM3, Supplemental Table S6) with significant overall survival differences. The IM1 subtype has the best OS, whereas IM3 has the worst survival.

Characteristics of the Immune Microenvironment Among Three Immune Subtypes

To explore the immune microenvironment of HCC, we calculated immune scores and stroma scores using the ESTIMATE algorithm (33) to predict the level of immune cells and stromal cells. The box plot shows that the immune score and the stroma score actually did not increase with the HCC prognosis (Fig. 2A and Supplemental Fig. S4A). IM3 and IM2 with worse survival had the highest and the lowest immune scores, respectively. The histological examination of immune cell infiltration from Jiang et al. (15) further supported these results (Fig. 2B and Supplemental Table S5).

Fig. 2.

Characteristics of the immune microenvironment of three immune subtypes.A, difference of immune scores of patients between three immune subtypes. B, difference of immune infiltration scores based on histological evaluation of patients among three immune subtypes. C, estimates of univariate logistic regression analysis and the 95% confidence interval (CI) are illustrated by forest plot to assess which immune signatures inferred by ssGSEA explain the poorest prognosis cluster (IM3) versus IM1 & IM2. EMT, epithelial–mesenchymal transition. D, heatmap showing log₂ fold-change of immunomodulatory genes that are significantly regulated when compared with normal samples derived from each immune subtype. The red cell denotes upregulation and the blue cell downregulation. The p-values in (C) were given by the two-tailed Student’s t test, while p-values in (A) and (B) were calculated by Kruskal-Wallis test. See Figure 1 for details on immune subtypes. Within each box in (A) and (B), the thick line represents the median value. The bottom and top of the boxes are the 25th and 75th percentiles (interquartile range).

To identify which signature explained IM3, we tested how each signature score contributes to IM3 versus IM1 and IM2 using generalized linear models. Exhausted CD8 T cell was a significant explanatory variable of IM3 (Fig. 2C). However, exhausted CD8 cell signature score was not of significant prognostic value according to Cox regression analysis (Supplemental Table S7). This result may be attributed to the limited sample size of the HCC cohort, which may have hindered the ability to identify significant differences. We also tested which signature explained IM1(Supplemental Fig. S4D) or IM2 (Supplemental Fig. S4C). We found neutrophil and natural killer cells contribute positively to IM1, while helper T cell and cell proliferation have a positive influence on IM2.

Considering the crucial role of immunomodulators in the formation of tumor microenvironment (TME), we examined the relative expression levels of major histocompatibility complex (MHC), immunostimulatory, and immunoinhibitory molecules among three immune subtypes compared to normal liver tissues (Fig. 2D). We found that these immunomodulators exhibited higher expression in IM3 while many of them were downregulated in IM1 and IM2. Overall, these results suggest that IM1 with the best prognosis exhibited a weaker tumor cell proliferation signal, IM2 exhibited a weak immune response accompanied with a strong tumor cell proliferation signal, and IM3 employed different tumor escape mechanisms while having the highest level of immune infiltration.

We also examined the above immune microenvironment characteristics of immune subtypes in three other available transcriptomics and proteomics datasets of early-stage HCC (“Roessler et al. transcriptomics dataset”, “TCGA-LIHC and Jiang et al. transcriptomics dataset”, and “Gao et al. proteomics dataset”, Supplemental Table S1). All the results of immune microenvironment characteristics were consistent with the results from “Jiang et al proteomics dataset” (Supplemental Fig. S3, B and C).

Metabolic Heterogeneity and its Correlation With Immune Infiltration

Given the observed link between T cell exhaustion and poor clinical outcomes in HCC, it is imperative to identify the factors that induce T cell exhaustion in HCC. As tumor progression is a dynamic process involving continuous interactions of tumor cells and other cells in TME including immune cells and stromal cells, tumor cells may profoundly influence various cell types within the tumor ecosystem to promote their survival and the dissemination of malignancy (44, 45). Numerous studies have indicated that metabolic changes in tumor cells can directly affect the composition and function of immune cells residing in the same microenvironment (46, 47). Meanwhile, metabolic reprogramming has been one of the crucial barriers to immunotherapy (48).

To investigate the impact of metabolic reprogramming on the immune microenvironment in HCC, we performed metabolic pathway analysis using an algorithm developed by Xiao et al. (36). The results showed distinguished metabolic pathway activities among the three immune subtypes (Fig. 3A and Supplemental Table S8). To identify the main contributors to metabolic heterogeneity, we compared metabolic pathway activity scores among three subtypes (Fig. 3B and Supplemental Fig. S5, A and B). IM1 patients showed higher activity in fatty acid metabolism and fatty acid degradation. IM2 patients had higher activity in aminoacyl tRNA biosynthesis and TCA cycle. IM3 patients exhibited higher activity in glycolysis and arachidonic acid metabolism. Of these metabolic pathways, glycolysis and arachidonic acid metabolism show high-risk scores for a mortality prognosis, suggesting that alteration of tumor metabolism may have a role in reprogramming immune cell function in HCC (Fig. 3A).

Fig. 3.

Different metabolism reprogramming among immune subtypes and its relationship with immune infiltration.A, significantly deregulated metabolic pathways in immune subtypes. Left, the activity of metabolic pathways. Right, prognostic risk scores of each pathway. B, the waterfall plot demonstrates the differences in metabolic pathway activities between IM3 and IM1&IM2. The t values from a linear model are shown, with positive values in blue and negative in green. C, Spearman correlation matrix of characteristic metabolic pathways and immune cell activities. D, Spearman correlation matrix of characteristic metabolic pathways and immune cell activities. Correlation coefficients are represented in the form of a heatmap using a colored scale ranging from blue (minimum correlation) to red (maximum correlation). The p values were calculated by the two-tailed Student’s t test (C and D). See Figure 1 for details on immune subtypes.

To assess the potential crosstalk pattern between immune and metabolic reprogramming, we calculated Spearman’s correlation coefficients between metabolic pathway activity and expression of immunomodulators (Fig. 3C). The majority of metabolic pathways showed a significant positive correlation with MHC, immunostimulatory and immunoinhibitory molecules, while primary bile acid metabolism, and fatty acid metabolism showed a negative correlation with these genes (t test, p-value < 0.05). These results imply a close relationship between metabolic alterations and immune response.

We further conducted a correlation analysis between metabolic pathways and immune signatures in HCC (Fig. 3D). Glycolysis and arachidonic acid metabolism positively correlated with epithelial-mesenchymal transition (EMT), exhausted CD8 T cells, and M2 macrophages, implying a strong association with immunosuppression in HCC. On the other hand, there were significant negative correlations between the overwhelming majority of immune signatures and pathways of lipid metabolism like fatty acid biosynthesis, fatty acid degradation, and primary bile acid metabolism, indicating that the low level of immune infiltration in IM1 patients is related to lipid metabolism.

Role of Glycolysis and Bile Acid Metabolism in the IM3 Immune Microenvironment

To gain a deeper understanding of the potential therapeutic approaches for HCC, we specifically focused on the IM3 subtype, which demonstrates the poorest prognosis and the most favorable response to immunotherapy (Fig. 4). Of note, our findings indicate the high levels of T cell exhaustion in IM3 correlate strongly with both the upregulation of glycolysis and downregulation of bile acid metabolism (Fig. 3D).

Fig. 4.

Glycolysis and hypoxia and their relationship with immune infiltration.A, Spearman correlation matrix of glycolysis genes and immune cell activities. Correlation coefficients are represented in the form of a heatmap using a colored scale ranging from blue (minimum correlation) to red (maximum correlation). B, forest plot illustrating the hazard ratio of glycolysis genes (tumor samples, n = 101). The red points indicate the log₂-based hazard ratio for each protein; endpoints represent upper or lower 95% confidence intervals. C, distribution of hypoxia scores among three clusters. D, scatter plots comparing activities of glycolysis and hypoxia in the HCC cohort. The colors of points indicate different immune subtypes. R value means the Spearman correlation coefficient between glycolysis pathway activity and hypoxia score. E, Spearman correlation matrix of immune cell activities and hypoxia scores calculated based on three methods in the HCC cohort. F, biological insight of glycolysis and hypoxia among three immune subtypes of HCC. The diagram summarizes relevant genes involved in glycolysis and hypoxia in HCC. Each gene alteration is represented as log ratios (fold change, expressed as log₂ ratio of average protein abundance in each immune subtype versus the non-tumor group). Red, upregulated genes; blue, downregulated genes. The yellow border indicates a gene that is significantly different among subtypes (F test, p-value< 0.05). The size of the dot in A and E indicates p value, and the outlined boxes indicate significance (p value < 0.05). The p-values were calculated by the two-tailed Student’s t test (A, D and E) or Kruskal-Wallis test (C and F). See Figure 1 for details on immune subtypes.

It has been shown that lactate from the glycolysis metabolism of tumor cells can be transported to the TME, where it inhibits the function of T cells (49). Lactic acid induces tumor-associated M2 macrophage polarisation (50, 51). According to our results, genes involved in lactate synthesis (such as LDHA and MCT4) were relatively upregulated in IM3 patients (Fig. 4F) and showed high-risk scores for a mortality prognosis (Fig. 4B). We subsequently calculated the correlation coefficients between their expression and exhausted CD8 T cell activity (Fig. 4A), and found that most of them were positive. Considering hypoxia in the TME promotes glycolysis in tumor cells (52), we calculated the hypoxia score for each sample based on the Hallmark gene set from the MSigDB by ssGSEA. Patients in IM3 showed significantly higher hypoxia scores than those in both IM1 and IM2 (Fig. 4C). Indeed, previous studies have demonstrated that HIF1α can promote T cell exhaustion via activation of PD-L1 (53). Correlation analysis showed positive correlations between hypoxia scores and glycolysis pathway activity, as well as exhausted CD8 T cells signature scores (Fig. 4, D and E), which is consistent with previous studies (52, 53).

Ursodeoxycholic acid (UDCA) is a secondary bile acid derived from microbial conversion via 7α/β isomerization of chenodeoxycholic acid (CDCA) (54). Previous studies have demonstrated that UDCA accumulation can alleviate immunosuppression within tumors by inducing TGFβ degradation, which in turn reduces regulatory T cell differentiation (55, 56). UDCA has been shown to inhibit the proliferation of HCC cells in both mouse models and in vitro experiments. In our result, we observed a significant negative correlation between the primary bile acid synthesis level and Treg activity (Fig. 5A). Additionally, patients of IM3 showed downregulations in genes related to bile acid synthesis (such as SLC27A4, BAAT, and ABCB11) (Fig. 5C), and showed high ssGSEA scores of Treg (Supplemental Fig. S4B). These genes also significantly correlated with patient survival (Fig. 5B), further indicating their robust prognostic value for potential clinical application.

Fig. 5.

Primary bile acid biosynthesis and its relationship with immune infiltration.A, Spearman correlation matrix of primary bile acid biosynthesis genes and immune cell activities. Correlation coefficients are represented in the form of a heatmap using a colored scale ranging from blue (minimum correlation) to red (maximum correlation). The size of the dot indicates p-value, and the outlined boxes indicate significance (p-value < 0.05). B, forest plot illustrating the hazard ratio of primary bile acid biosynthesis genes (tumor samples, n = 101). The red points indicate the log₂-based hazard ratio for each protein; endpoints represent upper or lower 95% confidence intervals. C, biological insight of primary bile acid biosynthesis in the S-III subtype of HCC. The diagram summarizes relevant signatures and signaling cascades involved in glycolysis and hypoxia in HCC. Alteration scores of each gene are depicted as log ratios (fold change, expressed as log₂ ratio of average protein abundance in each proteomic subtype versus the non-tumor group). Red, upregulated genes; blue, downregulated genes. Yellow border indicates a gene is significantly different among subtypes (F test, p-value < 0.05). See Figure 1 for details on immune subtypes.

Discussion

It is crucial to stratify patients with clinical HCC, as it may help to inform the targeted design of treatments that boost the efficacy of cancer immunotherapy. Here, we identified three immune subtypes (IMS: IM1, IM2, and IM3) based on the proteome of patients with early-stage HCC, and analyzed the correlation between the immune microenvironment and metabolic reprogramming. We found that IM3 with the poorest prognosis has the highest levels of immune infiltration and T-cell exhaustion. Furthermore, IM3 showed elevated glycolysis and reduced bile acid metabolism, which was strongly correlated with CD8 T cell exhaustion and regulatory T cell accumulation. These findings hold promise in providing insights into the immune landscape of HCC and guiding personalized treatment approaches for this disease.

Although there is a substantial overlap between IMS and JMS (Supplemental Fig. S2A and Supplemental and Tables S4 and S5), the gene sets on which IMS and JMS are based differ greatly. JMS is based on the top 25% most-variant proteins (1802 genes) in 101 HCC tumor samples, while IMS 1410 immune-related genes. Only 197 genes appear in both IMS and JMS gene sets, which account for 13.97% of the IMS gene set and 10.93% of the JMS gene set (Supplemental Fig. S2A). IMS examined 1213 immune-related genes that were not used in JMS. This provides the following new insights into early-stage HCC (added value of IMS over JMS). First, IMS has a lower Davies Bouldin index (40) and a higher Calinski Harabasz score (41) than JMS, which means IMS is of higher compactness and lower separability (Supplemental and Tables S9 and S10). In addition, only one immune signature (cell proliferation) score showed a significant difference between concordant and discordant samples in IMS (Supplemental Fig. S2B). In JMS, six immune signature scores (S-I: macrophage M2, helper cell; S-II: macrophage M2, mast cell, dendritic cell; S-III: cell proliferation) were found to be significantly different between concordant and discordant samples in JMS (Supplemental Fig. S2B). Second, IMS can better characterize the immune landscape for early-stage HCC compared to JMS. Specifically, IM1 with the best prognosis showed weaker tumor cell proliferation signals. In contrast, IM2 exhibited a weak immune response accompanied by strong tumor cell proliferation signals. Meanwhile, IM3 was supposed to employ different tumor immune escape mechanisms while having the highest level of immune cell infiltration. Patients classified as IM3 subtype are likely to benefit from immunotherapy. Third, IM2 patients had higher activity in aminoacyl tRNA biosynthesis and TCA cycle. IM1 patients showed higher activity in fatty acid metabolism and fatty acid degradation. IM3 patients exhibited higher activity in glycolysis and arachidonic acid metabolism. Specifically, the high levels of T cell exhaustion in IM3 strongly correlated with both the upregulation of glycolysis and the downregulation of bile acid metabolism.

To provide some possible future directions to improve immunotherapy efficacy in HCC, we presented multiple possible factors that may contribute to the immunotherapy response difference (Fig. 2 and Supplemental Fig. S4A), including immune cell infiltration, stromal cell infiltration, T cell dysfunction, and T cell exclusion. Among these factors, T cell exclusion is a process that stromal cells of the TME that restrict the accumulation of T cells in the vicinity of cancer cells (57), which has been reported to play a major role in cancer-associated fibroblasts in decreasing cancer immunotherapy outcome (58). In addition, we used the TIDE score (based on T cell dysfunction and exclusion) to predict the response of three immune subtypes (IM1, IM2 and IM3) to immune checkpoint blockade (ICB). A lower TIDE score indicates a lower likelihood of immunological evasion and a higher possibility of benefiting from immune checkpoint blockade therapy (34). We found that in both “Jiang et al. proteomics dataset” and “Roessler et al. transcriptomics dataset”, IM3 patients had significantly lower TIDE scores than IM1&IM2 patients (Wilcoxon rank-sum test; p-value = 1.9e-02 and 2.1e-02 respectively; Supplemental Fig. S6). These results suggest that patients in IM3 are likely to benefit from immunotherapy, especially from immune checkpoint blockade therapies. Notably the above analyses are based on prediction of immunotherapy response by TIDE software, only serving as clues for differences in immunotherapy responses between subtypes. Further experimental validation in larger cohorts is required to confirm these results.

By comparing metabolic pathway activities, we simultaneously observed decreased bile acid synthesis and increased glycolysis in HCC IM3 patients (Fig. 3B). In fact, it has been reported that bile acid can physiologically regulate glucose metabolism by activating nuclear farnesoid X receptor (FXR) (59, 60, 61, 62), and in colon cancer the silent forms of FXR accelerate tumor progression by promoting aerobic glycolysis (63). Meanwhile, in HCC, bile acids have been found to be associated with tumor cell proliferation (64, 65), and the inhibition of the antitumor immune response (66). Two secondary bile acids, named ursodeoxycholic acid (UDCA) and taurocholic acid (TCA), can cause HepG2 cell apoptosis by activating the mitochondrial cell death pathway, which may inhibit or even reverse EMT (64, 65). Additionally, Ma et al. discovered that bile acids can act as a messenger to regulate chemokine CXCL16 level and, as a result, control the phenotype of hepatic NKT cells, determining whether they exert an antitumor immunity response in liver tumors (66).

Data Availability

Data in this study are included in the supplementary.

Supplemental data

This article contains supplemental data (15, 20, 23, 26, 27, 28).

Conflict of interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.