Single-cell sequencing and bulk RNA sequencing reveal a cell differentiation-related multigene panel to predict the prognosis and immunotherapy response of hepatocellular carcinoma

To the Editor: Hepatocellular carcinoma (HCC) is the sixth most common tumor and the third leading cause of cancer death globally in 2020.^[1] The lack of early diagnosis and high intratumoral heterogeneity in HCC are important reasons for its high mortality. Genetic susceptibility and environmental stimuli lead to a tumor mass consisting of mixed tumor cells with different differentiation characteristics. Thus, intratumoral heterogeneity is the inevitable product of the differentiation of tumor stem cells.^[2]

Traditional bulk RNA sequencing (RNA-seq) can only provide an average level of gene expression in tumor tissues. However, single-cell sequencing (scRNA-seq) can identify cell clusters in diverse differentiated states and help to reveal cell heterogeneity in the tumor.^[3] Therefore, we combined bulk RNA-seq in conjunction with scRNA-seq to construct a prognostic signature and to explore the potential mechanisms and immunotherapy response populations.

scRNA-seq data file GSE146115 involving four HCC samples and 3200 cells was acquired from Gene Expression Omnibus. The bulk RNA-seq data of 49 normal and 373 HCC samples were accessed from The Cancer Genome Atlas (TCGA) served as the training dataset, and an additional 220 normal and 225 HCC samples were obtained from GSE14520 served as the validation dataset. The R package “sva” (https://www.r-project.org/) was used to eliminate the batch influence.

The flowchart of this study is shown in Supplementary Figure 1. We first performed quality control and normalized the scRNA-seq data and retained the top 1500 genes with highly variable features for further analysis [Supplementary Figure 2A–D]. Principal component analysis was used to screen out the significant dimensions [Supplementary Figure 2E]. The first 20 principal components were selected for dimensionality reduction through the t-distributed stochastic neighbor embedding algorithm, and the FindClusters function with a resolution of 0.5 was performed to aggregate HCC cells into 12 clusters [Supplementary Figure 2F, G]. The number of cells per cluster, per sample was shown in Supplementary Table 1. According to log2 fold change [FC] > 0.5 and false discovery rate (FDR) < 0.05, the marker genes in each cluster were identified, and the top 10% of marker genes in each cluster were shown in the heatmap [Supplementary Figure 2H]. According to the marker genes [Supplementary Table 2], 12 clusters were annotated as hepatocytes, monocytes, T cells, and natural killer (NK) cells with the “SingleR” package [Supplementary Figure 2I, J]; and pseudotime and cell trajectory analysis were performed with the “Monocle” package [Supplementary Figure 2K–M]. A differential expression analysis between different branches was conducted to select HCC cell differentiation-related genes (HDRGs). A total of 912 HDRGs were obtained [Supplementary Table 3] and shown in the heatmap [Supplementary Figure 3].

The gene expression profiles of the GSE14520 dataset were obtained to identify HDRG-related molecular subtypes with the “ConsensusClusterPlus” package [Supplementary Figure 4A, B]. We then compared whether the survival outcomes and clinical characteristics of the three subtypes differed [Supplementary Figure 4C–F]. We found that the survival probability and tumor staging show significant differences, which suggests that HDRG-based patient classification can be used to predict patient prognosis. We then investigated the discrepancy in the expression of common immune checkpoint genes (ICGs) [Supplementary Table 4] within different clusters and performed a survival analysis [Supplementary Figure 5].

Next, we screened the genes used to construct the prognostic signature. First, there are 723 HDRGs left after taking an intersection of genes between TCGA and the GSE14520 cohort [Supplementary Figure 4G]. The expression profiles of 723 HDRGs in training cohort and testing cohort were shown in Supplementary Table 5 and Supplementary Table 6. We performed weighted gene co-expression network analysis based on HCC samples in TCGA and five modules were identified. Three modules (yellow, blue, and turquoise) were closely related to the grade of HCC [Supplementary Figure 6A–C]. Because grade is a clinical trait associated with cell differentiation, so we selected them for further performing differential expression analysis between tumor and normal tissues from TCGA, and 136 genes were obtained [Supplementary Figure 6D]. Subsequent univariate Cox regression analysis of these genes yielded 91 HDRGs with prognostic value [Supplementary Figure 7]. The least absolute shrinkage and selection operator regression model was utilized to develop a novel HDRG prognostic signature [Supplementary Figure 8A, B]. The patients were stratified into two subgroups: high-risk and low-risk, with the training cohort's median risk score serving as the cutoff point [Supplementary Figure 8C–D]. The results showed that high-risk patients were more likely to die [Supplementary Figure 8E, F], and the prognosis of the high-risk subgroup was worse than the low-risk subgroup [Supplementary Figure 9A, B]. The area under the curve of the receiver operating characteristic (ROC) curve demonstrated that the risk score had a satisfied prognostic value [Supplementary Figure 9C, D]. Furthermore, the results in Supplementary Figure 9E, F showed the expression levels of the seven HDRGs in the 12 clusters.

Univariate and multivariate analyses suggest that risk score and stage can serve as an independent prognostic index [Supplementary Figure 10A, B]. Then, package “rms” was utilized to construct the nomogram with risk score and stage. The calibration and ROC curves indicate the good predictive power of our nomogram [Supplementary Figure 10C–E].

Immune infiltration and tumor immune microenvironment play significant roles in various cancers and could predict the responses to immunotherapy.^[4] In addition, it had been thought that immune cells and ICGs were prospective targets for the treatment of cancers. So, we utilized several algorithms to assess immune infiltration between high-risk and low-risk subgroups. Meanwhile, the differential expression analysis of common ICGs between two subgroups was conducted, and only the statistically significant results were shown. The results display that B cells, CD4⁺ T cells, neutrophils, macrophages, activated NK cells, and myeloid dendritic cells show a higher expression in the high-risk subgroup, indicating the presence of higher pre-existing anti-cancer immunity of patients in the high-risk subgroup [Supplementary Figure 11A]. However, most ICGs (26/27) were expressed higher in the high-risk subgroup [Supplementary Figure 11B], indicating that immune cells can be inhibited by the expression of ICGs, resulting in tumor immune escape, which leads to a worse prognosis. This also suggests that patients in the high-risk subgroup are more likely to benefit from checkpoint inhibitor therapy.

Tumor mutation burden (TMB) can influence the response to immunotherapy of cancer populations with different characteristics.^[5] Thus, we further explored the discrepancy in the expression of TMB between high-risk and low-risk subgroups. The result showed that higher proportion of TMB exist in high-risk subgroup, indicating that HCC patients with higher risk score might be beneficial from the immunotherapy [Supplementary Figure 12A]. Patients with low TMB have demonstrably better clinical outcomes [Supplementary Figure 12B, C]. The Tumor Immune Dysfunction and Exclusion (TIDE) web platform was designed to promote hypothesis generation, biomarker optimization, and patient stratification in immunooncology research through public data reuse methods. The lower the TIDE score, the smaller the possibility of immune escape, suggesting that patients are more likely to benefit from immune checkpoint blockade (ICB). In addition, macro-satellite instability, interferon-gamma, and PDL1 levels are all positive biomarkers of ICB reaction. The results suggest that immunotherapy is more effective for high-risk subgroup patients [Supplementary Figure 12D–K].

Gene set enrichment analysis software was used to analyze the potential functions between high-risk and low-risk subgroups in TCGA. Supplementary Figure 13, show the 16 crucial pathways in the high-risk population.

In conclusion, our research identified and validated a seven-gene signature associated with cell differentiation in HCC patients based on scRNA-seq and bulk RNA-seq. The signature had the potential to serve as a potential biomarker to predict the prognosis and immunotherapy response for HCC patients.

Funding

This work was supported by The Leading Talent of Hundred, Thousand and Ten Thousand Project of Xingliao Gifted Person Program of Liaoning Province (No. XLYC1905013), Type A Project of Leading Talent's Innovative Research of Dalian, Capacity-Building Projects of Clinical Discipline of Traditional Chinese Medicine of Health, and Family Planning Commission of Liaoning Province in 2018 (No. LNZYXZK201806).

Conflicts of interest

None.