Combining bulk and single-cell RNA-sequencing data to develop an NK cell-related prognostic signature for hepatocellular carcinoma based on an integrated machine learning framework

A total of 7 data sets, containing gene expression data and corresponding clinical data, were enrolled to perform the present study. The TCGA–LIHC data set (n = 370) from The Cancer Genome Atlas (TCGA) database (https://​portal.​gdccancer.​gov/​repository) was used to establish the prognostic model as the training group and to perform a series of analyses related to prognosis-related genes selection, somatic mutation, immune microenvironment, immune checkpoint inhibitor (ICI) response and chemotherapy sensitivity. GSE14520 data set (n = 242) and GSE76427 data set (n = 115) from Gene Expression Omnibus (GEO) database (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​), and ICGC–LIRI–JP data set (n = 260) from International Cancer Genome Consortium (ICGC) database (https://​dcc.​icgc.​org/​) were utilized to verify the predictive power of the prognostic model. The single-cell RNA-seq data were obtained from 3 HCC samples (GSM4955419, GSM4955421, and GSM4955426) in GSE162616 data set from GEO database [33]. Moreover, the two immunotherapy cohorts, GSE91061 and PRJEB23709 data set from GEO database, were also used to externally validate the predictive power of ICI response of the model.

The inclusion criteria used to perform analyses in the present study were as follows: 1. patients with a pathological diagnosis of HCC. 2. Patients with complete RNA-seq data and survival data (survival time and survival status). The exclusion criteria were: 1. duplicate samples. 2. Patients with preoperative anti-HCC therapy. 3. Samples from recurrent HCC patients. 4. Patients with other malignancy history. 5. Patients died within 1 month after surgery. After these data cleansing processes, the demographic and clinicopathological data of the 7 abovementioned data sets was summarized in Additional files 1 to 7, respectively.

Weighted correlation network analysis (WGCNA) was often used to identify clusters or modules of highly associated genes. Unlike the conventional method to screen out potential biomarkers only based on differentially expressed genes, WGCNA could also associate modules with one another and with phenotypes, and assess their correlation strength [34]. Using the expression profile and corresponding clinical data from TCGA–LIHC data set, a scale-free co-expression network was established by “WGCNA” package. The soft-threshold value was selected when the independence degree reached 0.9. Thirty was set as the minimum of module size. The correlations between marker genes in each module and survival status, and survival time were evaluated by Pearson’s correlation analysis. The optimal module was selected through comprehensively considering the coefficient and P value.

The prognostic performance of the gene markers in the selected module was analyzed by univariate Cox regression analysis. The genes with P value of less than 0.05 were defined as prognosis-related genes. Next, the expression matrices and corresponding survival data of these prognosis-related genes were brought into our integrated machine learning framework containing a total of 10 machine learning algorithms and 77 combinations to find out the optimal algorithm to build up the prognostic signature. Of these 10 machine learning algorithms, survival support vector machine (SurvivalSVM), CoxBoost, least absolute shrinkage and selection operator (LASSO), supervised principal components (SuperPC), elastic network (Enet), StepCox, Ridge regression, partial least squares regression for Cox (plsRcox), random forest (RSF) and gradient boosting machine (GBM) were classical algorithms, their statistical characteristics were described in the previous study [35]. RSF, LASSO, CoxBoost and StepCox algorithm could be utilized for feature selection. Hence, the combined algorithms included at least one of the abovementioned four classical algorithms. TCGA–LIHC data set worked as the training cohort, while GSE14520 data set, GSE76427 data set and ICGC–LIRI–JP data set worked as the validation cohorts. In this process, C-indices of corresponding algorithm in the 4 data sets was calculated. The optimal algorithm was chosen according to the highest average C-index among the 4 data sets. Based on the selected algorithm, each HCC patient in the 4 data sets was scored. Setting the median risk score as the cutoff, patients were classified into high- and low-risk groups. Afterward, the overall survival (OS) curves between these two groups in TCGA–LIHC data set and other 3 validation cohorts were compared by Kaplan–Meier method. Using R package “timeROC”, time ROC analysis and clinical ROC analysis were enrolled to evaluate the predictive accuracy of our model. Moreover, univariate and multivariate Cox regression analysis were also conducted to explore whether the risk score was an independent indicator to predict the prognosis of HCC patients.

The above integrated machine learning framework determined the optimal algorithm to build up the risk model. In addition, the power of the risk model in prognostic evaluation was assessed by survival curves, ROC curves and Cox regression analysis. For better correlating with clinical practice, a nomogram combining the risk score and clinical parameters, including age, gender and clinical stage, was plotted through “rms” package in R. The corresponding calibration plot was also constructed to compare the fitting degree between the ideal and the actual survival probabilities. In addition, using “DecisionCurve” package, decision curve analysis (DCA) was also conducted to evaluate predictive accuracy of the nomogram. We also focused on whether the predictive performance of our model surpassed that of published signatures. Harrell's concordance index (C-index) was widely used to describe the fitting degree between predicted value and actual value of Cox model in survival analysis. The C-indices of enrolled signatures were visualized by R package “CompareC”.

Continuous accumulation of somatic mutations induces cancer occurrence and promotes cancer development [36]. The overall somatic mutation categories and frequencies of the samples in TCGA–LIHC data set were acquired from UCSC Xena (https://​xena.​ucsc.​edu/​) and were visualized by R package “maftools”. Next, we downloaded the mutation annotation format files from TCGA database and then used “maftools” package to calculate tumor mutation burden (TMB) of each sample. TMB was defined as an applied biomarker to identify potential beneficiaries for ICI treatment, and the detailed statistical processing and calculation methods were described by Chalmers et al. [37]. Estimation of stromal and immune cells in malignant tumor tissues using expression data (ESTIMATE) algorithm was used to predict tumor purity and the infiltrating status of stromal cells and immune cells [38]. In the current study, we also calculated and compared the ImmuneScore, StromalScore, ESTIMATEScore and tumor purity in high- and low-risk group. To more precisely reflect the immune microenvironment in HCC tissues, the infiltration levels of 16 different kinds of immune cells and relative immune functional scores in high- and low-risk patients were then visualized via R package “GSVA”.

Different from traditional gene function analysis methods, gene set enrichment analysis (GSEA) derives its power through integrating the information from the genes with similar molecular characteristics or biological function, thereby flexibly and comprehensively revealing common biological pathways of the gene group [39]. Based on “GSVA” and “GSEABase” package in R, GSEA was adopted to evaluate different molecular functions and pathways between high- and low-risk group. Furthermore, the threshold of FDR and q value was set as 0.05 and 0.2, respectively.

As one of the important methods of cancer immunotherapy, ICI treatment has revolutionized treatment prospects of various cancer types, prolonging the survival time of many patients and offering more therapy options for cancer patients [40]. According to the expression profile in TCGA–LIHC data set and R package “ggpubr”, the expression levels of some reported immune checkpoint genes and human leukocyte antigen (HLA)-related genes were compared between high- and low-risk groups. In previous study, the tumor immune dysfunction and exclusion (TIDE) score was considered as an effective predictor for anti-PD1 or anti-CTLA4 therapy [41]. Immunophenoscore (IPS) was calculated according to the four main components, including effector cells, immune checkpoints, immunosuppressive cells and major histocompatibility complex molecules, and a higher IPS was related to increased immunogenicity [42]. We, respectively, downloaded TIDE score and IPS from the TIDE platform (http://​tide.​dfci.​harvard.​edu) and The Cancer Immunome Atlas (TCIA) (https://​tcia.​at/​home) and compared them between high- and low-risk patients.

Because publicly accessible cohort focusing on HCC patients’ ICI response was lacking, we could only enroll two immunotherapy cohorts comprising patients with advanced melanoma, GSE91061 data set (anti-PD1 treatment) and PRJEB23709 data set (anti-PD1 treatment or combined anti-PD1/anti-CTLA4 treatment), to externally verify the predictive power of the risk model on response of ICI therapy. GSE91061 cohort and PRJEB23709 cohort included 68 and 120 patients with advanced melanoma, respectively. Of these patients, complete follow-up records and RNA-seq data of 57 patients in GSE91061 cohort and 73 patients in PRJEB23709 cohort were provided. According to the therapeutic response defined by RECIST 1.1 criteria, patients were divided into responders (complete response (CR)/partial response (PR)) and non-responders [stable disease (SD)/progressive disease (PD)]. The similar study design was also shown in previous publications [43, 44].

Half maximal inhibitory concentration (IC50) was usually used in drug sensitivity assessment, and a higher IC50 value was correlated with lower drug sensitivity [45]. From Genomics of Drug Sensitivity in Cancer (GDSC) (https://​www.​cancerrxgene.​org/​), we collected IC50 data of several chemotherapy drug or molecular targeted drug used against HCC. Using R package “oncoPredict”, IC50 value of each drug was visualized and compared between high- and low-risk patients.

In the present study, R software version 4.1.2 and GraphPad Prism 9.5.0 was utilized in data processing, statistical analysis and data visualization. Comparisons between two groups were assessed by Student’s t test or Wilcoxon test. Kruskal–Wallis test was conducted in comparisons among three groups or above. A two-sided P value of less than 0.05 was considered significant.

The flow diagram concerning study design was shown in Additional file 8. The single-cell RNA-seq data were from 3 HCC samples, GSM4955419, GSM4955421, and GSM4955426, in GSE162616 data set. Numbers of detected genes and sequencing depth in each sample were visualized in Fig. 1A, and a strongly positive association based on Pearson’s correlation test was observed between the detected gene numbers and the sequencing depth (cor = 0.88, P value < 0.05, Fig. 1B). Then, the quality control processes were conducted to exclude the cells with low-quality, and 24,653 cells were finally selected out to perform subsequent analyses. In these cells, the standardized gene expression matrices were compared to find out top 1500 highly variable genes (Fig. 1C). The related data were dimensionally reduced by the PCA method, and top 20 PCs were distinguished with P value of less than 0.05 (Fig. 1D). The 20 PCs were then classified into 22 clusters through t-SNE algorithm, which were later translated to known cell types (Fig. 1E, F), including NK cells, T cells, monocytes, macrophages, B cells, hepatocytes and erythroblasts. The corresponding markers in each cell type were summarized in Additional file 9, which included 404 NK cell markers.

Using the genomic data from TCGA–LIHC cohort, the expression matrices of the 404 NK cell markers were obtained. According to the above data, weighted gene co-expression network analysis was performed to screen out the survival-related markers form the 404 NK cell-related genes. The index of scale-free topologies (R²) of 0.9 corresponded to a lowest soft threshold value of 6 (Fig. 2A). The 404 genes were then classified into different modules and eventually formed the blue module, the turquoise module and the grey module (Fig. 2B). The survival-related module was selected using Pearson’s correlation test, and corresponding correlation coefficient and P value of each module were visualized in Fig. 2C. Of these 3 modules, the turquoise module was positively correlated with survival status (cor = 0.14, P value = 0.005), and a total of 245 genes were included in the turquoise module (Additional file 10).

The prognostic performance of the 245 NK cell markers associated with survival status of HCC patients was assessed by univariate Cox regression analysis, and 59 markers were finally identified as prognosis-related markers using TCGA–LIHC data set (all P value < 0.05, Fig. 3). Integrated framework including 77 fundamental or combined machine learning algorithms was used to develop a consensus prognostic model based on the 59 prognosis-related markers. The C-index of each model was calculated in TCGA–LIHC cohort, GSE14520 cohort, GSE76427 cohort and ICGC–LIRI–JP cohort. With a leading average C-index of 0.671 among the 4 cohorts, a combination of LASSO and CoxBoost algorithm was considered as the optimal algorithm to construct prognostic model (Fig. 4). The 59 markers were first enrolled to perform LASSO regression analysis to selected out with the minimum partial likelihood deviance and markers with nonzero LASSO coefficients. These markers were next analyzed by CoxBoost proportional hazards regression, and 11 markers, including KLRB1, CFL1, LDHA, BSG, ATP1B3, SERBP1, UBE2L3, PCBP2, ENO1, OPTN and LMO4, were identified to construct prognostic model. The risk score of each patient was calculated according to the expression level weighted by corresponding regression coefficients.

In TCGA–LIHC cohort, HCC patients in high-risk group showed a worse prognosis than low-risk patients (P value < 0.001, Fig. 5A). Similar survival outcomes were also detected in GSE14520 cohort (P value < 0.001, Fig. 5B), GSE76427 cohort (P value = 0.011, Fig. 5C) and ICGC–LIRI–JP cohort (P value < 0.001, Fig. 5D). The risk score classification, survival status and gene expression of single patient in the above 4 cohorts were also shown. Next, ROC method was utilized to evaluate predictive efficacy of the prognostic signature. In TCGA–LIHC cohort, the AUCs for age, gender, clinical stage and risk score were 0.499 0.513, 0.703 and 0.816, respectively (Fig. 5E). In GSE14520 cohort, the AUCs for age, gender, clinical stage and risk score were 0.570, 0.535, 0.689 and 0.622, respectively (Fig. 5F). In GSE76427 cohort, the AUCs for age, gender, clinical stage and risk score were 0.603, 0.575, 0.753 and 0.662, respectively (Fig. 5G). In ICGC–LIRI–JP cohort, the AUCs for age, gender, clinical stage and risk score were 0.530, 0.421, 0.659 and 0.725, respectively (Fig. 5H). As for time ROC analysis, the 1-year survival, 3-year survival and 5-year survival AUCs in TCGA–LIHC cohort were 0.816, 0.741 and 0.722 (Fig. 5I). The 1-year survival, 3-year survival and 5-year survival AUCs in GSE14520 cohort were 0.622, 0.635 and 0.661 (Fig. 5J). The 1-year survival, 3-year survival and 5-year survival AUCs in GSE76427 cohort were 0.668, 0.662 and 0.617 (Fig. 5K). the 1-year survival, 3-year survival and 5-year survival AUCs in ICGC–LIRI–JP cohort were 0.780, 0.725 and 0.750 (Fig. 5L). These ROC results revealed our signature had a good performance in prognostic assessment of HCC patients. To verify whether the risk score was the independent prognostic indicator for HCC patients, univariate and multivariate Cox regression analysis considering age, gender, clinical stage and risk score were performed in TCGA–LIHC cohort, GSE14520 cohort, GSE76427 cohort and ICGC–LIRI–JP cohort, respectively (Fig. 6A, B). To further expand our results, we also conducted univariate and multivariate Cox regression analysis considering age, gender, race, clinical stage, Child–pugh grade, tumor grade and risk score in TCGA–LIHC data set (Additional file 11A, B), which showed the similar results with Fig. 6A, B. These analyses suggested risk score was one of the independent risk factors for HCC patients’ survival.

To offer clinicians with a more quantitative method for prognostic assessment, prognostic nomograms combining age, gender, clinical stage and risk score were established in TCGA–LIHC cohort. Each patient would get a score from corresponding clinical parameter, and a higher total score was related to a worse survival outcome (Fig. 7A). In 1-year, 3-year and 5-year calibration plots, there was a relatively good fitting between actual model and ideal model (Fig. 7B). The clinical ROC curve indicated the nomogram had a good predictive efficacy with the AUC value of 0.789 (Fig. 7C). Furthermore, the decision curves also indicated the high potential of our nomogram in clinical utility (Fig. 7D). To further enrich our study, we also built up a nomogram and corresponding 1-year, 3-year and 5-year calibration plots considering the 7 clinical parameters, including age, gender, race, clinical stage, Child–pugh grade, tumor grade and risk score in TCGA–LIHC data set (Additional file 12A and B). The AUC value of the nomogram was 0.798, which was comparatively high among these clinical factors (Additional file 12C). Furthermore, the decision curve also revealed the nomogram had a relatively good predictive power for HCC patients, prognosis (Additional file 12D).

To further verify the predictive efficacy of our prognostic signature, other 10 prognostic signatures were enrolled from previous publications [46‐55]. The processes of target gene selection, risk score calculation and risk model establishment all obeyed the corresponding original text, and the genomic and clinical data were from TCGA–LIHC data set. Survival plots and corresponding time ROC curves were listed in Additional file 13 and Additional file 14, respectively (all P value < 0.05). C-indices of the 11 prognostic signatures were also compared and visualized in Fig. 8A, B. A relatively high C-index of 0.730 was obtained from our NK cell-related signature, suggesting an advanced performance.

Mutation data from TCGA–LIHC data set was also downloaded and then visualized to reveal the overall somatic mutation status between the two groups. Both in high-risk group (Fig. 9A) and low-risk group (Fig. 9B), TP53, CTNNB1, TTN and MUC16 were identified as the most frequently mutated genes. Next, the overall TMB levels of the two groups was also calculated and compared, which showed no statistical difference between the two groups (P value > 0.05, Fig. 9C).

We also investigated the correlations between risk score and immune cell infiltration status in HCC. No significant difference was detected in tumor purity between these two groups (P value > 0.05, Fig. 9G), while high-risk patients had lower ESTIMATEScore (P value < 0.05, Fig. 9D) and StromalScore (P value < 0.05, Fig. 9E), but higher ImmuneScore (P value < 0.05, Fig. 9F). More precisely, we next assessed the infiltration levels of 16 kinds of immnocytes according to the ssGSEA algorithm. The infiltrating levels of aDCs, B cells, CD8 + T cells, mast cells, neutrophils, NK cells, T helper cells, Th2 cells and TIL cells in low-risk patients were higher than high-risk patients, while the infiltrating level of macrophages was increased in high-risk patients (all P value < 0.05, Fig. 9H). As for common immune functions, the ssGSEA score of corresponding items between these two groups was also compared. The levels of cytolytic activity, T cell co-stimulation and type II IFN response were decreased in high-risk patients, while APC co-stimulation and MHC class I level were increased in high-risk patients (all P value < 0.05, Fig. 9I).

GSEA was conducted to explore the potential biological functions and pathways of the prognostic model. The 5 most relevant KEGG items of the two groups were shown. High-risk group was mainly enriched in cell cycle, cytokine–cytokine receptor interaction, ECM receptor interaction, hypertrophic cardiomyopathy and neuroactive ligand receptor interaction (Fig. 10A), while low-risk group was primarily enriched in butanoate metabolism, glycine serine and threonine metabolism, linoleic acid metabolism, primary bile acid biosynthesis and tryptophan metabolism (Fig. 10B).

Previous publications emphasized the major role of NK cells in anti-tumor immunity and immunotherapy [23, 56]. Thus, we investigated whether the risk score could predict ICI response of HCC patients. We first compared the expression levels of HLA-related genes in the two groups and found that the expression levels of HLA–DMA, HLA–DQB2, HLA-J, HLA–DQA2, HLA-E, HLA–DMB, HLA–DPA1, HLA-H, HLA-L, HLA-C, HLA–DRA, HLA-A and HLA–DOA were significantly higher in low-risk patients (all P value < 0.05, Fig. 11A). Next, the expression levels of common immune checkpoints between the two groups were also compared. As shown in Fig. 11B, the TMIGD2 expression and the CD40LG expression were increased in high-risk group, while the CD200R1, HHLA2, ICOS, CD200, CTLA4, LGALS9, TNFRSF14, CD40, CD80, HAVCR2, CD86, ICOSLG, VTCN1, CD70, CD276, TNFSF15, TNFSF18, CD274, LAIR1, NRP1, CD28, TNFSF9, TNFRSF8, TNFRSF4, TNFRSF18, CD44, TNFRSF9, TNFSF4 and BTNL2 expression were significantly elevated in low-risk group (all P value < 0.05, Fig. 11B). Subsequently, we also compared TIDE score in the two groups and detected that TIDE score in high-risk group was significantly higher (P value < 0.001, Fig. 11C). Moreover, we also analyzed IPS–CTLA4 + PD1 blocker, IPS–CTLA4 blocker and IPS–PD1 blocker in the two groups. No significant statistical difference was detected in IPS–CTLA4 + PD1 blocker (P value = 0.081, Fig. 11D), while IPS–CTLA4 blocker (P value = 0.018, Fig. 11E) and IPS–PD1 blocker (P value = 0.031, Fig. 11F) were higher in low-risk patients. From the above results, patients in low-risk group might be more likely to benefit from ICI treatment.

In the abovementioned analyses, patients in low-risk group were detected to be more reactive to ICI treatment. Next, we also enrolled the two immunotherapy cohorts (GSE91061 cohort and PRJEB23709 cohort) to externally validate this result. Although there was no significant statistical difference of risk score between responders (CR/PR group) and non-responders (SD/PD group) in GSE91061 cohort (P value = 0.11, Fig. 12A), high-risk patients were still correlated with an adverse prognosis (P value = 0.009, Fig. 12C). The AUC values of 1-year, 2-year and 3-year survival rate were 0.784, 0.723 and 0.637, respectively (Fig. 12E). As for PRJEB23709 cohort, a significant lower risk score was detected in responders (CR/PR group) (P value = 5.1e−04, Fig. 12B), and high-risk patients showed a worse prognosis than those in low-risk group (P value < 0.001, Fig. 12D), which was similar with ours. In addition, the AUC values of 1-year, 2-year and 3-year survival rate in PRJEB23709 cohort were 0.750, 0.804 and 0.871, respectively (Fig. 12F).

The IC50 levels of some common targeted or chemotherapy drugs used against HCC were further compared between high- and low-risk patients. As shown in Fig. 13, IC50 values of irinotecan (P value = 6.1e−08, Fig. 13A), vorinostat (P value = 7.4e−06, Fig. 13B), axitinib (P value = 4.3e−06, Fig. 13C), cytarabine (P value = 4.8e−08, Fig. 13D), oxaliplatin (P value = 4.7e−05, Fig. 13E), leflunomide (P value = 0.0012, Fig. 13F), cisplatin (P value = 0.0019, Fig. 13G), topotecan (P value = 2.9e−06, Fig. 13H), mitoxantrone (P value = 0.02, Fig. 13I), sorafenib (P value = 1.5e−07, Fig. 13J), fludarabine (P value = 3e−06, Fig. 13K) and gemcitabine (P value = 6.6e−07, Fig. 13L) in low-risk group were significantly lower than those in high-risk group, which revealed that low-risk patients might be more reactive to targeted therapy and chemotherapy.