An immune-related signature for optimizing prognosis prediction and treatment decision of hepatocellular carcinoma

To identify and verify the prognostic significance of IGS risk model, we included three HCC publicly available cohorts (Additional file 1: Table S1). RNA-seq profiles for 363 HCC samples and 50 normal liver samples were obtained from TCGA-LIHC dataset (https://​portal.​gdc.​cancer.​gov/​repository), which were subsequently converted into the Transcripts Per Million (TPM) values. Survival data were achieved from TCGA Pan-Cancer Clinical Data Resource (TCGA-CDR). TPM normalized expression profiles of 226 normal liver tissues were downloaded from GTEx database. Liver cancer-RIKEN, JP project (LIRI-JP) included Gene profiles, somatic mutation data and clinical information of 229 HCC samples were downloaded from the International Cancer Genome Consortium (ICGC) portal (https://​dcc.​icgc.​org/​projects/​LIRI-JP). Normalized gene profiles and survival data of Chinese HCC (CHCC) cohort with 159 HCC tissues and adjacent tissues patients of Zhongshan Hospital (Shanghai, China) were downloaded from NODE (https://​www.​biosino.​org/​node) [13]. Somatic mutations were analyzed using R package ‘maftools’. Baseline patient characteristics of the three cohorts above were summarized in Additional file 1: Tables S2, S3 and S4. Two immunotherapeutic cohorts were included in this study to test the efficacy of IGS score in predicting immunotherapeutic response. The gene expression and clinical information of patients with locally advanced or metastatic urothelial carcinoma administrated with anti-PD-L1 antibody atezolizumab were obtained from IMvigor210 cohort (http://​research-pub.​Gene.​com/​imvigor210corebi​ologies) [14]. The raw gene expression data of IMvigor210 were normalized and transformed into TPM values using R packages “arrayQualityMetrics”, “voom”, and “limma”, respectively. The gene profiles and detailed clinical information of metastatic melanoma patients treated with an anti-PD-1 antibody pembrolizumab were achieved from GSE78220 cohort. The background adjustment, normalization, and logarithmic processing were performed using R package “affy” [15]. The list of immune-related genes was achieved from ImmPort database (https://​immport.​org/​shared/​home).

The differentially expressed genes (DEGs) between 363 HCC tissues (TCGA-LIHC database) and normal liver tissues (50 samples from TCGA-LIHC database and 226 samples from GTEx database) were identified using the R package “DESeq2” with a threshold of P < 0.05 and Fold change > 2. The entire TCGA HCC cases (n = 363) were randomly divided into a training cohort (n = 255) and testing cohort (n = 108) in a ratio of 3:1. Following overlapping the DEGs with immune-related genes from ImmPort database, the immune-related hub genes were further screened using univariate Cox regression analysis with a threshold of P < 0.05. Subsequently, the LASSO algorithm and 200-fold cross-validation were performed to construct the immune genes-based signature (IGS) model using R package “LASSO” and “cv.glmnet” in the training cohort. The IGS score of each sample was calculated as IGS score = ∑_i=1,2,…n (coefficients(gene_i) × gene_i expression). Based on the median value of IGS score, the patients in each cohort were stratified into high-risk and low-risk groups. The efficacy and predictive capacity of IGS was evaluated by the time-dependent receiver operating characteristic (ROC) curve and Kaplan–Meier (KM) curve by performing R package “survivalROC” and R package “survival”, respectively. Univariate and multivariate Cox regression analyses were performed to evaluated the IGS score as an independent prognostic factor for HCC OS. In additional to TCGA internal validation cohort and entire validation cohort, the IGS model was also validated in ICGC cohort and external CHCC cohort. A prognostic nomogram model was then established with calibration curve based on the IGS score and clinical parameters by performing “rms” and “nomogramEx” R packages. Four previously published immune-related signatures were referred to compare the predictive efficacy and accuracy in CHCC cohort [16‐19].

The infiltration of immune cells was quantified by performing CIBERSORT algorithm [20]. The infiltration scores of different groups were calculated by ImmuCellAI algorithm [21]. The correlation of IGS score with immune-related pathways was evaluated using single sample gene set enrichment analysis (ssGSEA). Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) were performed to conduct enrichment analysis. GSEA was performed to detect the difference in DEGs between the high-risk and low-risk group in the enrichment of the MSigDB Collection (h.all.v7.2.symbols.gmt). Adjusted P-value less than 0.05 was considered statistically significant. Somatic variant analysis was used to investigate differentially mutated genes associated with the ICI response by performing “maftools” R package [22].

Gene profiles of cancer cell lines were downloaded from the Broad Institute-Cancer Cell Line Encyclopedia (CCLE) project (https://​portals.​broadinstitute.​org/​ccle/​). Drug sensitivity data of CCLs were obtained from the Cancer Therapeutics Response Portal (CTRP v.2.0, https://​portals.​broadinstitute.​org/​ctrp, 481 compounds) and PRISM Repurposing dataset (19Q4, https://​depmap.​org/​portal/​prism, 1448 compounds). Missing area under the dose–response curve were imputed using Knearest neighbor (k-NN). Before imputation, compounds with more than 20% of missing data were excluded. Molecular profiles achieved from CCLE dataset were used for further CTRP and PRISM analyses.

The statistical tests analyses in this study were performed in R v3.6.1 software (Vienna, Austria). Student’s t-test, one-way analysis of variance, Wilcoxon rank-sum test or Kruskal–Wallis test were performed to compare the differences of continuous variables. The chi-square test was used for comparison of categorical variables. Correlation analysis was computed by performing Pearson algorithm. P-value < 0.05 was considered of statistical significance.

By comparing 363 HCC cases with 50 normal liver tissues in TCGA data portal and 226 normal tissues in GTEx database, 2371 significantly up-regulated genes and 544 down-regulated genes were identified with a threshold of P < 0.05 and Fold change > 2 (Additional file 1: Fig S1A, B). Following intersection to 1811 immune-related genes derived from the ImmPort database, 251 immune-related hub genes were extracted from the DEGs (Additional file 1: Fig S1C). Enrichment analysis indicated that these genes were implicated in immune and TME-related pathways including cell chemotaxis, antigen processing and presentation, receptor ligand activity, signaling receptor activator activity, cytokine activity, MHC protein complex, cytokine-cytokine receptor interaction, and chemokine signaling pathway (Additional file 1: Fig S1D, E).

Subsequently, HCC cases of TCGA dataset (363) were randomly divided into a training cohort (255) and testing cohort (108) in a 7:3 ratio. Of these immune-related hub genes, 19 candidate genes were further defined as significantly relevant with OS of HCC patients by performing Univariate Cox regression analysis (threshold: P < 0.05, HR > 1.25 or < 0.75; Fig. 1A) in the TCGA training cohort. The LASSO Cox algorithm was then conducted to identify the most robust prognostic genes among the immune-related DEGs in the TCGA training cohort. Following 200-fold cross-validation for variable selection, an optimal λ value of 8 was selected (Fig. 1B, C). Then, 8 Immune-related genes with individual coefficients were assembled to establish an Immune-related gene signature (Additional file 1: Fig S2A, B). Based on the IGS, an immune-related prognostic risk score normalized to Z-score was calculated for each patient in the TCGA training cohort and testing cohort (Fig. 1D, E).

Then, we evaluated the prognostic performance in different HCC cohorts. Based on the median value of the IGS score, each HCC cohort was stratified into a high-risk group and a low-risk group. As illustrated in Fig. 2A, HCC patients in high-IGS-risk group exhibited worse OS than that of low-IGS-risk group, which was coherently determined in TCGA training cohort (HR = 3.35, 95% CI = 2.19–5.13, P < 0.0001), TCGA internal testing cohort (HR = 2.25, 95% CI = 1.19–4.23, P = 0.01), and ICGC cohort (HR = 2.97, 95% CI = 2.09–4.22, P < 0.0001). Consistently, the time-dependent ROC curves indicated that the IGS risk model also displayed excellent AUC in the three cohorts at the time point of 1 year (training cohort, 0.701; testing cohort, 0.73; ICGC cohort, 0.727), 3 years (training cohort, 0.756; testing cohort, 0.727; ICGC cohort, 0.672), and 5 years (training cohort, 0.701; testing cohort, 0.658; ICGC cohort, 0.681), suggesting a favorable predictive capacity of the IGS risk model in short- and long-term follow-up of HCC patients (Fig. 2B).

In addition, the CHCC cohort, as an external independent HCC cohort, was used to further validate the prognostic value of the IGS risk model (Fig. 2C, D). HCC patients with high IGS risk score had poorer RFS (HR = 1.88, 95% CI = 1.15–3.06, P = 0.01) and OS (HR = 3.79, 95% CI = 2.23–6.44, P < 0.0001) than patients with low IGS risk score. It is worth noting that the time-dependent ROC curve also indicated that the IGS risk model had satisfactory performance in predicting OS in CHCC cohort patients at different time periods (2 year, 0.725; 3 year, 0.73; 4 year, 0.749). Subsequently, Univariate and Multivariate Cox regression analyses were conducted in the pooled TCGA cohort including IGS risk score and 5 clinical variables (Age, Gender, Grade, TNM stage, and TP53 status). The results revealed that TNM stage (P < 0.001) and IGS risk score (P < 0.001) were independent risk factors for OS of HCC patients (Fig. 2E). For the CHCC cohort, Age, Gender, tumor size, TP53 status, IGS risk score, and BCLC stage were enrolled into the COX regression analyses. It arrived at a similar conclusion that the TNM stage (P = 0.039) and IGS risk score (P = 0.015) were independent risk factors for HCC prognosis (Fig. 2F).

Subsequently, we conducted stratification analyses to evaluate the prognostic significance of the IGS risk model for HCC patients. The pooled TCGA or CHCC cohort was stratified into sub-groups based on clinical features such as Age (> = 60 or < 60), Gender, Stage (I/II or III/IV), and TP53 status (wide-type or mutated). In both of the two independent cohorts, the IGS risk model could profoundly predict the OS of HCC patients in all sub-groups (Additional file 1: Fig S3A, B). It suggested that IGS retained its prognostic value in differentiating high risk cases in different subgroups.

Based on the prognostic performance of the IGS risk model, we established a nomogram including IGS risk score and clinicopathological features such as age, tumor size, BCLC stage and gender in the CHCC cohort (Fig. 3A). As demonstrated in the calibration plot, the prediction lines of the nomogram for 1-, 2-, and 3-year survival probability were extremely close to the 45-degree reference line, suggesting the profoundly predictive accuracy for HCC patients (Fig. 3B). In addition, in contrast to the BCLC stage, the nomogram exhibited superior performance in predicting the OS of HCC patients (Fig. 3C). Furthermore, we compared the performance of the IGS risk model with 4 previous immune-related signatures in the CHCC cohort. At the time point of 3 year, the AUC of IGS (0.736) was obviously higher than that of Liu’s 7-gene signature (0.609), Wang’s 9-gene signature (0.535), Dai’s 11-gene signature (0.647), and Wang’s 13-gene signature (0.707), respectively. Moreover, the net reclassification improvement (NRI) was conducted to quantify the improvement in predictive performance, which showed a 10.1% improvement to Wang’s 13-gene signature by performing the IGS risk model (Fig. 3D). Consistently, with an NRI of 17.7%, the IGS risk model also had better performance in 4 year’s follow-up (0.749 vs. 0.556, 0.534, 0.638, and 0.658) compared with Wang’s 13-gene signature (Fig. 3E).

With a threshold of |log2(Fold Change)|≥ 1 and P < 0.01, 187 DEGs were identified in high- IGS risk samples and low- IGS risk samples of the TCGA cohort, respectively (Additional file 1: Fig S4A). Subsequently, these DEGs were subjected to GO enrichment analysis to predict their potential roles in biological processes. The result showed that they were implicated in pathways including response to stimulus, immune system process, metabolic process, cellular process, and multi-organismal process (Additional file 1: Fig S4B). In addition, GSEA was also conducted to investigate the IGS genes-related pathways in TCGA pooled cohort. The gene sets of the IGS-high samples were mainly enriched in DNA repair, E2F target, G2M checkpoint, and MYC targets (Additional file 1: Fig S4C), while the IGS-low group was correlated with IL-6/JAK/STAT3 signaling, inflammatory response and interferon α/γ response (Additional file 1: Fig S4D). Next, we identified the top 10 most frequently mutated genes in the IGS risk score-stratified subgroups. In both subgroups, the mutation rates of TP53, TTN, CTNNB1 and MUC16 were higher than 15%. The mutations of the ABCA13, OBSCN and SPTA1 genes were more frequent in the IGS-high subgroup. In contrast, the mutations of LRP1B, MUC4 and RYR1 were more commonly observed in the IGS-low subgroup (Fig. 4A). Accordingly, the IGS score of patients with OBSCN, TP53, DNAH8, SPTA1, or TTN mutation was significantly higher than that of wild-type patients (Fig. 4B). Interestingly, as shown in Fig. 4C, TP53 ranked the highest mutation frequency in both of the IGS-high and -low sub-group. More specifically, distinct mutation spots of TP53 were detected in the two sub-groups according to the lollipop analysis (Fig. 4D, E). Furthermore, the tumor mutational burden (TMB) was significantly elevated in patients with higher IGS risk score (P = 0.027; Fig. 4F). Moreover, the IGS score was negatively correlated with PDL1 expression in TCGA data portal (Additional file 1: Fig S5A). Interestingly, the protein expressions of MLH1 and MSH6, two typical MSI genes, were significantly elevated in high IGS risk group in CHCC cohort, suggesting the potentially promoting roles in TMB (Additional file 1: Fig S5B). As illustrated in Fig. 4G, patients with high IGS risk score/high TMB had shorter OS than patients with low IGS risk score/low TMB and high IGS risk score/low TMB (P < 0.001), respectively. In addition, for patients with high TMB, higher IGS score also suggested a poorer prognosis (P = 0.038; Fig. 4H).

Given the robust performance of IGS risk model in prognosis for HCC, we further investigated the underlying immune-related characteristics regarding IGS genes. Based on the CIBERSORT analysis in 363 TCGA HCC patients, the differences in immune infiltration of 28 immune cell types in high- and low- IGS risk groups are illustrated in Fig. 5A. Activated B cells, Memory B cells, immature B cells, Natural killer T cells, Natural killer cells, Activated CD8 + T cells, effector CD8 T cells, Natural killer cells were abundant in the low-risk group and negatively correlated with the IGS risk score (Fig. 5B). Furthermore, we calculated the abundance of immune cells in the TME between the high- and low- IGS risk subgroups by performing ImmuCellAI analysis. As shown in Fig. 5C, patients in the high-risk group had a lower proportion of B cells, dendritic cells (DCs), central memory T (Tcm) cells, T follicular helper (Tfh) cells, CD4/CD8 T cells, CD56 natural killer (NK) cells, Eosinophil cells, neutrophils, natural regulatory T cells (nTreg), effector memory T (Tem) cells, and T regulatory type 1/2/17 (Tr1/2/17) cells, suggesting an immunosuppressive microenvironment in HCC patients with high-IGS risk.

Then, we evaluated the prognostic value of IGS risk model in 13 independent TCGA cancer cohorts. As presented in Fig. 6A, in addition to LIHC (HR, 1.90; P < 0.001), IGS risk model displayed an excellent predictive capacity for cancer types such as ACC (HR, 1.53; P = 0.019), KIRC (HR, 1.59; P < 0.001), LIHC (HR, 1.90; P < 0.001), and THYM (HR, 2.88; P = 0.016). Interestingly, the IGS risk model also presented a robust prognostic value for pooled pan-cancer patients (HR, 1.90; P < 0.001). To further validate the potential capacity of IGS risk model, we chose a cohort administrated with anti–PD-L1 immunotherapy (GSE78220). High-IGS risk score was detected in non-responders rather than responders (Fig. 6B). In parallel, the low-IGS risk subgroup had a higher proportion of responders to anti–PD-L1 immunotherapy, suggesting a role of IGS risk score in predicting the response of PD-L1 treatment (Fig. 6C). Then we investigated the efficacy of IGS risk model in IMvigor210 urothelial cancer cohort with immunotherapy information. The IGS risk score was positively correlated with regulatory T cells, while it was negatively associated with effector memory/activated CD8 T cells and immature/activated B cells (Fig. 6D). As shown in Fig. 6E, patients with high-IGS risk score suffered from a significantly shorter OS than those with low-IGS risk score in IMvigor210 urothelial cancer cohort (P = 0.0017). In contrast to IGS-low subgroup, high IGS risk score led to a lower microsatellite instability (MSI) score and a higher T cell exclusion score, while there was no significant difference for T cell dysfunction (Fig. 6F). TIDE was further performed to evaluate the potential clinical efficacy of immunotherapy in different IGS subgroups. Patients with high-IGS risk score had a higher TIDE prediction score, which implied a higher possibility of immune evasion and less efficacy of ICB therapy (Fig. 6G). The ratio of clinical response to anti-PD-L1 immunotherapy was subsequently evaluated in high- or low-IGS risk subgroups in the IMvigor210 cohort. Compared with high-risk subgroup, Low-IGS risk resulted in a higher proportion of complete response [CR]/partial response [PR] and lower stable disease [SD]/progressive disease [PD] (Fig. 6H). Consistently, patients with status of SD/PD had a higher IGS risk score than patients with CR/PR (P = 0.048, Fig. 6I). In addition, ROC curves demonstrated that the AUC of the IGS risk score (0.64) was higher than that of the TIDE score (0.57) and TIS (0.585) at 20 months in the IMvigor210 cohort, indicating a better performance in predicting the efficacy of anti-PD-L1 immunotherapy (Fig. 6J).

The analyses above indicated that patients with high IGS risk score might have poor prognosis and a relatively suppressive immune status. Then we tried to screen potential agents with higher drug sensitivity for high-IGS risk patients based on the CTRP and PRISM datasets. After pre-exclusion, 1929 compounds (CTRP, 481; PRISM, 1448) were included for subsequent analyses (Fig. 7A). The candidate therapeutic agents screening was conducted with a two-step strategy (Fig. 7B). First, we performed differential drug response analysis between high IGS risk score sub-group and low IGS risk score sub-group to identify candidate compounds with lower estimated AUC values in high IGS risk score patients. Next, Spearman rank correlation analysis between AUC and IGS risk score was conducted to select potential compounds with negative correlation coefficient (r <  − 0.30/CTRP or r <  − 0.30/PRISM). Based on the analyses above, we obtained the candidate agents clofarabine and dasatinib from CTRP (Fig. 7C), while 6 compounds (epothiloneb, LY2606368, irinotecan, ispinesib, vinblastine and vindesine) from PRISM (Fig. 7D). The analyses above suggested that the 8 compounds might be candidates with higher drug sensitivity in high IGS risk score patients. Furthermore, multiple perspective analyses were conducted by comprehensive evaluation of clinical status, experimental evidence, differential mRNA/protein levels of candidates’ drug targets, and CMap score. Ultimately, vindesine, ispinesib and dasatinib were identified as therapeutic agents for HCC patients with high IGS risk score (Fig. 7E).