Prognostic model of patients with liver cancer based on tumor stem cell content and immune process

DEGs related to immune processes

In order to find important genes closely related to the immune process in LIHC patients, we used StromalScore and ImmuneScore in the Estimate algorithm to evaluate the stromal cell content and immune cell content of the samples. After that, we grouped the samples according to the median of StromalScore and ImmuneScore, screened the DEGs between the high score group and the low score group, and then intersected the DEGs of the two scores. StromalScore's high group has 1672 up-regulated genes (Supplementary Figure 1A), ImmuneScore's high group has 1421 up-regulated genes (Supplementary Figure 1B), and two sets have 1078 intersection genes (Figure 1B). StromalScore's high group has 222 down-regulated genes (Supplementary Figure 1A), ImmuneScore's high group has 160 down-regulated genes (Supplementary Figure 1B), and they have 62 intersection genes (Figure 1C). ImmuneScore has significant differences (Supplementary Figure 1C, P=0.029) in different stages, but StromalScore does not show significant difference (Supplementary Figure 1D, P=0.067) in different stages.

Results related to tumor stem cell scoring

The content of tumor stem cells in tumor tissues had a strong correlation with tumor recurrence, often leading to poor prognosis. The prediction results of tumor stem cell content (mRNAsi) of TCGA samples by Malta et al were used to evaluate the content of tumor stem cells in samples [16]. The mRNAsi was analyzed in different clinical traits. A significant difference (Wilcoxon test, P < 0.001) in mRNAsi was found between LIHC samples and normal samples (Figure 2A). mRNAsi did not differ significantly in different ages, genders, and tumor stages (Supplementary Figure 2A–2C). In different tumor grades, mRNAsi has significant differences (Figure 2B, Wilcoxon test, P=0.006), and with the increase of grade, the mRNAsi score gradually increases, which shows that mRNAsi can fully reflect the tumor cell differentiation status. The data of patients with LIHC obtained from TCGA were grouped according to the median mRNAsi value, and the Kaplan–Meier test was performed for survival analysis to explore the relationship between mRNAsi and patient prognosis. Patients with low mRNAsi scores had a better prognosis (Figure 2C, P=0.006). The results showed that it was reasonable to use mRNAsi to evaluate the tumor stem cell content in LIHC samples and normal samples.

DEGs in 50 normal tissues and 374 LIHC tissues were first screened to find genes in LIHC samples that were closely related to disease occurrence and tumor stem cell scoring. A total of 7273 upregulated genes and 394 downregulated genes were obtained in tumor tissues (Supplementary Figure 2D and 2E). Then, these DEGs were used for WGCNA and mRNAsi as the phenotype for analysis. In this study, when the soft threshold β = 13, the gene regulatory network better met the scale-free network (Figure 2D). First, 7667 DEGs were used to perform hierarchical clustering on 374 LIHC samples. Samples with “Height” greater than 1500 were regarded as outliers and excluded (Supplementary Figure 3A, 3B). Then, the genes were clustered and the gene modules were merged. The “mergeCutHeight” value was set to 0.4 to merge the different gene modules obtained (Supplementary Figure 4A). Figure 2E shows the results of gene clustering. “Dynamic Tree Cut” shows the gene modules before merging, and “Merged dynamic” shows the gene modules after merging. The correlation between the mRNAsi and the first principal component of the gene modules (Pearson's correlation coefficient) is shown in Figure 2F. The pink module (–0.41, P < 0.001) and the salmon module (–0.73, P < 0.001) were closely related to mRNAsi. Moreover, both correlation coefficients were negative, indicating that a high expression of genes in these modules meant lower mRNAsi scores. The relationship between gene expression, phenotype, and the first principal component of the pink and salmon modules are shown in Supplementary Figure 4B and 4C.

GO and KEGG enrichment analysis

The GO and KEGG enrichment analysis were applied to further evaluate the functions and mechanism of these DEGs. The DEGs obtained from the ESTIMATE algorithm were obviously associated with immune-related biological processes, including leukocyte migration and regulating cell surface receptor signaling pathway in immune response; Cellular component analysis showed that these DEGs significantly enriched in extracellular matrix, side of membrane and so on; Molecular function also indicated these genes involved in immune process, such as antigen binding, glycosaminoglycan binding and cytokine receptor activity (Supplementary Figure 5A). Meanwhile, KEGG analysis suggested DEGs were enriched in immune related pathway, like cytokine-cytokine receptor interaction, chemokine signaling, etc. (Supplementary Figure 5B).

WGCNA indicated that the salmon and pink modules were the significant gene modules related to tumor development. The GO analysis showed that genes in the salmon module were associated with glomerulus development, regulation of angiogenesis and vasculature development (Figure 3A, 3B) while genes in the pink module were predominantly involved in cellular protein localization and calcium ion transport in biological process (Figure 3C, 3D). Remarkably, angiogenesis related to cancer development and calcium homeostasis had an impact on cancer proliferation and metastasis [28].

Establishment of the prognostic model

The coefficients of the pink and salmon modules obtained by WGCNA were negative (Figure 2F), indicating that the high expression of genes in these two modules represented a decrease in tumor stem cells. The high expression of upregulated genes obtained by the ESTIMATE algorithm meant an increase in the content of immune cells. The reduction in stem cells and the increase in immune cell content were often closely related. Therefore, for finding genes closely related to tumor stem cell content and tumor immune processes, 210 genes in the salmon and pink modules were intersected with 1078 upregulated genes in the ESTIMATE algorithm and 28 key genes were obtained. These 28 genes were used in the construction of subsequent prognostic models. Besides these 28 genes, 14 types of clinical information were also included in the prognostic model to make the prognostic model more informative. A total of 225 patients were then grouped into a training set (n = 114) and a test set (n = 111). The clinical baseline data and grouping of patients are shown in Table 1. The differences in clinical variables between the training set and the test set were examined to illustrate the rationality of random grouping. The chi-square test was used for discrete variables, while the Wilcoxon test was used for continuous variables. Albumin showed a significant difference between the training and test sets (Table 1, P < 0.05), while the other variables showed no significant difference. It was reasonable to have few different variables when grouping because a variety of clinical information was included.

For 28 + 14 variables, lasso regression was first performed to eliminate collinearity between the variables. When the penalty coefficient λ = 0.064, the equation had the smallest error (Supplementary Figure 6A), and the coefficients of nine variables were not 0 (Supplementary Figure 6B). Nine variables were used to build a multivariate Cox regression model, and the independent variable selection method was the back-off method. Finally, eight variables were included in the Cox regression model (Figure 4A). model contains clinical information and genes related to the immune process and CSC, we call this model Immunity and CSC related prognosis (ICRP) score. The formula of the ICRP score was as follows:

Where Age is in year; Grade is an integer from 1 to 4; Asian takes values 1 (means “yes”) and 0 (means “no”); and Bilirubin is in mg/dL; the levels of the four genes are the normalized values of FPKM.

The training and test sets were grouped according to the median value (1.08) of the ICRP score in the training set. All patients with an ICRP score less than 1.08 were in the low-risk group, while all patients with an ICRP score greater than 1.08 were in the high-risk group. Moreover, Kaplan–Meier curves were plotted for survival analysis. The high ICRP score had less survival time in both the training set (Figure 4B, P < 0.001) and the test set (Figure 4C, P < 0.05). In the training set, the 1- and 3-year OS rate was 94.2% (95% CI = 88.0–100.0) and 85.5% (95% CI = 70.7–98.1) in the low-risk group and 76.6% (95% CI = 66.2–88.6) and 56.9% (95% CI = 44.0–73.6) in the high-risk group, respectively. In the test set, the 1- and 3-year OS rates were 92.2% (95% CI = 85.1–99.8) and 81.8% (95% CI = 70.8–94.6) in the low-risk group and 84.8% (95% CI = 75.6–95.1) and 48.9% (95% CI = 31.3–66.5) in the high-risk group, respectively.

The patients were sorted according to the ICRP score, and the distribution of patient survival status was plotted. In the training set (Figure 4D) and test set (Figure 4E), the survival time of patients gradually decreased and the number of death cases gradually increased with the increase in the ICRP score.

Verification of the ICRP score

In order to further verify the effectiveness of our prognostic model and compare it with other method, we used ICRP score, AJCC stage and ALBI score in the training set and test set to predict the survival status of patients at 1, 3, and 5 years, respectively, and plotted the receiver operator characteristic curves. The area under the curve (AUC) of ICRP score in the test set for 1, 3, and 5 years was 0.708, 0.723, and 0.765 respectively (Figure 5A–5C); the AUC of AJCC stage in the test set for 1, 3, and 5 years was 0.584, 0.660, and 0.701, respectively (Figure 5D–5F); the AUC of ALBI score in the test set for 1, 3, and 5 years was 0.654, 0.615, and 0.583, respectively (Figure 5G–5I). The AUCs of ICRP score in the training set for 1, 3, and 5 years was 0.840, 0.801, and 0.824 respectively (Supplementary Figure 6C–6E); the AUCs of AJCC stage in the training set for 1, 3, and 5 years was 0.592, 0.599, and 0.582 (Supplementary Figure 6F–6H); the AUCs of ALBI score in the training set for 1, 3, and 5 years was 0.641, 0.549, and 0.529 (Supplementary Figure 6I–6K). It can be seen that ICRP score’s prediction ability was better than AJCC stage and ALBI score. To further demonstrate this, the C-index of the ICRP score, AJCC stage, and ALBI score in the training and test sets were calculated. The related results are shown in Table 2. It was obvious that the C-index of the ICRP score was significantly larger than the C-index of the AJCC stage and ALBI score in the training set (P < 0.05) and the test set (P < 0.05). Carter et al [27] used the TCGA data to establish a signature that could measure chromosomal instability in tumor cells based on the expression of 25 genes, namely CIN25. They proved that CIN25 was an effective prognostic indicator for many cancers [29, 30]. The CIN25 score of each sample was calculated and the patient's prognosis was predicted. The C-index of CIN25 was significantly lower than that of the ICRP score in the training set (P < 0.05, Table 2) and test set (P < 0.05, Table 2). The aforementioned results showed that the proposed prognostic model was effective.

In addition, the genes selected in the ICRP score might have false-positive results. Therefore, a sensitivity test was conducted on the genes in the ICRP score. We used 28 randomly selected genes from DEGs in LIHC patients and 14 kinds of clinical information to construct a new prognostic model applying the same process and parameters, and repeated three times. The C-indexes of the three reconstructed prognostic models were calculated and compared with that of the ICRP score. The results of the ICRP score for the training and test sets were significantly greater than those of the three models (P < 0.05, Table 3), indicating that the genes selected using the mRNAsi score, WGCNA, and ESTIMATE methods were credible.

Next, qPCR experiments were used to check the change in gene expression level in normal cell line (THLE3) and LIHC cell line (SNU-423). As shown in Figure 6A, gene expression of KLHL30 and PLN significantly changed, while no significant difference was observed in the expression levels of LYVE1 and TIMD4 (Supplementary Figure 7). To some extent, the results of qPCR proved the difference in the expression of these genes in different cell lines. Moreover, the Human Protein Atlas database was used to further understand the functions of these significant genes in the ICRP score model. The related IHC results showed that the expression of PLN was upregulated in tumor tissues while the expression of LYVE1 and TIMD4 were down-regulated (Figure 6B–6D, KLHL30 was not found in this database), which indicated these genes do have expression variation in protein level during the development of liver cancer. Detailed information on IHC results was shown in Supplementary Table 2. Furthermore, the study explored whether the aforementioned significant factors had a strong correlation with the ICRP score. Hence, the chi-square test for discrete variables and U test for continuous variables were performed in both training and test sets. Age, Asian, bilirubin, and PLN had a strong difference between high–ICRP score and low–ICRP score groups, while grade was significant in the test set than in the training set (Figure 6E–6H). To observe the relationship of these four genes and immune infiltration, TIMER database (https://cistrome.shinyapps.io/timer/) was used to calculate the correlation between the genes and different immune cells to observe the relationship of these four genes with immune infiltration. As shown in Figure 7, KLHL30 expression was positively associated with the infiltration of B cells (Cor = 0.122, P = 0.023), CD8+ T cells (Cor = 0.228, P = 2e-05), CD4+ T cells (Cor = 0.398, P = 1.75e-14), macrophages (Cor = 0.421, P = 4.34e-16), neutrophils (Cor = 0.384, P = 1.52e-13), and dendritic cells (Cor = 0.271, P = 3.94e-07). The PLN expression was closely associated with the infiltration level of CD8+ T cells (Cor = 0.171, P = 0.001), CD4+ T cells (Cor = 0.326, P = 6.01e-10), macrophages (Cor = 0.236, P = 1.03e-05), neutrophils (Cor = 0.208, P = 0.0001), and dendritic cells (Cor = 0.247, P = 4.21e-06). The LYVE1 expression was positively related to the infiltration of CD8+ T cells (Cor = 0.176, P = 0.001), CD4+ T cells (Cor = 0.133, P = 0.013), macrophages (Cor = 0.303, P = 1.14e-08), neutrophils (Cor = 0.272, P = 3.03e-07), and dendritic cells (Cor = 0.164, P = 0.002). A positive correlation was found between the TIMD4 expression level and the infiltration of B cells (Cor = 0.312, P = 3.43e-09), CD8+ T cells (Cor = 0.361, P = 6.03e-12), CD4+ T cells (Cor = 0.198, P = 2.26e-4), macrophages (Cor = 0.303, P = 1.14e-08), neutrophils (Cor = 0.293, P = 3.02e-08), and dendritic cells (Cor = 0.397, P = 2.75e-14). The results indicated that the aforementioned four genes related to prognosis had a strong correlation with immune process, explaining why these genes could be used as biomarkers in LIHC prognosis. Previous studies [31] showed that the number of activated monocytes and plasma cells decreased and the numbers of B cells, CD4+ T cells, and CD8+ T cells increased in LIHC, compared with the healthy liver. This was consistent with the results of the present study. It is generally believed that B cells can be used as antigen-presenting cells to induce CD4+ T cell–dependent CD8+ memory T cells [32], thereby helping to control tumor invasion and metastasis.

Data collection and screening of differentially expressed genes

Liver cancer patient data from The TCGA (https://www.cancer.gov/tcga) database was used for analysis. RNA-seq data from 374 cancer tissue samples and 50 control tissue samples were obtained and then standardized by “Fragments Per Kilobase per Million” (FPKM). For genes with duplicate records, the average value of gene expression was calculated as the final value. In this process, the “limma” package in the R language was used. Genes whose average expression in all samples was less than 0.2 were excluded to make the screening of differentially expressed genes (DEGs) more reasonable. The Wilcoxon test was used to screen DEGs; genes with |log₂foldchange| >1 and P <0.05 were identified as DEGs.

Determination of tumor stem cell score

Malta et al [16] used one-class logistic regression machine learning algorithm to extract transcriptomic and epigenetic feature sets derived from nontransformed pluripotent stem cells and their differentiated progeny. They provided data on tumor stem cell scores from samples in the TCGA database. “mRNAsi” was the result based on all the expression profile data; “EREG-mRNAsi” was the result based on the expression of genes related to stem cell epigenetic regulation. These two scores ranged from 0 to 1, which was close to 1, indicating that the lower the degree of cell differentiation, the stronger the characteristics of stem cells. In this study, the samples were evaluated using mRNAsi scores, and subsequent WGCNA was performed.

WGCNA of DEGs

Corresponding studies showed that gene regulatory networks obeyed scale-free networks. The WGCNA method [17] was proposed to meet this requirement. This method was used to find the significant gene modules in the specific genome profile. Different from the “hard cutoff,” WGCNA obtained the scale-free network by calculating the power value β, also known as “soft cutoff.” The Pearson correlation coefficient was calculated to construct gene co-expression matrix (1), and then this matrix was transformed into adjacent matrix using exponential adjacent equation (2). Topological overlap matrix was used to obtain the degree of association between genes (3) by setting gene module parameters: MaxBlocksize = 9000; deepSplit = 2; minModuleSize = 40; and mergeCutHeight = 0.40. Dissimilarity degree d_ij (4) and dynamic fuzzy decision tree were first used to divide gene modules.

(1)

(2)

(3)

(4)

DEGs related to immune processes

Yoshihara et al. [12] established the ESTIMATE method. This method used immune cell–related gene expression profiles, considered the screened DEGs as background genes, and performed gene-set enrichment analysis [18] (GSEA) on external samples to obtain “ImmuneScore” scores for external samples, which were used to evaluate the immune cell content of the samples. Adopting a similar process, the ESTIMATE algorithm was also used to calculate the “StromalScore” of a sample to assess the content of stromal cells in the sample. The ESTIMATE algorithm was performed to analyze 374 liver cancer samples from the TCGA database. These samples were grouped according to the median score of “ImmuneScore,” and DEGs were screened between high- and low-score groups. Log₂|foldchange| >1 and P <0.05 were set as threshold values. For “StromalScore” scoring, the same process was repeated to screen out DEGs. The intersection genes of the DEGs of “ImmuneScore” and “StromalScore” were considered as genes closely related to the immune process of liver cancer.

Gene ontology and kyoto encyclopedia of genes and genomes pathway enrichment analysis

GO (Gene Ontology) and KEGG (Kyoto Encyclopedia of Genes and Genomes) databases are open to public and provide the function annotations of gene sets, rendering a better understanding of their biological functions. GO enrichment analysis included three aspects [19], namely biological process, cellular component (CC), and molecular function (MF), and KEGG was mainly used to conduct pathway enrichment analysis [20]. These two methods were applied to analyze the DEGs screened, setting P value <0.05 as the cutoff. The R package “enrichplot,” “clusterProfiler,” and “ggplot” were used for all aforementioned processes.

Establishment of the ICRP score

Of 374 liver cancer samples, samples with incomplete clinical information were removed, finally obtaining 225 samples containing complete clinical information. The gene expression data and clinical data of the samples were employed to perform prognostic analysis. A total of 225 samples were randomly divided into a training set (114 samples) and a test set (111 samples). The lasso regression [21] was carried out in the training set to eliminate the collinearity between factors, and the appropriate penalty coefficient was selected to minimize the partial likelihood deviance of the regression equation. The multivariate Cox proportional-hazards regression analysis [22, 23] was performed in the training set to build a prognostic model, named Immunity and Cancer-stem-cell Related Prognosis (ICRP) score, as follows:

where Fi is the value of the ith factor, β is the corresponding coefficient, and a high ICRP score represents a predicted poor prognosis.

Verification of the ICRP score

The samples in the training and test sets were grouped according to the median value of the ICRP score in the training set, and then Kaplan–Meier [24] survival curves with log-rank test were plotted. The following other verifications on the training and test sets were performed at the same time to verify the validity of the ICRP score: (1) survival status of patients after 1, 3, and 5 years were predicted, and the receiver operating characteristic [25] (ROC) curves were plotted to calculate the AUC. (2) The C-index [26] was calculated to evaluate the predictive effect of the prognostic model. Furthermore, ROC curves based on the AJCC stage and the Albumin–Bilirubin (ALBI) score were also drawn for comparison with the existing prognostic methods. The C-index of the AJCC stage, the ALBI score, and the signature of chromosomal instability (CIN25) from specific genes (containing 25 genes)[27] were calculated and compared with the results of the proposed model.

In the present study, quantitative real-time polymerase chain reaction (qPCR) assay was used to verify the expression level of four genes. Total RNA was extracted from normal liver cells (THLE3) and cancer cells (SNU-423) using the TRIzol extraction method, and then the concentration was measured using a NanoDrop 2000 spectrophotometer (Thermo Scientific, USA). cDNA was produced using the RevertAid First Strand cDNA Synthesis Kit (Thermo Scientific), and the reverse transcription reaction was conducted using the ABI7900 system (Applied Biosystems, USA). Finally, the relative gene expression was calculated by the 2^-ΔΔCt method. The primer sequences used in the qPCR are shown in Supplementary Table 1. Each sample was measured in triplicate.

Simultaneously, the Human Protein Atlas database (http://www.proteinatlas.org) was used to observe the gene expression at the protein level for further understanding the function of key genes in the proposed prognostic model. The TIMER database (https://cistrome.shinyapps.io/timer/) was used to explore the relationship between gene expression and immune cell content.

The workflow of this study is shown in Figure 1A.