Identification of subgroups along the glycolysis-cholesterol synthesis axis and the development of an associated prognostic risk model

Abstract

Background: Skin cutaneous melanoma (SKCM) is one of the most highly prevalent and complicated malignancies. Glycolysis and cholesterogenesis pathways both play important roles in cancer metabolic adaptations. The main aims of this study are to subtype SKCM based on glycolytic and cholesterogenic genes and to build a clinical outcome predictive algorithm based on the subtypes.

Methods: A dataset with 471 SKCM specimens was downloaded from The Cancer Genome Atlas (TCGA) database. We extracted and clustered genes from the Molecular Signatures Database v7.2 and acquired co-expressed glycolytic and cholesterogenic genes. We then subtyped the SKCM samples and validated the efficacy of subtypes with respect to simple nucleotide variations (SNVs), copy number variation (CNV), patients' survival statuses, tumor microenvironment, and proliferation scores. We also constructed a risk score model based on metabolic subclassification and verified the model using validating datasets. Finally, we explored potential drugs for high-risk SKCM patients.

Results: SKCM patients were divided into four subtype groups: glycolytic, cholesterogenic, mixed, and quiescent subgroups. The glycolytic subtype had the worst prognosis and MGAM SNV extent. Compared with the cholesterogenic subgroup, the glycolytic subgroup had higher rates of DDR2 and TPR CNV and higher proliferation scores and MK167 expression levels, but a lower tumor purity proportion. We constructed a forty-four-gene predictive signature and identified MST-321, SB-743921, Neuronal Differentiation Inducer III, romidepsin, vindesine, and YM-155 as high-sensitive drugs for high-risk SKCM patients.

Conclusions: Subtyping SKCM patients via glycolytic and cholesterogenic genes was effective, and patients in the glycolytic-gene enriched group were found to have the worst outcome. A robust prognostic algorithm was developed to enhance clinical decisions in relation to drug administration.

Keywords: Machine learning; Melanoma; Metabolic; Predictive model.

Fig. 1

Stratification of SKCM tumors based on expression of glycolysis and cholesterogenic genes. A The CDF curve under different values of k. The value of k represents the number of clusters during the consensus cluster. When the optimal k value is reached, the area under the CDF curve will not significantly increase with the increase of k value. B Heatmap depicting consensus clustering solution (k=6) for glycolysis and cholesterogenic genes in SCKM samples (n=469). C Scatter plot showing median expression levels of co-expressed glycolytic (x-axis) and cholesterogenic (y-axis) genes in each SKCM sample. Metabolic subgroups were assigned on the basis of the relative expression levels of glycolytic and cholesterogenic genes. D Heatmap depicting expression levels of co-expressed glycolytic and cholesterogenic genes across each subgroup. E PCA showed that patients in the different subtgroups were significantly different from each other. F Kaplan-Meier survival curves for patients in the different subgroups. Log-rank test P values are shown. The clinical outcome endpoint is OS. G Kaplan-Meier survival curves for patients in the different subgroups. Log-rank test P value is shown. The clinical outcome endpoint is PFI. SKCM, skin cutaneous melanoma; CDF, cumulative distribution function; PCA, principal components analysis; OS, overall survival; PFI, progression-free interval. And P < 0.05 is defined as statistically significant

Fig. 2

Mutational landscape across metabolic subgroups of SKCM. A The map of waterfall depicting the distribution of SNV events affecting frequently mutated genes in SKCM across the metabolic subgroups. B Bar plot illustrating the CNV of DDR2 across the metabolic subgroups. Chi-square test P value is shown. C Boxplot depicting the relationship of types of CNV and the expression of DDR2. D Boxplot depicting the expression levels of DDR2 across the metabolic subgroups. E Bar plot illustrating the CNV of TPR cross the metabolic subgroups. Chi-square test P value is shown. F Boxplot depicting the relationship of types of CNV and the expression of TPR. D Boxplot depicting the expression levels of TPR across the metabolic subgroups. SKCM, skin cutaneous melanoma; SNV, simple nucleotide variation; CNV, copy number variation. And P < 0.05 is defined as statistically significant

Fig. 3

Evaluation of tumor purity across metabolic subgroups by ESTIMATE. A Violin plot illustrating the immune score across metabolic subgroups. B Violin plot illustrating the stromal score across metabolic subgroups. C Violin plot illustrating the ESTIMATE score across metabolic subgroups. D Violin plot illustrating the tumor purity across metabolic subgroups. Kruskal-wallis test P values are shown. And P < 0.05 is defined as statistically significant

Fig. 4

The proliferation status across metabolic subgroups. A Scatter plot depicting the correlation between the proliferation score and median expression of cholesterogenic genes (left) and glycolytic genes (right). The Spearman’s test P values are shown. B Boxplot illustrating proliferation score across metabolic subgroups. Kruskal-wallis test P value is shown. C Scatter plot depicting the correlation between the MKI67 expression and median expression of cholesterogenic genes (left) and glycolytic genes (right). The Spearman’s test P values are shown. D Boxplot illustrating the MKI67 expression across metabolic subgroups. Kruskal-wallis test P value is shown. P < 0.05 is defined as statistically significant

Fig. 5

WGCNA to identify similar genes networks of cholesterogenic genes and glycolytic genes. A The relationship of soft threshold and TOM-based dissimilarity (left). The relationship of soft threshold and mean connectivity (right). B After the dynamic of cut and merged, a total of 6 gene modules were finally generated. C Heat map for the correlation of gene modules and phenotypes. D Scatter plot depicting the correlation between gene significance and module membership of genes in turquoise network. E Scatter plot depicting the correlation between gene significance and module membership of genes in yellow network. WGCNA, weighted correlation network analysis; TOM, topological overlap matrix. And P < 0.05 is defined as statistically significant

Fig. 6

The results of KEGG analysis in cholesterogenic genes and glycolytic genes. KEGG, Kyoto Encyclopedia of Genes and Genomes. P < 0.05 is defined as statistically significant

Fig. 7

Build the risk model by LASSO. A Cross validation based on C-index to determine the best choice of genes in the model. B Genes in the different combinations and their corresponding coefficients. C–E Patients of training set were arranged in the same ascending order of the risk score. F–H Patients of internal validation set were arranged in the same ascending order of the risk score. I–K Patients of GSE19234 data set were arranged in the same ascending order of the risk score. C, F, I Patients were divided into different risk levels according to the median of the risk scores in their respective data sets. D, G, J The relationship between the survival outcome and risk levels of patients. Low-risk patients were shown on the left side of the dotted line and high-risk patients were shown on the right side. E, H, K Heat maps for the genes in the signature. LASSO, least absolute shrinkage and selection operator. And P < 0.05 is defined as statistically significant

Fig. 8

Verification of the effectiveness of the model. A–E Kaplan-Meier curve for survival analysis. F–J The ROC curve of 5-year follow-up time. A, C, F, H The results in the training set. B, D, G, I The results in the internal validation set. E, J The results in GSE19234. The clinical outcome endpoint in A, B, E, F, G, J was PFI. The clinical outcome endpoint in C, D, H, I was OS. AUC, area under curve; PFI, progression-free interval; OS, overall survival. And P < 0.05 is defined as statistically significant

Fig. 9

The comparison between our model and Liao’s immune-related model. A Kaplan-Meier curve for Liao’s immune-related model in GSE19234 (n = 44). The clinical outcome endpoint was OS. B The ROC curve of 5-year follow-up time for our model and Liao’s immune-related model in GSE19234 (n = 44). AUC, area under curve; OS, overall survival. And P < 0.05 is defined as statistically significant

Fig. 10

Identification of candidate agents with higher drug sensitivity in patients with high risk score. A The results of Spearman’s correlation analysis and differential drug response analysis of three CTRP-derived compounds. B The results of Spearman’s correlation analysis and differential drug response analysis of three PRISM-derived compounds. Note that lower values on the y-axis of boxplots imply greater drug sensitivity