dbenedek / metabolicexpressr

Metabolic subtyping of tumors from gene expression data

MetabolicExpressR: Metabolic subtyping of patient tumors from gene expression data

A detailed tutorial can be found here.

Installation

devtools::install_github("dBenedek/MetabolicExpressR")

Features

1. Perform Gene Set Variation Analysis (GSVA) on cancer patient gene expression data using KEGG metabolic pathways. Potentially non-cancer data can be used as well.
2. Perform k-means clustering on the GSVA matrix, identify metabolic subtypes. Optimal number of k is defined based on data or user-specified k is used.
3. Summarize pathway "expression" per cluster: i.e. mean pathway activity per cluster.
4. Run PROGENy on the gene expression data of tumors and compare signature activity scores between the identified clusters.
5. Perform Kaplan-Meier (KM) analysis with clusters if survival data is present.

Usage

Gene Set Variation Analysis

Gene Set Variation Analysis (GSVA) allows us to measure the enrichment of gene sets on the single sample level. It is recommended to have at least around 10 samples in the data set. In all the examples we use the KEGG metabolic pathways collection, although any other curated gene sets might be used.
The input of GSVA is a normalized gene expression matrix, where row names are gene names (with the same gene nomenclature as in the gene set used), and column names correspond to sample IDs. The parameter kcdf defines the kernel to use during the non-parametric estimation of the cumulative distribution function (see GSVA documentation for details).

gsva_results <- run_gsva_metabolic(gene_exp_data = gene_exp_cptac_hnscc,
                                   kcdf = "Gaussian",
                                   kegg_gs = kegg_gs)
dim(gsva_results)

The output is an N x M matrix with row names as pathways and column names as sample IDs (same as in the gene expression matrix). The data values are the enrichment scores of the pathways per sample.

K-means clustering on the GSVA data

The next step is the k-means clustering on the GSVA matrix derived from the previous step using the function kmeans_gsva_metabolic().
This requires the GSVA matrix from the previous step, the gene set collection (kegg_gs, also used in the previous step). The user can either use the data set-specific optimal k for the clustering (default) when user_def_k = FALSE, or when user_def_k = TRUE then the user has to specify the parameter k used for k-means clustering. This way the resulting number of clusters is going to be already pre-defined.

kmeans_results <- kmeans_gsva_metabolic(gsva_data = gsva_results,
                                        kegg_gs = kegg_gs,
                                        user_def_k = FALSE,
                                        k = NULL)

head(kmeans_results$kmean_res)
kmeans_results$heatmap

The function returns a list with 1) a data frame with cluster membership per sample (kmean_res), and 2) a heatmap (heatmap) with GSVA enrichment scores visualized as a heatmap for the clusters.

Comparison of pathway "expression" between the identified clusters

Here we compare the pathway enrichment scores (pathway "expression") between the identified clusters. For this, we use the function plot_mean_pathway_activity() which compares and visualizes the mean pathway enrichment scores between the clusters. This gives us an overview on the pathway-level differences of the sample clusters.

plot_mean_pathway_activity(gsva_data = gsva_results,
                           kegg_gs = kegg_gs,
                           kmeans_res = kmeans_results$kmean_res)

PROGENy analysis and signature activity comparison between the identified clusters

From the website of PROGENy: "PROGENy is resource that leverages a large compendium of publicly available signaling perturbation experiments to yield a common core of pathway responsive genes for human and mouse. These, coupled with any statistical method, can be used to infer pathway activities from bulk or single-cell transcriptomics." Thus, it aims to quantify the activity of 14 signatures which are very often disturbed in several cancer entities. These signatures are androgen, EGFR, estrogen, hypoxia, JAK-STAT, MAPK, NFkB, p53, PI3K, TGFb, TNFa, Trail, VEGF, and WNT signaling.

progeny_results <- run_progeny(gene_exp_data = gene_exp, 
				kmeans_res = kmeans_clusters$kmean_res)
progeny_results$progeny_boxplot

The output is a boxplot (progeny_boxplot) with signaling activity scores per pathway compared between the identified clusters. In addition, differential testing P-value (Wilcoxon or Kruskal-Wallis test) is indicated in the facet header of each signature. The function outputs a data frame (progeny_res) as well, with PROGENy pathway activity scores per sample and signature.

Kaplan-Meier analysis with the identified clusters

Finally, we perform Kaplan-Meier (KM) analysis between the identified clusters. The function kmeans_metab_clust_surv() does exactly this, for which we need survival/clinical endpoint data with e.g. overall survival time and overall survival status of the samples (patients).
We have to specify the k-means clustering results with the kmeans_res parameter (just as in the previous step), we have to provide the clinical data table (clinical_data parameter). Moreover, we specify the column name of the sample IDs in the clinical data table (parameter sample_id_col), the endpoint status column name (surv_status_col parameter), and the endpoint time column name (surv_time_col parameter).

kmeans_metab_clust_surv(kmeans_res = kmeans_results$kmean_res, 
                        clinical_data = clinical_data_table,
                        sample_id_col = "case_id", 
                        surv_status_col = "overall_free_status", 
                        surv_time_col = "overall_survival")

The output is a KM survival curve of the sample clusters for the selected endpoint.

Test data sets

LinkedOmics is a great resource of several cancer types omics data. Data sets in the data/ folder were downloaded from there.

• CPTAC-HNSCC: data/CPTAC_HNSCC/
• CPTAC-COAD: data/CPTAC_COAD/
• TCGA-COAD: data/TCGA_COAD/
• TCGA-LAML: data/TCGA_LAML/
• TCGA-UVM: data/TCGA_UVM/
• TCGA-ACC: data/TCGA_ACC/