A joint analysis of transcriptomic and metabolomic data uncovers enhanced enzyme-metabolite coupling in breast cancer

Abstract

Disrupted regulation of cellular processes is considered one of the hallmarks of cancer. We analyze metabolomic and transcriptomic profiles jointly collected from breast cancer and hepatocellular carcinoma patients to explore the associations between the expression of metabolic enzymes and the levels of the metabolites participating in the reactions they catalyze. Surprisingly, both breast cancer and hepatocellular tumors exhibit an increase in their gene-metabolites associations compared to noncancerous adjacent tissues. Following, we build predictors of metabolite levels from the expression of the enzyme genes catalyzing them. Applying these predictors to a large cohort of breast cancer samples we find that depleted levels of key cancer-related metabolites including glucose, glycine, serine and acetate are significantly associated with improved patient survival. Thus, we show that the levels of a wide range of metabolites in breast cancer can be successfully predicted from the transcriptome, going beyond the limited set of those measured.

Figure 1

(A) The prediction pipeline: Step (1) A classifier predicting RGM triplets that are significantly associated: using Metabolomic and transcriptomic data to identify genes and metabolites that are connected via a metabolic reaction and are significantly associated with each other. This is obtained via building an RGM SVM classifier where each instance represents a unique RGM triplet and whose output is a confidence level signifying whether the gene expression and metabolite levels are significantly positively or negatively associated across all samples (Methods). Step (2) A regressor predicting metabolite levels from gene expression in a sample-specific manner: Confidence levels predicted by the classifier for each RGM triplet in the first step are utilized together with the expression and network features to build a generalized multiple linear regression predictor of metabolite levels from the pertaining enzymes’ gene expression levels (Methods).

Figure 2

(A) Top panels describe the mean AUC of the RGM predictors for all the breast data together and for cancer and noncancerous samples separately. The sensitivity, specificity and accuracy levels of the different classifiers are indicated as well. Bottom panels display the correlation between the confidence levels of the RGM predictions and the gene-metabolite correlations actually measured in the data, for the three cases studied here (see main text). Confidence levels range between −1 and 1 where 1 represents a highly confident positive association and −1 a highly confident negative association. (B) Scatter plot representing the association, for all genes, between (1) the absolute Spearman correlation between gene and metabolites associated with it across all samples (x-axis) and (2) the magnitude of the differential expression of that gene between noncancerous and cancer samples (y-axis). (C) Pathways that are predicted to be regulated in healthy (red) and cancer (blue) samples. The dashed line represents a hyper-geometric significance threshold of 0.05 (FDR-corrected for multiple hypotheses testing).

Figure 3

(A) The mean AUC of the RGM predictors for the Brauer BC dataset. The sensitivity, specificity and accuracy levels of the different classifiers are indicated as well. (B) The mean AUC of the RGM HCC predictors for all the data together and for cancer and healthy cohorts separately, and the sensitivity, specificity and accuracy levels. (C) A scatter plot describing the correlation between pathway enrichment p-values among the three datasets, when all significantly enriched pathways in the two datasets were considered (hyper-geometric p-value < 0.05). The left most panel displays the latter for the two BC datasets, and the middle and right most panels compare the HCC values to each of the BC datasets.

Figure 4. Kaplan-Meier survival plots for extracellular levels of glycine, acetate and serine.

The associated FDR-corrected log-rank P-values are 0.002, 5.18e-6 and 8e-4, respectively.