Distinct Urinary Metabolic Biomarkers of Human Colorectal Cancer

Abstract

Colorectal cancer (CRC) is one of the most commonly diagnosed cancers with high mortality rate due to its poor diagnosis in the early stage. Here, we report a urinary metabolomic study on a cohort of CRC patients (n =67) and healthy controls (n =21) using ultraperformance liquid chromatography triple quadrupole mass spectrometry. Pathway analysis showed that a series of pathways that belong to amino acid metabolism, carbohydrate metabolism, and lipid metabolism were dysregulated, for instance the glycine, serine and threonine metabolism, alanine, aspartate and glutamate metabolism, glyoxylate and dicarboxylate metabolism, glycolysis, and TCA cycle. A total of 48 differential metabolites were identified in CRC compared to controls. A panel of 12 biomarkers composed of chenodeoxycholic acid, vanillic acid, adenosine monophosphate, glycolic acid, histidine, azelaic acid, hydroxypropionic acid, glycine, 3,4-dihydroxymandelic acid, 4-hydroxybenzoic acid, oxoglutaric acid, and homocitrulline were identified by Random Forest (RF), Support Vector Machine (SVM), and Boruta analysis classification model and validated by Gradient Boosting (GB), Logistic Regression (LR), and Random Forest diagnostic model, which were able to discriminate CRC subjects from healthy controls. These urinary metabolic biomarkers provided a novel and promising molecular approach for the early diagnosis of CRC.

Figure 1

Metabolite profile of CRC patients (n =67) and healthy control (n =21). (a) Relative abundance of urinary metabolites classes in CRC and control group. (b) Relative abundance of urinary metabolites classes in CRC and control samples. (c) Heatmap of urinary metabolites concentrations (Z-score scale to -2~2)in CRC patients and controls.

Figure 2

PCA and OPLS-DA of CRC metabolism. (a) PCA model generated from CRC patients and healthy controls. (b) OPLS-DA model generated from CRC patients and healthy controls. (c) Correlation coefficient of permutation test. (d) Differential metabolites identified by correlation coefficient and VIP value.

Figure 3

Differential metabolites for CRC urine sample compared to control. (a) Volcano plot for differential metabolites identified by OPLS-DA in CRC patients vs controls (VIP>1, |correlation coefficient| >0.3). (b) Volcano plot for differential metabolites identified by univariate statistical analysis in CRC patients vs controls (p <0.05, significantly increased metabolites in CRC (FC>1, red dots) and significantly decreased metabolites in CRC (FC<1, blue dots). (c) Heatmap of differential biomarkers of CRC patients vs controls (Z-score scale to -2~2). (d) Box plot of representative differential metabolites with significantly different concentration in CRC sample vs control (p <0.05).

Figure 4

Metabolic pathway analysis based on differential metabolites of CRC patients. (a) Bar plot of metabolic pathway analysis based on SMPDB database (top 50). (b) Bubble plot of metabolic pathway analysis based on HSA database.

Figure 5

Identification of biomarkers for CRC by Random Forest, Support Vector Machine, and Boruta analysis. (a) Metabolite importance plot of Random Forest analysis calculated by Mean Decrease Gini for classification between CRC patients and healthy controls (top 10 metabolites). (b) Metabolite importance plot of Support Vector Machine analysis calculated by recursive feature elimination (RFE) for classification between CRC patients and healthy controls (top 10 metabolites). (c) Box plot of Boruta analysis for the relevant feature selection of potential biomarker (blue box correspond to minimal, average and maximum Z-score of a shadow attribute; green and red box correspond to the confirmed and the rejected attributes, respectively).

Figure 6

Validation of biomarkers for CRC by Gradient Boosting, Logistic Regression, and Random Forest diagnostic model. (a) The receiver operating characteristic (ROC) curve of Gradient Boosting diagnostic model with sensitivity of 0.991 and specificity of 1.00. (b) The precision recall curve of Gradient Boosting diagnostic model. (c) The receiver operating characteristic (ROC) curve of Logistic Regression diagnostic model with sensitivity of 0.885 and specificity of 0.983. (d) The precision recall curve of Logistic Regression diagnostic model. (e) The receiver operating characteristic (ROC) curve of Random Forest diagnostic model with sensitivity of 1.00 and specificity of 1.00. (f) The precision recall curve of Random Forest diagnostic model.