Novel personalized pathway-based metabolomics models reveal key metabolic pathways for breast cancer diagnosis

Abstract

Background: More accurate diagnostic methods are pressingly needed to diagnose breast cancer, the most common malignant cancer in women worldwide. Blood-based metabolomics is a promising diagnostic method for breast cancer. However, many metabolic biomarkers are difficult to replicate among studies.

Methods: We propose that higher-order functional representation of metabolomics data, such as pathway-based metabolomic features, can be used as robust biomarkers for breast cancer. Towards this, we have developed a new computational method that uses personalized pathway dysregulation scores for disease diagnosis. We applied this method to predict breast cancer occurrence, in combination with correlation feature selection (CFS) and classification methods.

Results: The resulting all-stage and early-stage diagnosis models are highly accurate in two sets of testing blood samples, with average AUCs (Area Under the Curve, a receiver operating characteristic curve) of 0.968 and 0.934, sensitivities of 0.946 and 0.954, and specificities of 0.934 and 0.918. These two metabolomics-based pathway models are further validated by RNA-Seq-based TCGA (The Cancer Genome Atlas) breast cancer data, with AUCs of 0.995 and 0.993. Moreover, important metabolic pathways, such as taurine and hypotaurine metabolism and the alanine, aspartate, and glutamate pathway, are revealed as critical biological pathways for early diagnosis of breast cancer.

Conclusions: We have successfully developed a new type of pathway-based model to study metabolomics data for disease diagnosis. Applying this method to blood-based breast cancer metabolomics data, we have discovered crucial metabolic pathway signatures for breast cancer diagnosis, especially early diagnosis. Further, this modeling approach may be generalized to other omics data types for disease diagnosis.

Figures

Fig. 1

The workflow of the pathway-based metabolomics data analysis. Step 1: conversion from metabolite- to pathway-based metabolomics data. The input data include the master file containing pathway-metabolite mapping information, the metabolomics profiling data, and the normal/tumor classification vector. The metabolomics-level data are transformed to pathway-level data by the pathifier algorithm. The output file of pathifier is the pathway dysregulation score matrix, within which each score measures the deregulation of a specific pathway for a specific sample. Step 2: model construction. Qualified COH plasma samples are split by 80/20 for training and holdout testing data. Correlation feature selection (CFS) is used for feature selection and the logistic regression model is used for classification. Tenfold cross-validation (10-fold CV) is applied with CFS feature selection in the plasma training data set. Two models are constructed: an all-stage diagnostic model and an early-stage diagnostic model. Step 3: model evaluation. The model performance is assessed using receiver operating characteristic (ROC) curves and various metrics, including AUC, MCC, sensitivity, specificity, and F1-statistic

Fig. 2

The performance of the all-stage diagnosis model for breast cancer. We used 80 % of the controls and cases in the COH plasma data set to train the model. The remaining COH plasma data (20 %) and the COH serum data set were used as the test set and validation set. a Receiver operating characteristic (ROC) curves for the all-stage breast cancer diagnosis from different data sets. b AUC, MCC, sensitivity, specificity, and F1-statistic to measure the performance of the all-stage diagnosis model. c Mutual information for pathway features selected by the all-stage diagnosis model. d. Log fold change of metabolites associated with the selected pathway features determined by comparing cases and controls across different data sets

Fig. 3

The performance of the early-stage diagnosis model for breast cancer. We used 80 % of the controls and early-stage (stage I and II) cases in the COH plasma data set to train the model. The remaining controls and early stage cases in the COH plasma data set, as well as controls and early stage cases in the COH serum data set, were used as the testing and validation set. a Receiver operating characteristic (ROC) curves for the early-stage breast cancer diagnosis from different data sets. b AUC, MCC, sensitivity, specificity, and F1-statistic to measure the performance of the early-stage diagnosis model. c Mutual information for pathway features selected by the all-stage diagnosis model. d Log fold change of metabolites associated with the selected pathway features determined by comparing cases and controls across different data sets

Fig. 4

Integrative analysis of pathway features and the associated metabolites. The key pathways and their intersections crucial for breast cancer diagnosis. Metabolites and enzymes are represented with nodes of different shapes and colors, and their relationships are represented by edges

Fig. 5

Receiver operating characteristic (ROC) curves comparison of pathway-based model and metabolites-based model among data sets. The same 80 % of early stage (stage I and II) cases and controls from the COH plasma data set used in the early-stage diagnosis model were used for the plasma training set. The remaining 20 % of early stage (stage I and II) cases and controls represent the test set. The metabolite-based model is based on the same tenfold cross-validation CFS selection used for the plasma training set. ROC curves for training and test sets are compared between the plasma-based model and the metabolite-based model among data sets