Andrea Rodriguez Martinez, Joram M. Posma, Rafael Ayala, Ana L. Neves, Maryam Anwar, Jeremy K. Nicholson, Marc-Emmanuel Dumas

May 26, 2017

1 Abstract

“MWASTools” is an R package designed to provide an integrated and user-friendly pipeline to analyze metabonomic data in the context of large-scale epidemiological studies. Key functionalities of the package include: quality control analysis; metabolite-phenotype association models; data visualization tools; metabolite assignment using statistical total correlation spectroscopy (STOCSY); and biological interpretation of MWAS results.

2 Introduction

Metabonomics is a powerful systems biology approach that targets metabolites from biofluids (e.g. urine or plasma) or tissues, providing metabolic patterns that correspond to the metabolic status of the organism as a function of genetic and environmental influences (Nicholson et al. 2002). Thanks to the recent developments in high-throughput platforms (i.e. nuclear magnetic resonance (NMR) and mass spectrometry (MS)), metabolic profiling is now being used for large-scale epidemiological applications such as metabolome-wide association studies (MWAS) (Holmes et al. 2008; Elliott et al. 2015).

Customized statistical modeling approaches and data visualization tools are essential for biomarker discovery in large-scale metabolic phenotyping studies. Several software packages have been developed to detect and visualize metabolic changes between conditions of interest (e.g. disease vs control) using multivariate statistical methods (e.g. OPLS-DA) (Gaude et al. 2013; Thevenot et al. 2015). However, a major limitation of these multivariate models from the epidemiological perspective is that they do not properly account for cofounding factors (e.g. age, gender), which might distort the observed associations between the metabolites and the condition under study. Here, we present a package to perform MWAS using univariate hypothesis testing with efficient handling of epidemiological confounders. Our package provides a versatile and user-friendly MWAS pipeline with the following key functionalities: quality control (QC) analysis; metabolite-phenotype association models (partial correlations, generalized linear models) adjusted for epidemiological confounders (e.g. age or gender); bootstrapping of association models; visualization of statistical outcomes; metabolite assignment using Statistical Total Correlation Spectroscopy (STOCSY) (Cloarec et al. 2005); and biological interpretation of MWAS results (Kanehisa & Goto 2000).

3 Installation instructions

Assuming that R (>=3.3) and Bioconductor have been correctly installed, MWASTools can be installed with:

if (!requireNamespace("BiocManager", quietly = TRUE)) install.packages("BiocManager")
BiocManager::install("MWASTools", version = "devel")

4 Case-study

4.1 Study subject

The FGENTCARD cohort is a Lebanese clinical cohort of patients with/without coronary artery disease (CAD) (Platt et al. 2015). Plasma samples from this cohort were profiled by 1H NMR, to identify metabolites associated with risk factors of CAD. In this case study, we describe how the “MWASTools” package was used to identify metabolites associated with type II diabetes (T2D).

4.2 Metabonomic data

1H NMR plasma spectra were acquired on a Bruker Avance III 600 MHz spectrometer, in a randomized design. Quality control (QC) samples, composed of identical aliquots of a representative pool of the experimental samples, were injected regularly throughout the run. Following phasing and baseline correction in TopSpin 3.2 software, the spectra were calibrated to the glucose alpha anomeric signal at δ 5.23 (Pearce et al. 2008) and aligned using recursive segment-wise peak alignment (Veselkov et al. 2009). For simplification purposes, this case study focuses on the analysis of the spectral region: δ 1.60 - 0.80

4.3 Clinical data

For each patient of the cohort, information regarding age, gender, T2D status, and body mass index (BMI) was recorded.

4.4 Load data

First we loaded the “MWASTools” package:


We then loaded the dataset required to perform the analysis:

## class: SummarizedExperiment 
## dim: 595 516 
## metadata(0):
## assays(1): metabolic_data
## rownames(595): 0.80006656 0.80141232 ... 1.59810224 1.599448
## rowData names(0):
## colnames(516): P1 P2 ... QC9 QC10
## colData names(5): Age Gender T2D BMI sample_type

metabo_SE is a SummarizedExperiment object, generated with the function “MWAS_SummarizedExperiment()”, and containing the following information:
- metabolic_data: matrix containing the 1H NMR profiles (δ 1.60 - 0.80) of the experimental samples (n = 506) and the QC samples (n = 10).
- clinical_data: matrix containing clinical data (age, gender, T2D and BMI) and sample type information (i.e. experimental or QC sample).

4.5 Quality control (QC) analysis

To ensure the stability and reproducibility of the analytical run, we performed QC analysis based on principal component analysis (PCA) and coefficient of variation (CV) (sd/mean) across the QC samples (Dumas et al. 2006).

# PCA model
PCA_model = QC_PCA(metabo_SE, scale = FALSE, center = TRUE)

# Plot PCA scores (PC1 vs PC2 & PC3 vs PC4)
par(mfrow = c(1, 2))
QC_PCA_scoreplot(PCA_model, metabo_SE, main = "PC1 vs PC2")
QC_PCA_scoreplot(PCA_model, metabo_SE, px = 3, py = 4, main = "PC3 vs PC4")

In both score plots, the QC samples appear tightly clustered in the center of the Hotelling´s ellipse confirming the absence of batch effects, and ensuring the reproducibility of the analytical run.

Following QC analysis via PCA, we calculated the CVs of the NMR signals across the QC samples. Notice that CV = 0.30 and CV = 0.15 are the thresholds established by the FDA (U.S. Food and Drug Administration) for biomarker discovery and quantification, respectively.

# CV calculation
metabo_CV = QC_CV(metabo_SE, plot_hist = FALSE)

# NMR spectrum colored according to CVs
CV_spectrum = QC_CV_specNMR(metabo_SE, ref_sample = "QC1")