Denoising Autoencoder Normalization for Large-Scale Untargeted Metabolomics by Gas Chromatography-Mass Spectrometry

Abstract

Large-scale metabolomics assays are widely used in epidemiology for biomarker discovery and risk assessments. However, systematic errors introduced by instrumental signal drifting pose a big challenge in large-scale assays, especially for derivatization-based gas chromatography-mass spectrometry (GC-MS). Here, we compare the results of different normalization methods for a study with more than 4000 human plasma samples involved in a type 2 diabetes cohort study, in addition to 413 pooled quality control (QC) samples, 413 commercial pooled plasma samples, and a set of 25 stable isotope-labeled internal standards used for every sample. Data acquisition was conducted across 1.2 years, including seven column changes. In total, 413 pooled QC (training) and 413 BioIVT samples (validation) were used for normalization comparisons. Surprisingly, neither internal standards nor sum-based normalizations yielded median precision of less than 30% across all 563 metabolite annotations. While the machine-learning-based SERRF algorithm gave 19% median precision based on the pooled quality control samples, external cross-validation with BioIVT plasma pools yielded a median 34% relative standard deviation (RSD). We developed a new method: systematic error reduction by denoising autoencoder (SERDA). SERDA lowered the median standard deviations of the training QC samples down to 16% RSD, yielding an overall error of 19% RSD when applied to the independent BioIVT validation QC samples. This is the largest study on GC-MS metabolomics ever reported, demonstrating that technical errors can be normalized and handled effectively for this assay. SERDA was further validated on two additional large-scale GC-MS-based human plasma metabolomics studies, confirming the superior performance of SERDA over SERRF or sum normalizations.

Keywords: GC–MS; data normalization; derivatization; primary metabolism; statistics.

Figure 1

Comparisons of normalizations of large-scale GC–MS human cohort plasma datasets.

Figure 2

Reaction schemes of MSTFA, MTBSTFA, and PCF derivations, and the chromatography of valine derivatized with MSTFA. (a) valine-1TMS and valine-d8-1TMS products; (b) valine-2TMS and valine-d8-2TMS products; (c) reaction schemes for MSTFA, MTBSTFA, and PCF derivatization.

Figure 3

Relative standard deviation (%) of 16 amino acids by three derivatization reagents. Raw: not normalized data; ISTD: normalization of each amino acid to its corresponding internal standard; fTIC: normalization to the sum of 13 fatty acid methyl esters; iTIC: normalization to the sum of all internal isotope-labeled standards; mTIC: normalization to the sum of all identified metabolites.

Figure 4

Sequence of sample acquisition and distribution of three sets of quality controls (QC) in large scale GC–MS metabolomics. (a) Sequence of injections of blanks, pooled sample quality controls, and BioIVT and NIST external plasma quality controls, plus blinded sample doublets. (b–d) Partial least square-discriminant analysis plots (PLS-DA) of (b) raw data, and effect of normalization by (c) SERDA and (d) SERRF normalization.

Figure 5

Correlation of blinded T2D cohort sample duplicates after SERDA normalization.

Figure 6

The relative standard deviation (RSD) of pool QC samples (red) and BioIVT qc samples (blue) for each set of 10, 20, 40, and 80 biological samples. With a smaller number of training QC samples, the performance of SERDA decreases, as indicated by the increasing of both the RSD of pool QC traing qc samples and BioIVT validation qc samples.

Figure 7

Comparison of ISTD absolute ratio normalization with QC-based and TIC-based normalization methods. The Friedman nonparametric test was used for significance comparison with raw. p-value threshold: 0.1234 (ns), 0.0001 (****). The Friedman nonparametric test was used for significance testing compared to SERDA. P value threshold: 0.1234 (ns), 0.0332 (#), 0.0002 (###), 0.0001 (####). (a) One-to-one: the absolute ratio was calculated by dividing the peak intensity of endogenous metabolite by the corresponding deuterated ISTD; (b) One-to-class: the absolute ratio was calculated by dividing the peak intensity of endogenous metabolite by a single deuterated compound as an analog ISTD for the entire class.

Figure 8

Effect of different normalization methods on residual errors in GC–MS-based metabolomics datasets. Left panel: this study, N = 413 quality control human plasma samples (QC), reporting 661 metabolites. Mid panel: 104 human plasma QC samples of the GeneBank study on 319 metabolites. Right panel: 30 QC plasma samples of the MPA study on 991 metabolites. For each panel, cumulative distributions of cross-validated relative standard deviations (RSD) are given using raw (black), batchwise-LOESS (blue), SERRF (green), and SERDA (red) normalized dataset. The coverage of metabolites achieving specific RSD levels is given as the y-axis.