Alignment and Analysis of a Disparately Acquired Multibatch Metabolomics Study of Maternal Pregnancy Samples

Abstract

Untargeted liquid chromatography–mass spectrometry metabolomics studies are typically performed under roughly identical experimental settings. Measurements acquired with different LC-MS protocols or following extended time intervals harbor significant variation in retention times and spectral abundances due to altered chromatographic, spectrometric, and other factors, raising many data analysis challenges. We developed a computational workflow for merging and harmonizing metabolomics data acquired under disparate LC-MS conditions. Plasma metabolite profiles were collected from two sets of maternal subjects three years apart using distinct instruments and LC-MS procedures. Metabolomics features were aligned using metabCombiner to generate lists of compounds detected across all experimental batches. We applied data set-specific normalization methods to remove interbatch and interexperimental variation in spectral intensities, enabling statistical analysis on the assembled data matrix. Bioinformatics analyses revealed large-scale metabolic changes in maternal plasma between the first and third trimesters of pregnancy and between maternal plasma and umbilical cord blood. We observed increases in steroid hormones and free fatty acids from the first trimester to term of gestation, along with decreases in amino acids coupled to increased levels in cord blood. This work demonstrates the viability of integrating nonidentically acquired LC-MS metabolomics data and its utility in unconventional metabolomics study designs.

Graphical Abstract

INTRODUCTION

Low molecular weight compound profiling by liquid chromatography–coupled mass spectrometry (LC-MS) is routinely applied in biomedical research to obtain comprehensive read-outs of the metabolic composition of biological samples. Computational analysis of untargeted metabolomics data analysis is challenging for multiple reasons. Raw mass spectral outputs are complex and must be processed to form a matrix consisting of n samples with p metabolomics features represented by a mass-to-charge ratios (m/z) and chromatographic retention times (RT).^1,2 Metabolite identification is a central hurdle, whereby only a small proportion of the detected metabolome can be unambiguously characterized.^3,4 The diversity of LC-MS instrumentation and processing software precludes standardization of methods between laboratories and leads to varying content in processed metabolomics feature lists, even for the same specimens. It is difficult to relate unidentified metabolites between experiments using different LC-MS instrumentation and protocols, largely due to differences in measured m/z and especially RTs. Moreover, metabolomics studies often require larger sample sizes than can be analyzed in a single batch to achieve sufficient statistical power, requiring that they be divided into multiple batches. This leads to significant technical variation in the form of intra- and interbatch effects, the former defined by changes in instrumental response throughout the course of a batch analysis, the latter characterized by systematic differences in measured signal as well as RT shifts between batches.^5,6 Together, these challenges pose many obstacles to the interpretation of LC-MS metabolomics results.

Previously, our group published metabCombiner,⁷ a method for aligning and harmonizing nonidentically acquired LC-MS metabolomics data pertaining to biologically similar specimens. metabCombiner determines the intersection of metabolomics features across differences in chromatographic gradients, columns, and laboratories, and concatenates per-sample abundance measurements of known and unknown metabolites into a single merged table. This presents an opportunity to perform analyses on an expanded sample set, where information from multiple experiments is pooled into one table. A key question left unaddressed is how to manage the significant heterogeneity from merging nonidentically acquired experimental data. Major differences in chromatographic, spectrometric, and other analytical factors between experiments induce significant technical variation that must be corrected for prior to analysis of the full merged data set.^8,9 Batch effects correction approaches are developed and routinely applied to large-scale metabolomics studies, though these are mostly tested within studies in which experimental parameters are replicated to the greatest extent possible between batches. Furthermore, it remains unclear whether multiexperimental alignment will enhance or dilute signals present in individual data sets. Numerous studies have assessed the reproducibility of untargeted LC-MS metabolomics data analyzed by different mass spectrometers,^10–13 preprocessing software,^14–16 and laboratories.^17–20 Generally, these have concluded that the relative quantitation performance of untargeted metabolomics assays is reproducible, despite differences in these factors, but the choice of methods could impact the ability to detect significant observations, including biomarkers or impacted metabolic pathways.

Here, we describe a protocol for merging disparately acquired untargeted LC-MS metabolomics data from the Michigan Mother–Infant Pairs (MMIP) cohort. Our group has previously used a lipidomics platform to identify maternal lipids associated with the cord blood lipidome and infant birth weight in this cohort.^21,22 Untargeted LC-MS metabolomics was performed on the first trimester maternal plasma (M1), delivery maternal plasma (M3), and umbilical cord blood (CB) for a total of 106 mother-infant dyads. A subset of these samples was analyzed in 2016, whereas the remaining data were acquired by the same laboratory in 2019, but with a different chromatography system, mass spectrometer, and experimental protocol. We merged these two data sets and performed normalization to eliminate significant technical variation. Our analysis of the aligned and normalized data demonstrates enhanced statistical power to uncover significant changes in maternal plasma metabolome between the first trimester (M1) and third trimester (M3), and between maternal plasma (M3) and infant umbilical cord plasma (CB).

METHODS

Study Design

Pregnant women were recruited at their first prenatal appointment to the Michigan Mother Infant Pairs (MMIP) cohort. Recruitment occurred during 2010–2018. Eligibility criteria for MMIP includes age between 18 and 42 years old, had a spontaneously conceived singleton pregnancy, and intended to deliver at the University of Michigan Hospital. A subset of mother–infant dyads was selected for untargeted metabolomics measures. Inclusion criteria were complete demographic, survey, and health information at their initial study visit and availability of all biospecimens at all time points from mother and infant. The initial study visits occurred at 8–14 weeks gestation (M1), where participants provided blood samples, weight, and height. Women were recontacted prior to delivery. Maternal blood samples (M3) and umbilical cord blood samples (CB) were collected at delivery. The MMIP study was approved by the University of Michigan Institutional Review Board (HUM00017941). Written informed consent was received from all participants, and all plasma samples and metadata were deidentified prior to analysis. Characteristics of the study population can be found in Table 1 of our previously published work.²¹ At baseline, mothers were on average 32.1 years with an average BMI of 25.8 kg/m². Most women delivered vaginally (72%), and women with a higher BMI at baseline gained less weight across gestation. All women delivered at term. The average birth weight was 3.51 kg with 51 male and 55 female infants. Study samples were divided into two subsets, referred to as ex616 and ex946, with distinct experimental protocols performed on each subset.

Table 1.

Feature Counts from Batch and Experimental Alignment^a

MMIP Feature Counts from Alignment and Filtering Steps
Ionization Mode	Positive					Negative
Experiment	ex616 (N =197)			ex946 (N = 167)		ex616 (N = 198)			ex946 (N = 167)
Batch number	1	2	3	1	2	1	2	3	1	2
Batch feature counts	8254	7474	7149	16216	13391	4830	4905	5135	9549	8773
Batch-merged feature counts	3971			10466		2616			5643
Experiment-merged feature counts	2343					1583
Filtered by Missingness	342					202
Degenerate Features (adducts, fragments, …)	512					327
Annotated and Unannotated Features	1489					1054
Annotated Features	199					129

^aExperimental batches in positive and negative ionization modes were preprocessed and then merged within and between experiments. Successive intersections followed by filters for highly missing (50% tolerance per experiment) and degenerate features (adducts, fragments, multimers, and contaminants) reduce feature counts to 1489 and 1054, respectively.

Experimental Methods for ex616

ex616 consists of M1, M3, and CB time point samples for 56 subjects separated into three batches, with 10 pooled aliquots of each batch analyzed at regular intervals. 100 μL of each plasma sample was combined with 400 μL methanol:acetonitrile:acetone (1:1:1) extraction solvent containing 20 μM of internal standards (l-¹⁵N-anthranalic acid, l-¹⁵N₂-tryptophan, gibberelic acid, and l-epibrassinolide) in microcentrifuge tubes. Samples were vortexed for 5 min and centrifuged for 10 min at 15 000 rpm. 250 μL supernatant was transferred to another clean vial, dried under a N₂ stream, and reconstituted with addition of 100 μL methanol:H₂O (2:98) containing Zeatin. A 1290 Infinity Binary LC system from Agilent is used for LC separation together with a Water Acquity UPLC HSS T3 1.8 μm × 100 mm column. Mobile phase A is 100% water with 0.1% formic acid, and mobile phase B was 100% methanol with 0.1% formic acid. The gradient is 0–2 min 2% B, 2–20 min 2–75% B (linear), 20–22 min 75–98% B (linear), held at 98% for 22–30 min, followed by re-equilibration. Flow rate is set to 4.5 mL/min with a column temperature of 55 °C. An Agilent Technologies 6530 QTOF mass spectrometer with a dual Agilent Jetstream ESI source was used as the mass detector in the positive and negative ionization modes. Mass spectrometer settings were as follows. Ion source: gas temperature, 325 °C; drying gas flow, 10 L/min; nebulizer pressure, 45 psi; sheath gas temperature, 400 °C; sheath gas flow, 12 L/min; capillary voltage, 4000 V; skimmer voltage, 65 V; mass range, 50–1000 m/z; acquisition rate, 2 spectra/s with internal mass correction. Sample run order was not fully randomized, as M1 samples were run as a group toward the beginning or the end of each batch, separately from M3 and CB samples.

Experimental Methods for ex946

ex946 consists of M1, M3, and CB time point samples for 48 subjects separated into two batches, with 14 master pooled aliquots of all ex946 samples included within each batch. Data acquisition occurred in 2019, nearly three years after the ex616 samples. Sample preparation procedures were largely the same as in ex616, with minor differences. As in ex616, internal standard compounds included l-¹⁵N-anthranalic acid, l-¹⁵N₂-tryptophan, gibberellic acid, and l-epibrassinolide, along with Zeatin. Liquid chromatography column type and mobile solvents A and B were the same, but with an altered gradient: 1–16 min 0–99% B (linear), 16–20 min 99% B (hold), 20 min return to 1% B, followed by a 4 min re-equilibration at starting conditions. Mass spectrometry was performed with an Agilent 6545 QTOF instrument acquiring full-scan mass spectra over the range 50–1000 Da. Source parameters were as follows: drying gas temperature 350 °C, drying gas flow rate 10 L/min, nebulizer pressure 30 psi, sheath gas temp 350 °C and flow 11 L/min, and capillary voltage 3500 V. Sample order arrangement was randomized. Iterative Data Dependent (iDDA) MS/MS analysis was performed on pooled plasma. iDDA captures ms/ms in stepwise fashion, with rolling excluded precursors. For our untargeted platforms, we collect 8 rounds of iDDA at 3 different collision energies (10, 20, and 40 eV). Analysis of iDDA spectra using msPepSearch and the NIST20 spectral library was performed to provide MSI Level II and III identifications for selected features.

Initial Batch Data Processing

Data from both experimental sets were preprocessed using Agilent Profinder software, creating separate feature tables for each batch in both ionization modes. Adducts, in-source fragments, and multimer labels were annotated on a representative batch using Binner 1.1.0.²³ Known metabolites were annotated using in-house m/z and RT libraries.

Overview of Alignment and Normalization Workflows

The main computational workflow for this study was performed as depicted in Figure 1. Initial data analysis steps corrected for variations in retention times and acquired signal abundance values. The alignment steps merge the batch feature lists of each experiment into a cohesive table (Figure 1A). The normalization procedure consists of data filtering, imputation, batch effects removal, and scaling within each experimental set before reassembly into the final matrix (Figure 1B).

Figure 1.

Analytical workflow. Two experiments, ex616 and ex946, consist of three and two batches, respectively. (A) Alignment steps necessary to concatenate the LC-MS measurements, first between batches of the same experiment followed by between experiments (using metabCombiner). (B) Normalization steps necessary to eliminate unwanted sources of variation consist of missing value handling, batch effects correction, and Z-transformation on separate experimental sample sets, followed by reassembly.

Disparate Multibatch Alignment with metabCombiner

Metabolomics alignment between batches and experiments was performed using the metabCombiner R package.⁷ metabCombiner operates on processed metabolomics feature tables in a pairwise manner by generating a mapping between chromatographic RTs, then matching features through a score weighing differences in m/z, relative abundance (Q), and predicted RT differences. Within metabCombiner, the quality control (QC) samples of each batch were selected for relative quantitation comparisons, with M1, M3, and CB samples designated as “extra” columns. The program filters features missing in over 50% of each set of batch QC samples before analysis. In each experiment separately, feature tables corresponding to each batch were aligned in an iterative and stepwise manner to construct batch-merged tables. Subsequently, features from the two experimental tables were overlapped to construct a single cohesive table of sample measurements consisting of features detected in all batches. More details describing the use of metabCombiner, along with associated parameters, are described in the Supporting Information.

Missing Value Filtering and Imputation

Multivariate analyses often require complete data as input, hence missing feature values in the data set constructed over the preceding steps must be imputed. In this study, any features missing from over 50% of all experimental samples from either ex616 or ex946 were eliminated. For the remaining features, the data were log-transformed and missing data imputation was applied to feature quantities from the two experiments separately, using random-forest based imputation implemented in the MissForest R package,²⁴ which has reported advantages over other imputation methods for LC-MS metabolomics data.^25,26

Batch and Cross-Experimental Effects Correction

Numerous methods have been published for batch effects correction in conventional LC-MS metabolomics studies, including QC-based,^27,28 matrix factorization,^9,29 and location-scale^30–32 approaches. Merging of nonidentically acquired data adds an additional layer of technical variation due to changes in instrumentation and other analytical factors between the experiments. To address these problems, separate approaches for batch effects correction were first applied to the two experimental sample sets, attuned to their specific study designs. For ex616, which lacks uniform QC samples and contains nonrandomness in its sample run order assignment, the QC-RLSC method corrected for intrabatch effects by normalizing to LOESS curves of feature values from batch QC samples for each individual batch,²⁷ and then ComBat implemented in the sva R package was applied for the removal of interbatch effects.³² On the other hand, intra- and interbatch effects are handled in ex946 with the WaveICA method.⁹ Subsequently, quantities in both experiments were Z-transformed, followed by reassembly based on the previously determined feature matches. To determine the efficacy of this approach, Principal Components Analyses (PCA) and Principal Components Partial R-Square (PC–PR2)³³ analyses were performed. Projection of data onto the first two principal components enables a visual inspection of the major sources of variability. The PC–PR2 method merges PCA with multivariate linear regressions to express a measure of the proportion of variability explained by variables, which are chosen to be experiment, batch, and sample type in this application.³³

Following normalization and reassembly of aligned experimental data sets, annotations from prealignment steps were harmonized to obtain a list of known and unknown compounds for analysis. Metabolites were named when annotated in ex616 or ex946. Features annotated as adducts and fragments or derived from internal standard compounds were removed at this stage. The aligned and normalized data set is presented in Table S1.

Statistical and Bioinformatics Analysis

We conducted differential analyses comparing normalized metabolite levels in maternal plasma at the first and third trimesters (M1 and M3) and the umbilical cord blood (CB). Paired t test statistics, p-values, and Bonferroni-adjusted p-values were obtained for the merged experimental data, as well as ex616 and ex946 separately. Thresholds for nominal and statistical significance were set to 0.05 for unadjusted and adjusted p-values, respectively. Partial correlation networks were generated with the CorrelationCalculator program,³⁴ using annotated compounds from negative and positive mode data sets across the three time points. The resulting networks were visualized in Cytoscape³⁵ using Metscape.³⁶ Network nodes represent metabolites, while thick edges represent significant (p_adj < 0.05) partial correlations.

RESULTS

Disparate Data Alignment with metabCombiner

Experimental alignment consisted of within-batch preprocessing, merging between batches of the same experiment, and disparate alignment between experiments using metabCombiner (Figure 1A). Retention times of ex616 (30 min run) were mapped to ex946 (20 min run) by selecting m/z and abundance quantile (Q)-matched ordered pair anchors through which basis splines curves are fit. Figure 2A shows the computed RT mapping, alongside a visual confirmation of matched features for an example Extracted Ion Chromatogram (m/z = 365.1–365.11) from both experiments (Figure 2B). Table 1 lists the derived feature counts for the individual experimental batches and the final set of intersected analytes found across all batches. Metabolomics alignment steps consist of initial preprocessing, interbatch, and interexperimental merging, deriving 2343 positive and 1583 negative mode features in common. These lists were subsequently reduced to 1489 and 1054 features, respectively, after filtering for excess missingness and redundancy.

Figure 2.

Retention time mapping and feature matching. (A) Plotted spline fit generated by metabCombiner based on m/z and abundance quantile-matched feature pair anchors between ex616 (30 min total chromatography) and ex946 (20 min). (B) Selected peaks for the two experiments in m/z range 365.1–365.11 (negative mode), with matching colors for identical compounds, as assigned by metabCombiner.

Normalization of Merged Experimental Data

Figure 3 illustrates the pre- and postnormalization of features projected onto the first two principal components, colored by experimental batch (Figure 3A,C) or sample type (Figure 3B,D). PC–PR2 bar plots represent the proportion of overall metabolomics variability explained by experiment, batch, and sample type, along with the total variation explained by these covariates (R²) (Figure 3E,F). In the prenormalized data, interexperimental effects constitute the most significant source of variability, followed by sample type and within-experiment batch effects (Figure 3E). In the postnormalization data, metabolomics variability arising from experiment and batch effects was eliminated, whereas variation due to sample types was largely preserved (Figure 3F). After normalization, separation was observed between the maternal (M1 and M3) and the CB metabolite levels (Figure 3D), indicating substantial metabolic differences between mother and infants. These visuals highlight the success of the normalization procedure in significantly reducing the influences of non-biological study variables, while retaining biological differences.

Figure 3.

Pre- and postnormalization plots. (A,B) Prenormalized aligned data projected onto the first two principal components, colored by (A) experimental batch and (B) sample type. Initially, interexperiment and interbatch technical variability overshadow differences between sample types. (C,D) PCA plot of the postnormalized data set colored by (C) experimental batch and (D) sample type, showing the elimination of interexperimental and interbatch effects as a significant factor while biological variability between sample types remains prevalent. (E,F) PC–PR2 plots generated (E) before and (F) after normalization, showing the amount of metabolomics variation by experiment, batch, sample type, and the total variation explained by the covariates (R²).

Identifying Changes between Early Gestation/Late Gestation/Infant Metabolome

Using the aligned and normalized data, we identified differential metabolites between M1 and M3 (Figure 4A,B), representing change in the metabolome during gestation, and M3 and CB (Figure 4C,D), representing both transfer of metabolites to support fetal development in late gestation and maternal–child metabolic differences. Comparing M1 and M3 samples across both experiments, 32% of aligned positive mode features and 47% of aligned negative mode features changed significantly (adjusted p-value <0.05). Overall, 72% of the significant features increased from M1 to M3, demonstrating an increase of metabolite availability later in gestation to support fetal growth.³⁷ Of the 968 significantly differential features in both ionization modes, 158 were annotated compounds. In a similar comparison M3 and CB time points, 64% of aligned positive mode features and 73% of aligned negative mode features exhibited significant changes (adjusted p-value <0.05). Overall, 58% of the significant features were higher in CB compared to M3. Only 215 of the 1718 differential features were annotated compounds.

Figure 4.

Differential analysis vs retention time. Negative log-transformed p-values multiplied by the sign of the t statistic indicate direction and significance of changes between time points. M3 vs M1 (A) positive and (B) negative mode data sets, CB vs M3 (C) positive and (D) negative mode data sets represent greater differences between these sample types. The black dotted line indicates the Bonferroni significance cutoff (q = 0.05).

Differential analysis results were compared for the merged metabolomics data set with those obtained from experiment ex616 and ex946 separately, using the same set of 1490 and 1054 overlapping compounds in the positive and negative ionization modes, respectively. There was an increase in the number of significant features in both M3 vs M1 and CB vs M3 for the merged metabolomics data set vs the experiments separately (Table S2A). This illustrates the advantage of increased sample sizes obtained by merging experimental data, even when acquired under disparate conditions. Among significant differential features found in common between ex616 and ex946, the majority changed in the same direction, with 99–100% consistency in M1 vs M3 and 96–97% consistency in CB vs M3 (Table S2B). See Tables S2 and S3 for the full report of differential analysis results.

Data-Driven Network Analysis

To further analyze the observed metabolite changes we constructed a partial correlation network (PCN) of all identified compounds (Figure 5). For adequate statistical power, PCNs require high sample (n) to predictor (p) count ratios,³⁴ which could only be attained by assembling the complete sample sets from both experiments in this study. Despite being agnostic with respect to the biological pathway or chemical class information the resulting networks grouped together similar metabolites, enabling visualization and biological interpretation of the data. Individual subnetworks containing shared chemical classes are depicted in Figure S1.

Figure 5.

Partial correlation network constructed from metabolomics data. Nodes represent metabolites and edges represent partial correlations computed from the experimental data. Nodes representing significant metabolites (q-value < 0.05) have bold borders and node colors are based on t statistics. (A) M3 vs M1; (B) CB vs M3. The dotted lines outline subnetworks that include metabolites from different chemical classes. See Figure S1 for a detailed view of each subnetwork.

Early vs Late Gestation

Between M1 and M3, we observed an increase of long chain and very-long chain free fatty acids (FFA, 80% significantly increased) (Figure S1a) and lipid species containing long chain fatty acids, including diacylglycerides (DGs, 87% significantly increased) (Figure S1b), ceramides (CER, 80% significantly increased), and sphingomyelins (SM, 50% significantly increased) (Figure S1c). However, not every lipid species increased between M1 and M3. Only 41% of phosphatidylcholine (PC) and phosphatidylethanolamine (PE) lipids changed between M1 and M3. Furthermore, most of the differential PCs/PEs were polyunsaturated very-long chain metabolites that decreased between M1 and M3 (Figure S1d). Several medium and long chain acylcarnitines increased between M1 and M3 (AC 10:3, 11:1. 12:1, 14:0, 14:1, 16:2) (Figure S1e). There were decreases in essential amino acids, including tryptophan and BCAA metabolites (ketoleucine, AC 4:0, and AC 5:1) (Figure S1g). Variations in bile acid fluctuation across pregnancy were observed, with significant decreases from M1 to M3 in glycocholate and glycochenodeoxycholate (primary bile acids), and deoxycholate, glycodeoxycholate, and glycoursodeoxycholate (secondary bile acids), countered by significant increases from M1 to M3 in taurocholate and tauro-alpha/beta-muricholate (Figure S1f). Parturition-induced increases in cortisol and 11-deoxycortisol were also observed.²¹

MS/MS library searches were applied for the most significantly differential unnamed features between M1 and M3 in both ionization modes, using NIST Hybrid Search³⁸ and the NIST20 library. Top matches for most of these searches consisted of steroids with associated sulfidated and glucorinidated modifications. The top feature (adjusted p = 3.3 × 10⁻⁵²) in the negative ionization mode (m/z = 463.1967 Da, RT = 6.64 min in ex946) returns a match to estriol-16-glucoronide, a glucosiduronic acid metabolite of estriol that has been previously isolated in the urine and amniotic fluid of mothers during pregnancy.^39,40 Other matches include epiandrosterone sulfate, testosterone glucoronide, 7α,17α-dimethyl-5β-androstane-3α,17β-diol glucuronide, and 5α-pregnane-3α,17α-diol-20-one 3-sulfate. MS/MS head–tail plots for features with compound match NIST Hybrid Search scores above 600 are illustrated in Figure S2; their associated sample vs normalized abundance plots can be found in Figure S3.

Late Gestation vs Infant Metabolome

Between M3 and CB, differential lipids depended on lipid species, chain length, and saturation. Overall, most long chain fatty acids were higher in M3 compared to CB (77%), except two very long chain polyunsaturated fatty acids that were elevated in CB (FA 26:4 and FA 26:5) (Figure S1a). Limited trends were observed for phospholipids, lysophospholipids, and diacylglycerols, apart from several long-chain and very-long chain monounsaturated and polyunsaturated PCs and PEs being elevated in CB (Figure S1b–d).

Cord blood levels of amino acids were consistently higher than M3, including branched chain amino acid metabolites, such as AC 4:0, AC 5:1, isovalerylcarnitine, and l-gamma-glutayl-l-isoleucine, and aromatic amino acid metabolites, such as tryptophan, phenylalanine, glutamyl-phenylalanine, indole lactate, and kynurenine (Figure S1g). Methionine, a major contributor to one carbon metabolism, was higher in CB. Variations in differential bile acids were observed with lower levels in CB of glycochenodeoxycholate (primary bile acid), deoxycholate, glycodeoxycholate, glycoursodeoxycholate, ursodeoxycholate (secondary bile acids) and higher levels of CB of glycocholate alpha-muricholate (primary bile acids), and taurocholate and taurodeoxycholate (secondary bile acids) (Figure S1f). Like the M3 vs M1 comparison, MS/MS spectral searches for the most significantly differential features between CB and M3 yielded hits to sulfidated and glucorinidated steroids. Notable hormonal differences include higher levels of androsterone sulfate, 11-deoxycortisol, and cortisol in M3 and higher levels of dehydroepiandrosterone sulfate, pregnenolone sulfate, 11-beta-hydroxyandrost-4-ene-3-17-dione, and 17-alpha-20-alpha-dihydroxypregn-4-en-3-one in CB (Figure S1g).

DISCUSSION

We developed a framework for the alignment of known and unknown features between experimental data sets and normalization to remove interbatch and interexperiment variation. Multiple analytical factors differed between the two metabolite profiling experiments, ex616 and ex946, staged 3 years apart. First, a more sensitive QTOF instrument used in ex946 derived substantially more extracted features, with most extracted signals inevitably lacking one-to-one matches with the smaller list of ex616 features. Second, the shorter total chromatography run time in ex946 led to gaps of up to 13 min in measured RTs for identical compounds between experiments. Alignment of these extracted feature matrices required mapping between chromatographic RTs or bypassing RT comparisons altogether.⁴¹ metabCombiner achieves an accurate and comprehensive correspondence of features, despite the differences in analytical methods. LC-MS metabolomics meta-analysis studies^42–44 and methods^45,46 typically require replicated experimental conditions or limit their scope to shared known compounds. To our knowledge, this is the first study in which metabolomics measurements of known and unknown compounds acquired with altered experimental protocols were aligned and analyzed together in an unbiased manner.

Due to hardware limitations, individual batches of both experiments are treated as distinct tables prior to their alignment with metabCombiner. While it is possible to simultaneously align LC-MS metabolomics batches of the same experiment using conventional open-source preprocessing tools, such as XCMS,⁴⁷ this has some disadvantages, such as a tendency to misalign features in large-scale experiments due to excess chromatographic drift^5,48 and generation of false positive features due to peak detection inaccuracies.⁴⁹ The application of our methods reduced the feature lists to a common intersection of compounds found across all experimental batches, facilitating a well-powered statistical analysis.

Harmonization of metabolomics data pooled across experiments and studies requires novel normalization strategies to account for significant variation between laboratories, instruments, and protocols, while preserving biological variability. In this study, conventional intrabatch and interbatch effects correction approaches were first applied to two untargeted metabolomics experimental sets acquired by the same laboratory through different protocols, followed by Z-transformation and reassembly of the matched features. This framework requires a similarly balanced study design between experimental sets as Z-transform normalization assumes similar abundance distributions.⁹ The composition of M1, M3, CB, and quality control samples is proportionally similar between the experiments, though differences in how QC samples were prepared (per-batch pooled plasma in ex616 vs all ex946 sample pooled plasma) led to the use of alternative batch effects correction approaches. A recent publication suggests a linear mixed model approach for reducing interstudy variation, demonstrating higher Intraclass Correlation Coefficient values for shared quantified sample measurements among eight pooled studies.⁵⁰ Successfully adapting such approaches to untargeted metabolomics data pooled from disparate sources may need to consider additional factors beyond those found in targeted metabolomics.

We assessed differences in the maternal plasma metabolome (M3 vs M1) and between the maternal and infant plasma (CB vs M3). Given the analytical differences between ex616 and ex946, it was important to assess whether merging would enhance a differential signal between sample groups or dilute it. The results show that the number of significant results increased with sample count, and that 96–100% significant changes shared between the two experiments occurred in the same direction (increasing or decreasing). Changes that occur in inconsistent directions could be indicative of inaccurate alignment between nonidentical analytes, improper normalization between or within batches, or quantitation inconsistencies arising in the experimental or computational preprocessing stages.

The increase of lipid species containing long chain fatty acids from M1 to M3 is consistent with previous analyses by our group using a lipidomics platform,²¹ suggesting the mobilization of FFA to support fetal brain development late in gestation.⁵¹ This phenomenon is a result of an increase in maternal insulin resistance and elevation of lipoproteins and triglycerides (TG) to support the metabolic changes for pregnancy and fetal growth.³⁷ The decrease in polyunsaturated long-chain metabolites between M1 and M3 suggests specific roles of lipid types, as previous research has demonstrated that lysated PC/PE lipids transport essential polyunsaturated fatty acids through the fetus to support fetal brain development,⁵² as the fetus has limited desaturase activity.⁵³ Of the metabolite classes considered in this study, PCs/PEs were least consistent between the two experiments, which may explain why we did not observe increases in saturated long chain PCs/PEs across pregnancy, as reported previously.²¹ Previously, Luan et al. (2014) similarly explored differences in the maternal plasma across pregnancy trimesters, uncovering changes in metabolites related to biopterin and phospholipid metabolism, as well as fatty acid oxidation.⁵⁴ Their study is cross-sectional, whereas our study consists of longitudinal observations. A potential future direction would be to align data from these two studies and determine how consistently overlapped compounds change between the first and third trimesters of pregnancy.

This study expands on our previous work to reveal variations in fatty acid oxidation intermediates between M1 and M3. Increases in acylcarnitines between M1 and M3 are consistent with a targeted analysis conducted for a pilot study on a subset of these women⁵⁵ and indicates an increased reliance on fat metabolism for energy later in gestation.⁵⁶ Decreases in essential amino acids is also consistent with other cohorts,⁵⁷ as protein anabolism is favored during pregnancy.⁵⁸ Lastly, increases in cortisol and deoxycortisol between M1 and M3 probably reflects collection of the M3 sample during parturition.⁵⁹ Using MS/MS spectral search, several of the features that most significantly increased from M1 to M3 were identified as steroid hormones. Our group has previously demonstrated that steroids increase during pregnancy to promote fetal and placenta development and ultimately parturition.⁶⁰ In our previous work,²¹ multiple lipid species—PCs, PEs, SMs, and triacylglycerols (TG)—containing long and very long chain polyunsaturated fatty acids were higher in CB than M3, provisioning essential fatty acids for the developing fetal brain.⁵¹ The untargeted metabolomics platform supported these results, demonstrating significant increases of several long chain polyunsaturated PCs and PEs in CB. There was no discernible pattern in lysophospholipid fluctuations between M3 and CB, with 50% exhibiting no change, 20% increasing in CB, and 30% decreasing in CB. We have previously demonstrated that lysophospholipids with varying chain length and saturation were elevated in CB compared to maternal plasma.²¹

Higher levels of cord blood levels of amino acids and branched amino acid metabolites in CB relative to M3 is consistent with previous studies⁶¹ and with the high levels of branched-chain aminotranferases in human placenta tissue.⁶² Elevated methionine may have implications in DNA methylation supporting fetal development.⁶³

CONCLUSION

In summary, we present a framework for merging LC-MS metabolomics data sets and overcoming substantial technical variation due to changes in protocols, instrumentation, and analytical conditions. Statistical analysis of the aligned data set revealed significant metabolic alterations between M1, M3, and CB, including coordinated changes observed between the first and third trimesters of gestation as well as infant cord blood consistent with previous studies. This work establishes the viability and utility of merging disparately aligned metabolomics data sets, providing greater biological insight with a combined sample set than with separate experimental sets alone. The data acquired for this study were generated within one institution using two similar, but nonidentical, protocols. The vast landscape of LC-MS technologies and methods for metabolite profiling among laboratories creates many challenges for data set merging and meta-analysis. Overcoming these challenges unlocks opportunities for the field of metabolomics that would be inaccessible through the traditional assumption of replicated conditions.

Data Availability Statement

The LC-MS metabolomics data for this study can be found in the Metabolomics Workbench with the data set identifiers ST002134 and ST002135.