An enriched biosignature of gut microbiota-dependent metabolites characterizes maternal plasma in a mouse model of fetal alcohol spectrum disorder

Prenatal alcohol exposure (PAE) causes permanent cognitive disability. The enteric microbiome generates microbial-dependent products (MDPs) that may contribute to disorders including autism, depression, and anxiety; it is unknown whether similar alterations occur in PAE. Using a mouse PAE model, we performed untargeted metabolome analyses upon the maternal–fetal dyad at gestational day 17.5. Hierarchical clustering by principal component analysis and Pearson’s correlation of maternal plasma (813 metabolites) both identified MDPs as significant predictors for PAE. The majority were phenolic acids enriched in PAE. Correlational network analyses revealed that alcohol altered plasma MDP-metabolite relationships, and alcohol-exposed maternal plasma was characterized by a subnetwork dominated by phenolic acids. Twenty-nine MDPs were detected in fetal liver and sixteen in fetal brain, where their impact is unknown. Several of these, including 4-ethylphenylsulfate, oxindole, indolepropionate, p-cresol sulfate, catechol sulfate, and salicylate, are implicated in other neurological disorders. We conclude that MDPs constitute a characteristic biosignature that distinguishes PAE. These MDPs are abundant in human plasma, where they influence physiology and disease. Their altered abundance here may reflect alcohol’s known effects on microbiota composition and gut permeability. We propose that the maternal microbiome and its MDPs are a previously unrecognized influence upon the pathologies that typify PAE.


Results
Litter characteristics. The alcohol dose used (3 g/kg) caused a mean blood alcohol concentration of 211 ± 14 mg/dl at 30 min post-gavage, and the mice were inebriated but did not pass out. Alcohol exposure (ALC) did not affect maternal food intake 41 or overall weight gain (Supplementary Table S1 online); the ALC dams had a non-significant trend to reduced weight gain during the alcohol exposure period (embryonic day (E) 8.5-E17.5) compared to controls (CON, 11.22 ± 0.47 g; ALC, 9.98 ± 0.44 g; p < 0.07). Prenatal alcohol exposure did not affect litter size (p = 0.31) or fetal survival (p = 0.69) at E17.5, and no adverse outcomes were observed in dam or fetus. Alcohol-exposed maternal plasma is enriched in MDPs. Untargeted metabolite analysis identified 813 biochemicals in maternal plasma, 733 with known chemical structures and 80 that were unknown. Of the 813 metabolites, 146 had significantly altered representation (q < 0.05 by Mann-Whitney U-test followed by Benjamini-Hochberg correction) in response to PAE. Principle Component Analysis (PCA) of the metabolite profiles showed that alcohol-exposure is a clear driver of variance within the metabolic profiles, and placed one dam (ALC-6) as an outlier (Supplementary Figure S1A,B online). This dam's plasma ethyl glucuronide level was just 14.0% of that for the other alcohol-exposed dams, suggesting a gavage error, and she was removed from further analysis. Repeating the PCA with omission of ALC-6 revealed that the metabolomic profile explained the separation of the samples by intervention. PC1 explained 23.1% of the sample variance, and PC2 explained an additional 15.9% (Fig. 1a), and visual inspection of the data set revealed that MDPs were among the strongest drivers of PC1 and PC2 (Fig. 1b). Analysis of the log-fold change q-values using T-statistic similarly found that MDPs were over-represented (Fig. 1c), and of 146 metabolites having a q ≤ 0.05, 28.1% (N = 41) were MDPs, although they comprised 10.5% of the 733 known metabolites. Housing assignment can also affect enteric microbiome composition 42 ; remapping the PCA results against housing assignment affirmed that cage assignment did not influence analyte distribution or abundance (Supplemental Fig. S1C).
MDPs comprise a significant biosignature in plasma of ALC dams. We utilized Hierarchical Clustering of Principal Components (HCPC) to understand the relationships among these metabolite features. This placed the 813 metabolites into five, evenly distributed clusters using Ward's method. The MDPs were unevenly distributed across the clusters, and a majority segregated into cluster 1 (57.8%, 41 of 71 MDPs), followed by clusters 3 (11/71) and 5 (11/71) (Fig. 2a). Clusters 2 and 4 were dominated by endogenous metabolites. This clustering of MDPs primarily reflected their contribution to PC1, which captured the greatest class separation between ALC and control (Fig. 2b). The MDPs in Cluster 1 were predominantly plant-derived aromatics, and all were enriched in ALC (Table 2; Supplementary Table S2 online). In contrast, the MDPs in cluster 5 were enriched in control plasma and they were mostly (9/11) secondary bile acids. Clusters 3 and 4 were highly skewed towards other Principal Component Dimensions, and repeating the PCA and PLSDA analysis according to PC2 (Supplementary Figure S2 online) suggested that dimension 2 modeled the time of plasma collection following the intervention. This cluster separation was relevant only for the ALC dataset and not the controls. Because this influence did not involve the MDPs, it is the focus of a separate investigation described elsewhere. In summary, the HCPC and PLSDA analysis revealed that the MDPs clustered by treatment response based on the PC loadings. These MDPs had a disproportionate influence in explaining exposure variance within the metabolite dataset, and their influence was defined by their molecular structure, in addition to their relative abundance and p-value. Correlation analysis further informs whether influences in addition to exposure drive the metabolite relationships and variance across the dataset. We repeated the hierarchical clustering using Spearman's correlation (Fig. 2c). The strong association between the plant phenolics was retained, and 47/71 were correlated within Cluster 5, comprising 27.8% of that cluster, and these were all enriched by alcohol exposure (Table 3, Supplementary Table S3). New associations also emerged, and Cluster 1 (13/71) included a mixture of phenolics, indoles, and secondary bile acids largely unaffected by alcohol-exposure. These findings further support that alcohol exposure strongly influenced plasma MDP content, and their relationships further depended on chemical structure and metabolic fate.
Metabolite classes that tightly cluster together in the correlation analysis share a consistent response to treatment. Similarly, metabolites that are downstream products of a shared cellular process affected by alcohol also will maintain equilibrium with each other during the analysis. Thus, highly conserved correlations likely represent a metabolite set that exists within a molecular equilibrium. The consistent clustering of MDPs in both the HCPC and Spearman's correlation suggested the existence of such relationships. To investigate this, we filtered the correlation matrix to correlations greater than 0.9, and subjected these to network correlation analysis, and segregated by treatment. This yielded very different network structures for Control and ALC maternal plasma (Fig. 3). The network architecture of the controls was comprised of two dense hubs connected by tightly-linked interactions, and each joined to separate satellites that were in turn weakly linked (Fig. 3a). In contrast, the ALC plasma network architecture was dominated by just a single dense hub that was more loosely connected with a Figure 1. Plasma microbial-derived metabolites distinguish alcohol-exposed and control dams. (a) Dimension 1 and 2 of the PCA on the scaled metabolite profiles within plasma of nine control (CON) and eight alcoholexposed (ALC) dams. (b) PCA biplot with microbial-derived metabolites overlaid onto the samples plot. PCA values for the MDPs are presented in Supplemental Table S2. (c) T-statistic plot of all 813 metabolites (arranged alphabetically along the x-axis) against their log10 q-values. The black dashed line indicates the cut-off for the FDR adjusted value of q < 0.05, and the red dashed line indicates q < 0.01. Red dots indicate microbial-derived metabolites having q < 0.05. Sample size is n = 9 control and n = 8 alcohol-exposed dams. www.nature.com/scientificreports/ second, more diffuse hub, each with an adjoining smaller satellite (Fig. 3b). The MDPs held quite different relationships within these two architectural structures, and the plant phenols formed a dense mini-structure in ALC, whereas no such network appeared in controls; the latter's largest MDP set was a mix of phenolics, indoles, and sugar acids. A parallel analysis that focused on Spearman correlations less than 0.9 revealed similarly divergent networks, such that the Control network ( Fig. 3c) featured far fewer metabolites than did ALC (Fig. 3d), signaling that the plasma metabolite profile of ALC was characterized by a loss of tight regulatory control. This is endorsed by the more dispersed structure of the ALC network and suggests an overall weakening of metabolite relationships that may be a product of dysregulated metabolism. Overall, the analyses revealed that alcohol-exposure altered the relationships between MDPs and endogenous metabolites, and endorsed the MDP biosignature for alcohol-exposed maternal plasma. Additional insight was obtained by merging the Control and Alcohol plasma datasets, and again performing network analysis filtered by the Spearman's correlations. For the negative correlations, this only yielded twonode subnetworks, none of which contained MDPs. For the positively correlated metabolites, this yielded a network dominated by MDPs enriched in ALC (Fig. 4). This included a tightly correlated subnetwork of nineteen plant phenolics (hippurates, catechols, salicins, phenols) and a smaller, linked subnetwork of sugar acids that was further linked with endogenous-derived sugar acids. Also included were several unknowns including one Table 1. Microbial-derived products detected in maternal plasma. N = 9 control and N-8 alcohol-exposed dams; p-value by Mann-Whitney U-test; q-value by Benjamini-Hochberg FDR correction. Relative Abundance presents median ± median absolute deviation (MAD). Alc Alcohol-treated, Cont control, F-C foldchange.

Metabolite
HMDB ID  (Table 4) and was dominated by secondary bile acids and  (c), a positive principal component score indicates the metabolite has increased abundance in response to alcohol; negative scores signify reduced abundance. Sample size is n = 9 control and n = 8 alcohol-exposed dams.  (Fig. 5a) affirmed their lower representation among the significantly altered biofeatures in maternal liver (10/205), suggesting that MDPs were not major drivers of metabolite variance in response to alcohol for this tissue. Many metabolites within maternal plasma exchange readily across the placenta and become bioavailable to the fetus. Thirty-one MDPs were detected in placenta, including plant phenolics, indoles, and secondary bile acids (Table 4). Of these, alcohol significantly altered the abundance of 11 MDPs, and seven plant-derived phenolics and several betaines were enriched, while indolepropionate and gluconate were reduced.

p-Value q-Value
The placental data suggested that MDPs might circulate within the fetus. Although fetal plasma was too scant for analysis, other fetal tissues were readily characterized (Table 4). Thirty MDPs were detected in fetal liver and/or brain. The fetal liver profile largely replicated that of maternal liver, and all but five (enterolactone sulfate and four secondary bile acids) of the 32 metabolites present in maternal liver were also detected in fetal liver. Responses to alcohol also trended similarly, with enrichments in hippurate (2.43-fold), catechol sulfate (1.83-fold), salicylate (1.46-fold), phenol sulfate (1.34-fold), and ergothioneine (1.29-fold). Although not MDPs, the phytosterols beta-sitosterol (1.76-fold) and campesterol (1.31-fold) were also elevated.

Discussion
The enteric microbiome generates a complex spectrum of biochemicals that have a substantial influence on the host [14][15][16][17] . This is the first study to document that PAE significantly alters this biochemical profile within the maternal-fetal dyad. This metabolite profile is derived from and influenced by the composition of the enteric microbiota, and the changes documented here are consistent with alcohol's known ability to alter that composition 36,37 . Growing evidence demonstrates that microbe-derived products are mechanistic in the pathologies that underlie alcohol-related organ damage [33][34][35] , and our findings suggest that parallel mechanisms may operate during PAE with similar pathological consequences. Importantly, we show that these biochemicals cross the placenta and circulate within the fetus, where they could directly impact development. Indeed, many of these biochemicals having significantly altered representation, mostly plant-derived phenolics and secondary bile acids, have welldocumented effects upon host physiological and cellular processes 16,18,44,45 . Our demonstration that alcohol alters their abundance within both mother and fetus introduces a novel mechanism by which PAE could alter fetal development, and thus these findings have clinical relevance.
Moreover, these biochemicals form a plasma biosignature that distinguishes the PAE pregnancies. There is substantial interest in identifying biomarkers of alcohol exposure as these enable clinicians to focus interventions on those pregnancies at greatest risk. In addition to the established markers phosphatidylethanol and ethyl-glucuronide 8,9 , recent studies have identified microRNA 46 and cytokine-chemokine 47 plasma signatures that may be selective for alcohol-exposed pregnancies. An MDP signature could complement those measures, as some features were enriched seven-to 13-fold in this model. Although the composition of any microbial biosignature is shaped by considerations including diet, the enteric community's taxonomic structure, exome Table 3. MDPs by cluster from the hierarchical clustering on Spearman's correlation. Spearman's coefficients for each metabolite are presented in Supplemental Table S3, and are depicted in Fig. 2c. N = 9 control and N = 8 alcohol-treated dams.

Cluster 5 metabolites (N = 45)
12-Dehydrocholate, 2,6-dihydroxybenzoic acid, 2-hydroxyhippurate (salicylurate), 3-(3-hydroxyphenyl)propionate sulfate, 3-(4-hydroxyphenyl)propionate, 3-ethylphenylsulfate, 3-formylindole, 3-hydroxypyridine sulfate, 3-indoleglyoxylic acid, 3 www.nature.com/scientificreports/ profile, and host species and sex 15,48,49 , our data lend proof-of-concept for the existence of such a biosignature that may complement existing markers and further enhance their specificity to detect PAE. How might these microbial metabolites impact fetal development and alcohol-related pathologies? In this study, the dominant microbial compounds enriched by alcohol were plant-derived aromatics, mostly phenolic acids that originate from the fermentation of ingested lignin bedding and starch-bound flavonoids 18,43,50 . These same phytochemicals are abundant in edible plants, and the human enterocyte and microbiota have similar capabilities to release, convert, and absorb these compounds 18 . Humans consume an estimated two-plus grams daily of plant phytochemicals from foods and beverages, and plasma levels typically range in the nanomolar to low micromolar range 51,52 . This is the first report that these compounds circulate within the fetus. These compounds have a short half-life due to their rapid excretion 44 ; that alcohol enriched both their aglycone and conjugated forms suggests that it enhanced and/or prolonged their enteric metabolism and intestinal absorption, as well as their phase II conversion. In human studies, these compounds are typically associated with improved health outcomes and reduced all-cause mortality. Mechanistically, phenolic acids improve vascular tone through stimulation of endothelial Nrf2 and nitric oxide signaling, and have anti-inflammatory actions through their inhibition of pro-oxidant enzyme-signaling cascades 18,44 ; thus, their elevation in PAE may potentially mitigate www.nature.com/scientificreports/ some of alcohol's damage to the mother-fetal dyad. They also act as prebiotics and directly alter the microbiome composition 52 . Highly processed Western-style diets are low in lignins and flavonoids, and their enrichment here represent a novel means by which diet may modulate FASD outcome. The microbiota-derived secondary bile acids also defined the ALC dams, but were negative drivers within the Principal Components and Pearson Correlation analyses. Along with their parent primary bile acids, they comprised a correlated network that suggests a shared mechanistic response to PAE. Bile acids and the enteric microbiota operate in a two-way interaction that governs both bile acid metabolism and microbiota composition 45,53 . Our data suggest that alcohol altered that regulatory relationship, and this is consistent with its known effects on host-microbial bile acid pools, wherein chronic alcohol abuse elevates secondary bile acid levels 54,55 , perhaps Positive correlation analysis identifies a network that is enriched in microbial metabolites and is at statistical equilibrium within maternal plasma. Images were generated in Cytoscape (version 3.7.2) using R-Cy3 (version 2.6.3), wherein the network edges represent between-node Spearman's correlations > 0.90; the distance between nodes indicates strength of interaction. Colors as in Fig. 2; red indicates MDPs, blue indicates endogenous compounds, and green are unknown compounds. The compound X-17010 is likely the MDP 4-vinylcatechol sulfate, based on its molecular mass.  www.nature.com/scientificreports/ through dysregulation of hepatic bile acid synthesis 56 . The reductions here may reflect our shorter exposure (days vs. months) and perhaps influences from the pregnancy state 57 . We could not infer which microbial populations mediated these reductions because secondary bile acid metabolism is redundant across phyla 45 . The elevated taurine conjugates in the alcohol-exposed maternal liver implicate reduced microbial deconjugation and/or hepatic amidation as additional modifying mechanisms. Secondary bile acids modulate numerous processes. They stimulate the production of antimicrobial peptides that suppress the growth of proinflammatory, gramnegative microbes 58,59 , and their reductions here suggest a means by which alcohol promotes the proinflammatory environment that worsens fetal development 47,60 . Bile acid interactions with their RXR, FXR, LXRα, and GPBAR1 receptors affect insulin sensitivity, adiposity, and lipid metabolism, conditions that independently worsen gestational outcomes 61 . Secondary bile acids were recently detected in the porcine fetus 62 , suggesting their fetal presence is not unique to rodents; however, any biological impact upon fetal development is currently unknown. Taken together, these data suggest that alcohol disturbs microbiota-bile acid interactions in a manner that could negatively impact maternal-fetal health. Additional MDPs altered by PAE included indoles and betaine derivatives. The betaine-like compounds ergothionine and hercyine scavenge free radicals and reduce oxidative damage 63 , and their elevation in PAE may confer some protection. Indoles are generated by microbial tryptophanase and can influence brain and behavior. Oxindole, which was elevated in plasma from ALC dams, promotes anxiety-like behaviors in the open field and elevated plus-maze tests in rats 26 , behaviors also seen in PAE. Conversely, indolepropionate confers protection against neuroinflammation and TLR4 signaling through interactions with the microglial arylhydrocarbon receptor 22,28 , and sustains gut integrity through the pregnane X receptor (PXR) 64 ; its sharp reduction in ALC plasma and placenta is consistent with alcohol's pro-inflammatory actions 38,60 . Indoles also compete with amino acid and neurotransmitter efflux at the blood-brain barrier, and are functionally linked with anxiety, depression, cognitive impairment, and Parkinson's disease 20,26,65,66 . As physiologically relevant agonists for the arylhydrocarbon receptor, they modulate not only immune function but also xenobiotic responses and insulin sensitivity 67 . We detected at least four indoles in fetal brain (indoleacetate, indolelactate, indolepropionate, 3-formylindole), and because many indole derivatives have yet to be characterized functionally, their impact upon neurodevelopment merits additional investigation.

F-C q-value F-C q-value F-C q-value F-C q-value F-C q-value
This study has several important limitations. The first is that not all MDPs could be investigated. Many remain unannotated and likely comprise some of the unknown metabolites detected here. We also did not analyze the gut microbiota, and thus do not know if and how alcohol affects its composition in pregnancy. This study was not designed to distinguish those metabolites having a dual microbial-host origin, such as lipids, organic acids, and polyamines, nor does the methodology detect the larger MDPs that contribute to alcohol's proinflammatory actions, such as LPS 33,38,64 . Finally, we cannot distinguish the relative contributions of enteric synthesis and cecal permeability to the elevated MDP abundance. As alcohol promotes both dysbiosis and gut permeability [33][34][35][36][37][38] , both mechanisms likely contribute; additional studies will inform this question. Table 4. Comparison of MDP profiles and their abundance in maternal plasma and liver, placenta, and fetal liver and brain. F-C, fold change; q-value by Mann-Whitney U-test, followed by Benjamini-Hochberg FDR adjustment. "-" indicates not detected. N = 9 control and N = 8 alcohol-treated dams and their fetuses.

Biochemical name
Maternal plasma Maternal liver Placenta Fetal liver Fetal brain www.nature.com/scientificreports/ In summary, alcohol alters the maternal plasma MDP profile, and by inference perhaps the microbiota composition that produced them. Several of the MDPs elevated by PAE (catechol sulfate, 4-ethylphenylsulfate, erythritol, indolepropionate, oxindole, p-cresol, salicylate) have been implicated in neuroinflammation, depression, anxiety, and autism 22,26,28,30,65,66 , outcomes also characteristic for PAE 1,2 . Other compounds may confer benefits through effects on vascular tone and anti-inflammatory actions, and thus could mitigate some of alcohol's damage to the fetus. These MDPs circulate within the fetus, where their impact is unknown. Their enrichment, particularly in phenolic acids, constitutes a characteristic biosignature that distinguishes the PAE pregnancies, and their enrichment might also signal the presence of proinflammatory MDPs such as LPS. Together, these data suggest the novel hypothesis that the maternal microbiome may be an important mechanistic driver in the pathologies that underlie FASD.  Experimental blocking. We evaluated maternal plasma, maternal liver, placenta (with decidua removed), fetal liver, and fetal brain, from 9 Control and 9 ALC dams and their litters. To obtain sufficient fetal tissue for analysis, it was necessary to pool the fetuses. Specifically, for each litter, we held uterine position constant and defined Fetus 1 as occupying the position closest to the right ovary. Fetuses were numbered consecutively thereafter. Selecting fetuses 1 through 4, we combined half of fetal livers 1 through 4, and half of fetal brains 1 through 4, and submitted each pooled sample for metabolome analysis. Thus, each dam is an individual biological sample, and each fetal sample is the pool from Fetus 1, 2, 3, and 4. Each placental sample was derived from half of placental 1 and 2, because this tissue was larger. For each group (ALC, Control), we subjected nine individual dams and nine fetal pools to metabolome analysis.

F-C q-value F-C q-value F-C q-value F-C q-value F-C q-value
Metabolite analysis. Untargeted metabolite analysis was performed by Metabolon (Morrisville, NC), and their detailed methods are presented in Supplemental Methods S1. To summarize, samples were treated with methanol to remove protein, and then divided into five aliquots for reverse phase (RP)/UPLC-MS/MS with positive ion mode electrospray ionization (ESI, 2 samples), RP/UPLC-MS/MS with negative ion mode ESI (one sample), and HILIC/UPLC-MS/MS with negative ion mode ESI (one sample); a fifth sample was reserved for back-up. Quality controls include technical replicates of pooled experimental samples, extracted water and solvent blanks, addition of recovery standards to monitor variability and efficiency, and internal standards that assessed instrument variability and aided chromatographic alignment. Sampling order was randomized across each platform run. Compounds were identified by comparison to library entries of purified standards or recurrent unknown entities. Identification was based on the criteria of retention time/index, match to a mass to charge ratio ± 10 ppm, and chromatographic data (MS/MS spectrum). Proprietary visualization and interpretation software were used to confirm peak identities. Peaks were quantified using area-under-the-curve.

Statistical analyses of metabolites.
In the initial analysis, we tested for unequal variance between the Control and ALC groups using Shapiro-Wilks test, and tested for normality using the Levine's test, followed by analysis for significance using the Mann-Whitney U-test. For values that were missing, we imputed the minimum value obtained for that metabolite in that tissue, and Supplemental Table S5 presents the raw LC-MS/MS dataset, as provided by Metabolon. P-values were adjusted for multiple testing correction using the Benjamini-Hochberg False Discovery Rate (FDR) correction 69 , and are presented as q-values. Analyses were performed in ArrayStudio on log transformed data 70 . For analyses that were not standard within ArrayStudio, the program R (version 3.6.1) 71 ) was used. Fold-change was determined as the difference between group averages, placing ALC in the numerator, and is reported in log2 values. For the discriminant analysis between ALC and Controls, the plasma data were scaled to a zero mean with a standard deviation of one for each metabolite, and were then run through multivariate analysis. Principal Component Analysis (PCA) was performed using FactoMineR (version 2.3) 72 to test for separation between the treatment and control groups, and to identify outliers and question trends that supported class separation of the experimental design; findings were visualized using Factoextra R (version 1.0.7) 73 . Metabolite-metabolite correlation analysis was calculated using Spearman's correlation on un-scaled data, and were analyzed and visualized in ggplot2 (version 3.3.0) 74 using hierarchical clustering for comparison with Hierarchical Clustering of Principal Components (HCPC). The metabolite-metabolite correlation matrix was used to construct a network visualization to explore similarly affected metabolites. The correlation networks were constructed in Cytoscape (version 3.7.2) 75 using the RCy3 package (version 2.6.3) 76 and aMatReader (version 1.1.3) 77 . Only Spearman's correlations above 0.9 were included in the network, and nodes were overlaid with descriptive statistics including q-value and log-fold change (logFC). To evaluate metabolites acting in concert, we explored the loadings plot of the sample PCA again using FactoMineR on transposed scaled data and visualized in factoextra. Hierarchical clustering analysis was used to cluster the regression factor scores of the loadings identified in the PCA using Ward's minimum variance method and squared Euclidean distance in FactoMineR. This analysis, including the final k-means clustering, was identically applied to the correlation matrix and visualized in ggplot.