The Gut Microbiome and Metabolomics Profiles of Restricting and Binge-Purging Type Anorexia Nervosa

Alterations in the gut microbiome and fecal metabolites have been detected in anorexia nervosa (AN), but differences in those profiles between restricting AN (ANR) and binge-purging AN (ANBP) type have not been explored. We made a secondary analysis of our previous data concerning microbiome and metabolomics profiles of 17 ANR women, six ANBP women and 20 healthy controls (HC). Twelve fecal metabolites differentiating ANR patients, ANBP patients and HC were identified. Both patient groups showed decreased intra-individual bacterial richness with respect to healthy controls (HC). Compared to ANR subjects, ANBP patients had a significant increase in relative abundances of Bifidobacterium, Bifidobacteriaceae, Bifidobacteriales, and Eubacteriacae and a significant decrease in relative abundances of Odoribacter, Haemophilus, Pasteurellaceae, and Pasteurellales. The heatmaps of the relationships of selected fecal metabolites with microbial families showed different structures among the three groups, with the heatmap of ANBP patients being drastically different from that of HC, while that of ANR patients resulted more similar to HC. These findings, although preliminary because of the relatively small sample size, confirm the occurrence of different gut dysbiosis in ANR and ANBP and demonstrate different connections between gut microorganisms and fecal metabolites in the two AN types.


Introduction
It is widely acknowledged that in both animals and humans, the gut microbiota plays a pivotal role in several physiological processes, which may have a potential influence on human behavior and are essential for the maintenance of body homeostasis [1][2][3]. Indeed, increasing evidence indicates that alterations in the composition of gut microbiota are associated with various human diseases [4,5], and bidirectional communication between the gastrointestinal tract and the brain has been documented with a pivotal modulatory role played by the gut microbiota [6]. Several environmental factors affect the composition of gut microbiota and, among them, the type and the composition of the individual's diet seem to be major determinants. Indeed, variations in both short-and long-term dietary in the last two months. None of the participants was using oral contraception or was receiving hormonal therapies; none of them were smokers. Data from six ANBP women, 17 ANR women, and 20 healthy women were available for this secondary analysis.
Participants underwent a clinical assessment including the measurement of eating disorders and general psychopathology using the validated Italian version of the Eating Disorder Examination (EDE) interview [19,20] and the validated Italian version of the Brief Symptom Inventory (BSI), which is a short version of the Symptom Checklist-90 [21,22]. The SCID-5-CV was used to assess current psychiatric comorbidity [23]. Patients collected a fecal sample within one week from their admission into the inpatient unit and before starting any treatment program when they were receiving a standardized diet. Healthy controls (HC) collected fecal samples after one week of a standardized diet. Fresh stool samples were aliquoted in two vials (one vial with RNA later solution and one without solution) and immediately frozen at −80 • C until assayed.
In the week preceding the collection of the fecal sample, AN patients received a standardized diet of 1500 kcal/day with 54.44% carbohydrate, 17.00% protein, and 28.56% lipid, while healthy women received a standardized diet of 2000 kcal/day with 45% carbohydrate, 18% protein, and 35% lipid.
Each participant signed a written informed consent form. The study was carried out in accordance with the ethical principles of the declaration of Helsinki and approved by the University of Salerno's ethics committee (reference number: 31 r.p.s.o., 12 April 2017).

Microbiome Analysis
Microbiome sequencing was performed at the Massachusetts General Hospital NGS Core Facility in Boston (MA, USA). Detailed methods were reported in Monteleone et al. [15]. Briefly, after DNA extraction, V4 regions of the 16S rRNA gene were amplified using a dual barcode system using a fusion primer 515F and 806R. Electrophoretic confirmed amplified were purifies using Quiaquick PCR purification kit (Qiagen, Germantown, MD, USA). After amplicon concentration measurement using Quant-IT Picogreen dsDNA Assay Kit (Thermo Fisher, Waltham, MA, USA), they were sequenced on the Illumina MiSeq v2 using 500 cycles.

Metabolomic Analysis
Stool metabolomic profiling analysis was performed according to Monteleone et al. [15,16]. Briefly, fecal metabolome was extracted and purified using a MetaboPrep GC kit (Theoreo srl-Montecorvino Pugliano, SA, Italy). Profiles were acquired using Gas Chromatograph Mass Spectrometry GCMS-2010 (Shimadzu, Kyoto, Japan). A total of 224 endogenous metabolites involved in energy metabolism, lipid metabolism, and amino acid metabolism were detected sequentially, according to an untargeted metabolomic approach.

Statistical Analysis
Demographic and clinical data were analyzed by using R software [24] (version 4.0.3) using R-study IDE (version 1.3.959). The data distribution was tested using the Shapiro-Wilks test. Since data resulted normally distributed, intergroup differences were tested by one-way ANOVA followed by Tukey's post hoc test. The significance was set at p < 0.05.
Microbiome data analysis was performed according to Monteleone et al. [16]. Briefly, reads were preprocessed using the MICCA pipeline (ver. 1.7.2) [25,26]. Sequence clusterization (97% pairwise identity cutoff) allows assigning the operational taxonomic units (OTUs) that were classified using RDP classifier version 2.12 on 16S data [27]. PyNAST (ver. 1.2.2), based on the Greengenes database [28], was used for sequence alignment. Rarefaction was used to reduce the sampling inhomogeneity. Alpha-(within-sample richness) and beta-(between-sample dissimilarity) diversity were estimates using MicrobiomeAnalystR package (ver. 0.0.0.9000) [29]. Linear discriminant effect size (LEfSe) analysis was also performed to find taxa differentially represented among groups. The supervised machine learning model was based on partial least square discriminant analysis (PLS-DA) according to our previous report [15]. Briefly, analyses were performed using internal standard peak area normalized and chromatogram using MetaboPredict software version 1.2.1 (Theoreo srl). Semi-quantitative data were first aligned, then autoscaled and log-transformed. Models were cross-validated and subjected to a permutation test to verify the lack of overfitting. To manage the class imbalance (six ANBP women vs. 17 ANR women vs. 20 healthy controls) the metacost algorithm [30] was used using the class dimension as a cost matrix. Data dimensionality reduction was performed using genetic algorithm-based selection [31]. Moreover, based on PLS-DA models, variables important in projection (VIP) scores were calculated for each metabolite. Furthermore, one-way ANOVA with Tukey's post hoc test was used to evaluate metabolite concentration differences among patients and healthy participants. Metabolites with a VIP-score higher than 1.5, or with a p-value < 0.05 evaluated using the ANOVA were further confirmed with an independent analytical standard as reported in the Metabolomics Standard Initiative (level metabolics standards initiative (MSI) = 1) [32].
Pathway analysis was conducted using the MetPa algorithm [33]. In the context of pathway analysis, we tested if compounds involved in a particular pathway were enriched compared to random hits using the hypergeometric test. Node centrality (a measure estimation of node importance) was done using the betweenness centrality. It measures the number of shortest paths going through the node. Since the metabolic networks are directed, we used the relative betweenness centrality for a metabolite as an important measure. The betweenness centrality measure focuses on global network topology.
Microbiome and metabolomics data were correlated using heatmaps representation based on Spearman rank correlation coefficients for each pairwise combination of microbial Genus abundances and metabolite semi-quantitative results. Metabolite selection was carried out according to the genetic algorithm as previously reported [15]. Both metabolites and bacteria were clustered.

Clinical Characteristics
As shown in Table 1, patients with ANR did not significantly differ from those with ANBP in body mass index (BMI) and severity of both eating disorder and general psychopathology. ANR women were younger than ANBP ones. Both patient groups had significantly lower BMI and higher psychopathological measures compared to HC.

Metabolomics
As shown in Figure 1, the PLS-DA score plots clearly differentiated ANBP and ANR patients and HC with no overlap. The PLS-DA performance, evaluated by means of cross-validation and permutation test, showed no overfitting: Q 2 = 0.763, R 2 = 0.961, Accuracy = 0.824, p < 0.0001 for the 3-class model, and Q 2 = 0.798, R 2 = 0.912, Accuracy = 0.973, p < 0.0001 for the 2-class model.
compared with the original data, and the sum of squared error is calculated. The better prediction performances are provided by the model, with three components providing for the higher Q 2 (marked with a red star). (C,F) Permutation test results based on 1000 iterations.
The 3-class model separating ANBP and ANR patients and HC identified 13 metabolites with a VIP score >1.5: α-ketoisovaleric acid, tyrosine, deoxycytidine, xylose, tagatose, arabinose, valine, rhamnose, p-hydroxyphenylacetic acid, isoleucine, glycerol, leucine, and n-acetyl glucosamine (Figure 2A). The 2-class model, separating ANBP and ANR patients identified 12 metabolites with a VIP score >1.5 ( Figure 2A): rhamnose, xylose, deoxyadenosine, threonic acid, arabinose, acetic acid, lactose, γ-aminobutyric acid, pyroglutamic acid, succinic acid, sebacic acid, and scyllo-inositol. Based on Principal Component Analysis (PCA) on the ANR and ANBP patients (data not shown), the first 22 principal components explained 95% of the total variance. All relevant metabolites with VIP-score higher than 2.0 and/or significantly different among ANR and ANBP patients and healthy controls using ANOVA are reported in Table 2.  Based on Principal Component Analysis (PCA) on the ANR and ANBP patients (data not shown), the first 22 principal components explained 95% of the total variance. All relevant metabolites with VIP-score higher than 2.0 and/or significantly different among ANR and ANBP patients and healthy controls using ANOVA are reported in Table 2.
The metabolic pathway analyses of the VIP-selected metabolites are summarized in Figure 2B,D. The PLS-DA 3-class model showed that the most relevant involved metabolic pathways were: valine, leucine and isoleucine biosynthesis and degradation, aminoacyl-tRNA biosynthesis, pantothenate biosynthesis, pentose and glucuronate interconversions, phenylalanine, tyrosine and tryptophan biosynthesis, ubiquinone and other terpenoidquinone biosynthesis, and tyrosine, glycerolipid, phenylalanine, galactose, and pyrimidine metabolism. In the PLS-DA 2-class model, there was a clear-cut interaction of several pathways involving pentose and glucuronate interconversions, citrate cycle, glycolysis/gluconeogenesis, and the metabolism of butanoate, alanine aspartate, glutamate, pyruvate, propanoate metabolism, galactose, glutathione, glyoxylate, dicarboxylate, arginine, proline, and purine. ↑ increased or ↓ decreased compared to healthy controls; ↓↓ decreased compared to both healthy controls and the other AN subtype. Differences were statistically significant for those metabolites selected by analysis of variance (ANOVA). PLS-DA = partial least square discriminant analysis; ANR = restricting type anorexia nervosa; ANBP = binge-purging type anorexia nervosa.

Microbiome
With respect to HC, alpha-diversity measured by Chao index ( Figure 3A) showed a significantly reduced value in ANR patients (p = 0.02) and a trend toward a significant reduction in ANBP ones (p = 0.07) while no significant difference emerged between ANR and ANBP groups (p = 0.88). Beta diversity, reported in terms of non-metric multidimensional scaling, showed no significant difference among the analyzed samples resulting in a p = 0.72 based on Permanova evaluation ( Figure 3B).
Analysis of bacteria amounts at phylum levels showed that Actinobacteria were significantly less abundant in ANR compared to ANBP patients (p = 0.009), and Verrucomicrobia resulted higher in ANR women than in healthy controls (p = 0.02). No significant difference in terms of Bacteroidetes to Firmicutes ratio was detected among the groups. Bacteria amounts at the genus level in the three study groups are shown in Figure 4A. LEfSe analysis revealed that ANBP patients had a significant increase in relative abundances of Bifidobacterium, Bifidobacteriaceae, Bifidobacteriales, and Eubacteriacae as well as a significant decrease in relative abundances of Odoribacter, Haemophilus, Pasteurellaceae, and Pasteurellales ( Figure 4B,C).  Analysis of bacteria amounts at phylum levels showed that Actinobacteria were significantly less abundant in ANR compared to ANBP patients (p = 0.009), and Verrucomicrobia resulted higher in ANR women than in healthy controls (p = 0.02). No significant difference in terms of Bacteroidetes to Firmicutes ratio was detected among the groups. Bacteria amounts at the genus level in the three study groups are shown in Figure  4A. LEfSe analysis revealed that ANBP patients had a significant increase in relative abundances of Bifidobacterium, Bifidobacteriaceae, Bifidobacteriales, and Eubacteriacae as well as a significant decrease in relative abundances of Odoribacter, Haemophilus, Pasteurellaceae, and Pasteurellales ( Figure 4B,C).

Microbiome-Metabolomics Integration
Correlation coefficient analysis of the relationships between metabolites selected by VIP-score and ANOVA on the one side and microbial families on the other side showed different frameworks among the groups ( Figure 5). In particular, 43%, 44%, and 47% of the 616 analyzed correlations were positive for healthy subjects, ANBP and ANR patients,

Microbiome-Metabolomics Integration
Correlation coefficient analysis of the relationships between metabolites selected by VIP-score and ANOVA on the one side and microbial families on the other side showed different frameworks among the groups ( Figure 5). In particular, 43%, 44%, and 47% of the 616 analyzed correlations were positive for healthy subjects, ANBP and ANR patients, respectively. Moreover, in healthy controls, 3.2% of the analyzed correlations were statistically significant (p < 0.05); this figure was 3.7% in ANR patients and 5.5% in ANBP patients. Finally, in ANR and ANBP patients, 53% and 49% of the 616 correlations, respectively, showed the same trend (positive or negative) as in healthy controls in terms of rho coefficient, while this figure was 56% in the comparison between ANR and ANBP patients.

Discussion
In the present study, we made a secondary analysis of data from previously characterized fecal microbiome and metabolomics profiles of women with AN and HC [15,16] to verify whether those profiles differed between patients with acute ANR type and those with acute ANBP type. We found that a PLS-DA 3-class model identified 13 fecal metabolites with a VIP score higher than 1.5 that were able to differentiate between patients with ANR, those with ANBP, and HC with no overlap among the three groups. Moreover, a PLS2-DA 2-class model identified 12 fecal metabolites with a VIP score higher than 1.5 that were able to differentiate between patients with ANR from those with ANBP. Furthermore, by adopting two separate metabolite selection strategies with different statistical approaches (PLS-DA and ANOVA), differences in some fecal metabolites emerged among the three groups. In particular, focusing on fecal metabolites showing a statistically significant difference by ANOVA, some substances (i.e., deoxycytidine, isoleucine, malic acid, n-acetylglucosamine, palmitic acid, rhamnose, sorbose, tagatose, and xylose) showed a statistically significant reduction in both patient groups with respect to healthy controls while some others (i.e., allose, glycerol, homogentistic acid, p-hydroxyphenylacetic acid, succinic acid, and threonine) showed statistically significant opposite changes in ANR and ANBP with respects to controls. Finally, compared to ANR patients, ANBP ones showed significantly lower levels of deoxycytidine, isoleucine, malic acid, rhamnose, and xylose but significantly higher levels of n-acetyl-glucosamine, palmitic acid, sorbose, and tagatose. Previous studies that have investigated the metabolomic signature of AN have been focused on targeted metabolites, which are believed possible biomarker of the disease on the basis of specific pathogenetic hypotheses, and have reported significant changes in very long-chain fatty acids [34], steroid metabolome [35], polyunsaturated fatty acids [36,37] and other chemicals [38]. Here we report fecal metabolomics profiles of ANR and ANBP patients as resulting from an untargeted metabolomic approach, which, differently from the targeted methodology, measures as many metabolites as possible. Amid the fecal metabolites that we found statistically different among the groups, the sugars rhamnose, tagatose, sorbose, and xylose resulted uniformly decreased in both patient groups compared to HC with rhamnose and xylose being significantly lower in ANBP than in ANR patients and sorbose and tagatose being significantly higher in ANBP than in ANR patients. These differences may be ascribed to the abnormal feeding of AN patients that relies on energy sources different from those of healthy individuals since they typically avoid the ingestion of sugar-rich food and/or high-fat food [39][40][41]. Moreover, the metabolic pathway analyses of the VIP-selected metabolites identified several metabolic pathways potentially and differently deranged in ANR and ANBP patients. All these findings confirm that ANR and ANBP patients show a profound perturbation in fecal metabolites with respect to healthy women and demonstrate for the first time that the two AN types differ in their fecal metabolomics profiles.
As suggested previously [15], the detected changes in fecal metabolites in patients with AN may result from either their chronic malnutrition and/or changes in their gut microbiota composition. With respect to the first assumption, it has been demonstrated that the diet of women with acute ANR or ANBP differs from that of healthy women in both quantitative and qualitative composition [41]; furthermore, ANBP women not only ingest large amounts of nutrients in their binge-eating episodes but also regurgitate amounts of ingested food [17], which may lead to peculiar nutritional aberrations. These nutritional diversities between the two AN types and HC and between ANR and ANBP individuals may explain part of the observed differences in fecal metabolomics profiles of our ANR and ANBP women. Moreover, present findings also support our second assumption, since profound and distinct changes in the gut microbiome composition and in the relationships between gut bacteria and fecal metabolites were detected between ANR and ANBP women and healthy controls.
In particular, with respect to HC, the intra-individual bacterial richness (alpha-diversity) was reduced in both ANR and ANBP groups, although only in the former reached a statistically significant threshold likely because of the relatively low number of ANBP individuals. The inter-individual difference in gut microbiome composition (beta-diversity) did not statistically differ among the three groups. These data are in line with previous findings from both our group [16] and other research groups [42][43][44] and show for the first time that acute ANR and ANBP patients do not differ in these gut microbiome indices.
At taxonomy levels, the most relevant findings emerged in the comparison of the gut microbiome composition between ANR and ANBP patients. In fact, with respect to ANR patients, ANBP patients exhibited a significant increase in relative abundances of Bifidobacterium, Bifidobacteriaceae, Bifidobacteriales, which represent the genus, family, and order levels in the Actinobacteria phylum, and a significant decrease in the relative abundances of Haemophilus, Pasteurellaceae and Pastuerellales, which represent the genus, family and order levels in the Proteobacteria phylum. Moreover, ANBP patients had a significantly higher abundance of the genus Odoribacter and a lower abundance of the family Eubacteriaceae, which belong to the Bacterioidetes and Firmicutes phyla, respectively. Although these last findings support changes in the relative abundances of those phyla, contrary to our previous results [16], no significant differences emerged in the Bacteroidetesto-Firmicutes ratio between ANR and ANBP group and between each patient group and healthy controls. This discrepancy was likely due to the relatively small sample sizes of our patient groups resulting from this secondary analysis. Anyway, consistent with our findings, Mack et al. [18] reported a different gut microbiome structure between acute ANR and ANBP patients, but they did not provide a comprehensive analysis of that difference. Therefore, our findings add to the current literature on microbiome composition in AN and provide, for the first time, a detailed microbiome differentiation between the two subtypes of AN.
The heatmaps of the relationships of the 28 selected fecal metabolites with bacteria families showed different structures among the three groups. Indeed, by global inspection, it is evident that the heatmap of ANBP patients was drastically different from that of healthy controls, whereas that of ANR patients was more similar to healthy controls, which suggests a profound perturbation of the gut bacteria and fecal metabolites interactions in AN with some relevant differences between ANR and ANBP patients. Indeed, 44% of the interactions between microbial families and fecal metabolites had an opposite direction in the two AN subtype groups, which supports divergent relationships between gut bacteria and metabolite production (positive correlation) or consummation (negative correlation) in patients with ANR or ANBP. Previous studies have explored concomitantly the composition of the gut microbiome and the fecal concentrations of short-chain fatty acids (SCFA), such as propionate, valerate, butyrate, and/or branched-chain fatty acids (BCFA), such as isovalerate and isobutyrate in AN patients and, although results were not consistent, have suggested possible links between alterations in the gut microbiome and changes in the final products of carbohydrates (SCFA) and protein (BCFA) fermentation by gut bacteria [18,44,45]. Our findings expand such knowledge to the interactions of gut bacteria with several different metabolites and demonstrate for the first time that these interactions not only change in AN but they also differ between the two types of AN.
The main limitation of this secondary analysis study is that subgrouping patients according to the AN subtype resulted in two relatively small patient samples, and this could have affected present results. Thus, these findings should be considered preliminary and need confirmation by future studies with larger patient samples. Notwithstanding, this study is the first one that investigated the differences in the gut microbiome and fecal metabolites between ANR and ANBP patients and present findings, if confirmed, may contribute to further differentiate the two types of AN, adding to the current evidence of clinical and biological diversities of binge-purging and restrictive types of eating disorders [46][47][48].

Conclusions
We have found distinct perturbations in fecal metabolomics and microbiome profiles of acute ANR patients compared to acute ANBP patients. Moreover, we have provided evidence that the connections between gut bacteria and some fecal metabolites significantly differ between ANR and ANBP women and with respect to HC. These findings, although preliminary because of the relatively small sample size, confirm the occurrence of different gut dysbiosis in ANR and ANBP type and prove for the first time that different connections between gut microorganisms and fecal metabolites can be traced in the two subtypes of AN. Further studies need to establish whether these changes are an epiphenomenon of the altered eating behaviors or if they may have a role in the pathophysiology of AN representing possible therapeutic targets. Indeed, preliminary findings suggest that intervention aiming at correcting the microbiota modifications of AN patients by fecal microbiota transplantation may improve the fecal metabolome and/or the underweight condition [49,50]. Present data support the idea that such therapeutic interventions should also be targeted to the type of AN.  Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.

Data Availability Statement:
The data presented in this study are available on request from the corresponding author.