Integrated multi-omics of feces, plasma and urine can describe and differentiate pediatric active Crohn’s Disease from remission

Abstract

Background

This study aimed to obtain a holistic view of remission in pediatric Crohn’s Disease (CD) by integrating six omics datasets from three anatomical compartments.

Methods

Patients with fecal calprotectin below 250 mg/kg were considered in remission (n = 27), above 250 mg/kg as having active disease (n = 31). Proteome and microbiomes (fungi and bacteria) were analyzed in feces. Metabolomes in feces, urine, and plasma. Datasets were integrated into a multi-omics model.

Results

The use of individual datasets shows multiple differences between remission and active disease. Integration yielded a good model (AUC of 0.8) predicting remission. The most important features in this model are fecal bacteria (40%), fecal metabolites (22%), fecal proteins (16%), plasma metabolites (12%), fecal fungi (6%), and urine metabolites (4%). The interactome reveals Ruminococcaceae and Faecalibacterium as key players, with a correlation between antifungal urine hydroxyphenyllactic acid and fecal fungi. Pathway analysis shows an association of purine metabolism with remission, independent of thiopurine use. Changes in purine metabolism are confirmed in a pediatric CD public dataset.

Conclusion

The pathways and correlations identified as playing a role in remission may remain undetectable if individual omics datasets or single anatomical compartments are used, highlighting the need for a holistic approach that integrates multiple datasets from multiple anatomical compartments.

Koopman et al. integrated six omics datasets per patient to better understand the reason for remission in pediatric Crohn’s disease. The findings highlight the role of gut bacteria (Ruminococcaceae, Faecalibacterium), metabolic shifts (purine metabolism), and fungal interactions, offering insights into disease mechanisms.

Introduction

Pediatric Crohn’s disease (CD) is a chronic disorder characterized by abdominal pain, bloody diarrhea, and growth failure due to inflammation that can be present in the entire gastrointestinal tract^1,2. CD can have a large impact on quality of life and development of affected children¹. The etiology of CD is complex, and environmental factors such as diet and lifestyle, but also patients’ genetics and the gut microbiome play a role^{3, 4–5}.

The disease often manifests in “flares” of inflammation, also called active disease, in which intensified inflammation and symptom exacerbation can be seen, which requires careful management to restore control of the disease. Flares can be alternated with periods in which the disease is more or less in remission without observable endoscopic inflammation. Determining whether a patient is in remission is a challenge because the different parameters to determine remission do not always correlate. This is especially difficult in a, by definition, vulnerable pediatric population in which monitoring should be limited.

In pediatric patients, endoscopy is not routinely used to confirm inflammation as this poses a burden to the patient. Endoscopy is used in follow-up only if there is a discrepancy between biochemical markers and clinical disease activity, if a patient is therapy-resistant, to evaluate the mucosal condition, or if there is suspicion of another additional disease which might explain the complaints. Instead, fecal calprotectin levels are used as a biomarker for inflammation as it is a direct reflection of the infiltration of immune cells in the mucosa^6,7. Fecal calprotectin is a useful tool in the management of CD in children because it provides information on disease activity, treatment response, and potential relapses⁸. Its non-invasive nature makes it suitable for regular monitoring with a limited burden for the patient, which is crucial for managing CD in a pediatric population. However, there is intra-individual variability in fecal calprotectin values, which makes standardization of the collection procedure important⁹. In addition to this biomarker, the Pediatric Crohn’s Disease Activity Index (PCDAI) or variations on this are used by clinicians as a clinical disease activity measure taking into account patient experiences and physical examination^10,11. However, the PCDAI correlates weakly with endoscopic severity in children with CD^12,13. In rare cases, remission can occur spontaneously but often therapies such as exclusive enteral nutrition, steroids, or anti-TNF agents are needed^{14, 15–16}. Approximately one-third of the patients do not respond or become non-responders during treatment². Together, this highlights the need for the development of new treatment strategies. More insight into the mechanisms connecting the microbiome, immune system, and metabolome determining the shift from remission to active disease is needed.

To improve the mechanistic understanding of CD, studies obtained various omics datasets such as genomics, microbiomics, immune- and metaproteomics, (immune-)transcriptomics, lipidomics, and metabolomics from different anatomical compartments such as feces, blood, and tissue. These approaches have identified key genes, inflammatory cytokines, microbes, and metabolic pathways. For example, in a longitudinal study from our own center in a cohort of pediatric CD, we showed that metabolites such as trimethylamine may be important in the response to nutritional therapy¹⁵. However, up till now, most of these omics studies were unable to offer an integrated and holistic overview of the complex etiology and progression of the disease. To obtain an integrated and holistic view of the molecular consequence of a CD flare-up, integration of omics data from different anatomical compartments of the patient is needed¹⁷.

In the current cross-sectional study, disease status of pediatric CD patients is stratified into active disease or remission based on fecal calprotectin levels. In all patients, 6 different omics datasets from three different anatomical compartments (feces, plasma and urine) are integrated in a machine learning model to obtain a holistic and molecular overview of patients with active CD and patients in remission.

This approach allowed us to better understand which anatomical compartments, which individual features, and which molecular mechanistic pathways are important in active CD. Bacteria, metabolites, and proteins from the fecal compartment were most important in discriminating between active disease and remission. Urine metabolites and fecal fungi are not likely important parameters in the remission of CD.

Ruminococcaceae and Faecalibacterium stand out as key species in the microbiome while purine metabolism emerges as an important differentiator between active disease and remission. Additionally, we reveal possible cross-talk between bacteria and fungi mediated by the bacterial metabolite hydroxyphenyllactic acid.

Methods

Patients and ethics

In this study, we used samples collected from children (9–18 years old) with an established diagnosis of CD according to the revised Porto criteria¹⁸. Children who were treated at the outpatient pediatric gastroenterology clinic of the Amsterdam University Medical Center, Amsterdam, the Netherlands, between May 2014 and April 2017 were included. There was no selection of patients, all children who presented at the clinic and were able to provide samples were included. A partial description was published in the study of Diederen et al.¹⁹. Approval of the local Medical Ethics Committee was obtained (METC AMC 2014-024). In accordance with this approval, the children and their parents both gave informed consent for the use of pediatric samples in this study.

Biochemical and clinical disease activity

Disease activity was determined biochemically using fecal calprotectin levels determined by the Amsterdam University Medical Center clinical chemistry laboratory. Patients were considered to be in remission when they had a calprotectin below 250 mg/kg feces. Clinical disease activity was assessed using the abbreviated Pediatric Crohn’s Disease Activity Index (aPCDAI). Localization and disease behavior were classified using the Paris classification²⁰.

Biospecimen collection, preparation and storage

Urine and plasma samples were collected at the hospital and stored at −20 °C and −80 °C, respectively. Fecal samples were either collected at the hospital or patients were instructed to collect samples in provided containers and store the samples in the freezer as soon as possible, at least within 2 h of collection, and deliver these samples frozen in a cooled bag to the hospital. Aliquots of fecal samples were stored at −20 °C until analysis.

Preparation of fecal water

Fecal samples were suspended in 3 parts MiliQ, vortexed thoroughly and spun down at 13,000 rcf for 4 min. The supernatant was collected and 500 µL was stored at −80 °C for metabolomics. Another 500 µL supernatant was collected for proteomics. To the fecal samples for proteomics, 25× complete™ protease inhibitor cocktail and sodium dodecyl sulphate (1% final concentration) were added. Samples were incubated for 20 min at 95 °C. Then, ammonium bicarbonate (100 mM final concentration), tris(2-carboxyethyl) phosphine (10 mM final concentration) and 2-chloroacetamide (CAA) (30 mM final concentration) was added followed by incubation of 30 min at 60 °C. Samples were stored at −80 °C before using single tube solid phase sample preparation

Metabolomics

Metabolomics of fecal, plasma, and urine samples was performed using Zwitterionic Hydrophilic Interaction Liquid Chromatography and high-resolution mass spectrometry exactly as earlier described²¹. A panel of 200 different metabolite standards was used (Supplementary data).

Single-tube solid-phase sample preparation for proteomic analysis

Protein concentrations were determined by Bicinchoninic acid (BCA) assay (Pierce™ BCA Protein Assay kit) according to the manufacturer’s instructions, and samples were diluted in 1% sodium dodecyl sulfate to obtain 10 µg protein for protein clean up and digestion. The protocol for protein clean up and digestion was adapted from Hughes et al. and Sielaff et al.^22,23. Two types of carboxylate-modified magnetic beads (Sera-Mag™ Magnetic carboxylate modified particles hydrophobic and Sera-Mag™ Magnetic carboxylate modified particles hydrophilic) were mixed 1:1 (v/v), washed, and resuspended in water at a concentration of 20 µg/µl. Twenty-five microliters of sample lysate containing 10 µg of protein was mixed with 2 µl magnetic bead mixture in a 200 µl PCR tube. Immediately, acetonitrile was added to a final percentage of 50% (v/v). Samples were incubated for 18 min at room temperature, placed in a magnetic rack, and incubated for another 2 min. Supernatant was discarded and beads were washed twice with 200 µl 70% ethanol. Then, beads were washed with 180 µl of acetonitrile and air dried at 37 °C. Next, the beads were reconstituted in 10 µl 100 mM ammonium bicarbonate, sonicated for 2–5 min in a water bath sonificator and briefly spun down. Trypsin (1:50 enzyme to substrate ratio) was added, and the samples were incubated overnight at 37 °C. The next day, 2 µl of 10% formic acid was added, the samples were spun down and placed in the magnetic rack. The supernatant was transferred into a new tube and stored at −20 °C until LC-MS/MS.

Proteomics

Peptides were dissolved water containing 0.1% formic acid and 3% acetonitrile and then 200 ng (measured by a NanoDrop at a wavelength of 215 nm) of the peptide was injected by an Ultimate 3000 RSLCnano UHPLC system (Thermo Scientific, Germeringen, Germany). Following injection, the peptides were loaded onto a 75 µm × 250 mm analytical column (C18, 1.6 μm particle size, Aurora, Ionopticks, Australia) kept at 50 °C and flow rate of 400 nl/min at 3% solvent B for 1 min (solvent A: 0.1% FA, solvent B: 0.1% FA in ACN). Subsequently, a stepwise gradient of 2% solvent B at 5 min, followed by 17% solvent B at 24 min, 25% solvent B at 29 min, 34% solvent B at 42 min, 99% solvent B at 33 min held until 40 min returning to initial conditions at 40.1 min equilibrating until 58 min. Eluting peptides were sprayed by the emitter coupled to the column into a captive spray source (Bruker, Bremen Germany) which was coupled to a TIMS-TOF Pro mass spectrometer. The TIMS-TOF was operated in PASEF mode of acquisition for standard proteomics. In PASEF mode, the quad isolation width was 2 Th at 700 m/z and 3 Th at 800 m/z, and the values for collision energy were set from 20 to 59 eV over the TIMS scan range. Precursor ions in an m/z range between 100 and 1700 with a TIMS range of 0.6–1.6 Vs/cm² were selected for fragmentation. 10 PASEF MS/MS scans were triggered with a total cycle time of 1.16 s, with target intensity 2e4 and intensity threshold of 2.5e3, and a charge state range of 0–5. Active exclusion was enabled for 0.4 min, reconsidering precursors if ratio current/previous intensity >4.

Proteomics data processing and bioinformatics

Collected MS data were analyzed using the MaxQuant environment (version 1.6.14.0). The data was mapped against the human proteome retrieved from the Uniprot database (proteome ID = UP000005640) on the 10th of May 2021. Type of group-specific parameters was set as Bruker TIMS-DDA, digestion enzyme was Trypsin with a maximum of two missed cleavages. In MaxQuant, the LFQ method was selected for label-free quantification. Fixed modifications included in protein quantification were carbamidomethylation of cysteine, oxidation of methionine and acetylation of the N-terminal. Other parameters were set by default. Protein quantifications were analyzed in R version 4.1.0²⁴ with the DEP package²⁵. Potential contaminants and identified reverse peptides were removed from the results and samples with >80% missing values were excluded. LFQ-values of proteins that were present in at least 10% of the samples within one group were used as input for the model.

Microbial profiling by 16S and ITS1

Fecal DNA was extracted with PSP® Spin Stool DNA kit (Isogen LifeSciences, Utrecht, The Netherlands), as previously described²⁶. For bacterial profiling a single-step PCR procedure targeting the V3-V4 region of the 16S rRNA gene was performed and the fungal composition was determined by ITS1 amplicon sequence analysis. Both procedures were carried out at the Microbiota Center Amsterdam (MiCA) as described previously²⁷. The obtained library size was 3558 ASVs for 16S rRNA sequencing and 909 OTUs for ITS sequencing after removal of singletons.

16S data processing and analysis

The data was preprocessed to ASVs with the UNOISE3 pipeline and taxonomy was assigned using the DADA2 implementation of the RDP classifier and SILVA 16S reference database V132²⁸. Sequences classified as chloroplasts, mitochondria, contaminants, and unclassified sequences were removed.

Data was analyzed using phyloseq²⁹ and DESeq2³⁰ packages in R software (version 4.1.0)²⁴. Normalization with DESeq2 was performed before beta-diversity analysis. The DESeq analysis was restricted to ASVs that were present with >2 reads in more than 10% of the sample population. Non-normalized counts were used as input for the model.

ITS1 data processing and analysis

The data was preprocessed to OTUs with the DADASNAKE pipeline³¹ and taxonomy was assigned using mothur³² and DECIPHER³³ (& ITS detection—using ITSx & blastn + BASTA) and the unite_8.2_fungi database³⁴. Non-fungal sequences were removed. Data was analyzed using R software (version 4.2.2)²⁴. Normalization was performed before beta-diversity analysis. The analysis was restricted to OTUs that were present with >2 reads in more than 10% of the sample population. Non-normalized counts were used as input for the model.

Machine learning model for data integration

The different data sets (16S, ITS, proteomics, and 3 metabolomics data sets) were integrated in a Manifold Mixing for Stacked Regularization model³⁵ according to our Multi-omics Early-Integration Workflow, in which a combination of data preprocessing, feature selection, data integration, as well as model training techniques, is used³⁶. For each fold in our cross-validation, feature selection was performed only on the training data. The selected features were then applied to the validation set within that fold. We used stratified 5-fold cross-validation repeated 10 times (50 shuffles in total) to ensure robustness. Each fold maintained the class distribution of the original dataset.³⁵

Initially, the data from 16S and ITS were preprocessed by computing relative abundances and performing abundance-based selection of the most abundant 500 bacterial and fungal ASVs³⁷. Next, the data was cleaned, and median imputation for missing values was performed in the proteomics and metabolomics datasets. Only limited imputation was needed in the metabolomics datasets. We performed imputation separately within the high calprotectin group and the low calprotectin group. This approach preserves the class-specific distributions and avoids blending information across classes. A univariate feature selection approach was used, aiming to reduce dimensionality based on biochemical activity labels, which is crucial for effective data integration³⁸. Selected features from the different data sets derived from the same samples were merged. The model was trained by using the ExtraTreesClassifier, with default parameters unless specified, for classification purposes³⁹. Key parameters included the number of trees (n_estimators), maximum depth, and the criteria for splitting. A robust cross-validation process with 50 shuffles was implemented during which the AUC ROC score was calculated and recorded for each iteration. Additionally, a Permutation Importance test was conducted to assess feature importance⁴⁰. The mean area under the curve receiver operating characteristic (AUC ROC) score and feature importance were calculated and reported across all shuffles, ensuring a comprehensive understanding of our model’s performance. For computing correlation among the top features of each omics level Pearson correlation coefficient was used. For computing Python⁴¹ version 3.10 with packages Scipy⁴², Scikit-learn⁴³ was used and Numpy⁴⁴ for model development and visualizations.

Pathway analysis

Metabolites that were significantly different within a single metabolomics dataset, metabolites in the by the model selected top 50 features and metabolites in the by the model selected top 10 for each anatomical compartment were used for pathway analysis in MetaboAnalyst⁴⁵. Homo sapiens (KEGG) was selected as pathway library and all other parameters were set as default.

Statistics and data visualization proteomics, metabolomics, and microbial profiling

For the analysis of the 16S and ITS data, groups were compared using a Welch two-sample t-test for alpha diversity and PERMANOVA for beta diversity. For the bacteria, differential expression was determined with a Wald test within the DESeq2 package³⁰. A false discovery rate of 0.05 was taken into consideration and Benjamini–Hochberg was used as multiple-interference correction. For proteome analysis, differential enrichment was tested using protein-wise linear models and Bayes statistics using limma⁴⁶, and false discovery rates were calculated using fdrtool⁴⁷. T-tests were performed on the metabolomics data. DESeq2³⁰ (v1.40.2) was utilized to conduct differential expression in the published RNA sequencing data. For data visualization tidyverse⁴⁸ and ggplot2⁴⁹ packages were used in R²⁴. Volcano, PCA (metabolomics, proteomics), and PCoA (16S and ITS1) plots show results from non-imputed data.

Anaerobic microbial culture with lactoferrin

To analyze the effect of lactoferrin on bacteria, YCFA medium was modified by supplementing with two concentrations of iron saturated human lactoferrin (1 g/L and 0.01 g/L) (Sigma-Aldrich). The growth of Faecalibacterium prausnitzii (DSM 107838) and Collinsella aerofaciens (DSM 3979) was examined over a 48-h period. The experiments were conducted in a Don Whitley workstation using an absorbance reader (Byonoy) at 600 nm under anaerobic conditions (10% CO₂, 10% H₂, and 80% N₂) at 37 °C. Additionally, growth analysis was performed in the standard YCFA medium for reference. Data analysis was carried out in R²⁴, maximum growth rates were obtained from the absorbance reader data with a window size of 60 min using the growthrate package⁵⁰.

Public single-cell RNA-sequencing data analysis

Processed raw count data in H5AD format was downloaded from the Gut Cell Atlas as published originally by Elmentaite et al.⁵¹ and imported into the R statistical environment (v4.3.1) using the Zellkonverter⁵² (v1.10.1), SingleCellExperiment⁵³ (v1.22.0), and Seurat⁵⁴ (v5.0.1) packages. Differential expression analyses were conducted using the pseudobulk method⁵⁵. In short, counts were summed per gene per sample per provided cell type. DESeq2³⁰ (v1.40.2) was utilized to conduct differential expression, comparing patients with pediatric CD with healthy controls.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Results

Description of clinical cohort

We analyzed the fecal, urine, and plasma samples of 58 pediatric CD patients. The study cohort characteristics can be found in Table 1. Patients with fecal calprotectin levels of <250 mg/kg were considered biochemically in remission and higher levels, up to the detection limit of 1800 mg/kg, were considered to have active disease. In total, 31 patients with active disease and 27 patients in remission were included. The groups were comparable with respect to most characteristics with the exception of a higher average clinical disease score and prednisolone use in the patients with active disease.

Table 1. Patient characteristics (n = 58)

	Calprotectin < 250 mg/kg N = 27	Calprotectin > 250 mg/kg N = 31	Test
Age, years (median, IQR)	15 (13–16)	14 (13–16)	0.353e
Males (n,%)	16 (59%)	19 (61%)	1.000f
Disease duration (months, median, IQR)	41 (17–57)	14 (8–39)	0.084e
Current medication (n,%)	27 (100%)	31 (100%)
• MTXa (n, %)	3 (11%)	4 (13%)	0.494f
• Thiopurinesb (n, %)	10 (37%)	17 (55%)	0.701f
• anti-TNFαc (n, %)	11 (41%)	12 (39%)	0.513f
• Prednisolone	0 (0%)	4 (13%)	0.017f
Previous Crohn’s disease-related surgery (n, %)	3 (11%)	7 (23%)	0.219f
Crohn’s disease: age at diagnosis (Paris classification)			0.804g
• A1a: 0–<10 years (n, %)	8 (30%)	6 (19%)
• A1b: 10–<17 years (n, %)	19 (70%)	25 (81%)
Crohn’s disease: localizationd (Paris classification)	3 (11%)	5 (16%)	0.749g
• L1 (n, %)	9 (33%)	7 (23%)
• L2 (n, %)	14 (52%)	19 (61%)
• L3 (n, %)	10 (37%)	13 (42%)
• L4a (n, %)	1 (4%)	1 (3%)	0.683g
• L4b (n, %)
Crohn’s disease: behavior (Paris classification)
• B1: non-stricturing, non-penetrating (n, %)	22 (82%)	24 (77%)	0.646h
• B2: structuring (n, %)	2 (7%)	5 (16%)
• B3: penetrating (n, %)	2 (7%)	2 (7%)
• B2B3: penetrating and stricturing (n, %)	1 (4%)	0 (0%)
• p: perianal disease (n, %)	6 (21%)	8 (26%)	0.797f
Crohn’s disease: growth impairment (Paris classification)
• Evidence of growth delay (n, %)	4 (15%)	5 (16%)	1.000f
Calprotectin mg/kg (median, IQR)	(92, 30-160)	(1348, 816-1800)	<0.001e
Clinical disease activity (aPCDAI)			0.017h
• Remission (n, %)	22 (82%)	19 (61%)
• Mild (n, %)	5 (19%)	11 (36%)
• Intermediate (n, %)	0	1 (3%)
• Severe (n, %)	0	0

anti-TNFα anti-tumor necrosis factor alpha, IBD inflammatory bowel disease, NA not applicable.

^aMethotrexate.

^bAzathioprine / 6-mercaptopurine / 6-thioguanine.

^cWith or without immunomodulators.

^dL1: distal 1/3 ileum ± limited cecal disease; L2: colonic; L3: ileocolonic; L4a: upper disease proximal to ligament of Treitz; L4b: upper disease distal to ligament of Treitz and proximal to distal 1/3 ileum.

^eMann–Whitney U test.

^fFisher’s exact test.

^gChi-square test (trend: multiple groups).

^hChi-square test (Linear-by-linear association).

Fecal analysis reveals differences in bacterial composition between patients with active disease and in remission, while fungal compositions show variation among all individuals

To investigate differences within the gastrointestinal milieu of this cohort, we analyzed the bacterial and fungal compositions by 16S rRNA gene and ITS sequencing, and the metabolome and proteome by mass spectrometry (Fig. 1).

[See PDF for image]

Fig. 1

Multi-omics in Crohn’s disease patients with active disease and remission.

a Alpha diversity for bacteria (Shannon Diversity index) does not show a significant difference. b Alpha diversity for fungi by the Shannon Diversity index does not show a significant difference. c Relative abundance at family level, left panel patients with active disease, and right panel patients in remission. d Principal coordinates analysis (PCoA) on 16S species level Bray-Curtis dissimilarity shows a significant dissimilarity (P = 0.029). e PCoA on ITS species level Bray-Curtis dissimilarity does not show two distinct groups. f Significant bacterial and fungal ASVs/OTUs with corresponding Log2Fold change, positive values correlate with an increase in active disease while negative values correlate with an increase in remission. Volcano plot (g) fecal metabolomics, h plasma metabolomics, i urine metabolomics and j proteomics: a positive log2FoldChange corresponds with an increase in active disease while negative values correspond with an increase in remission. Non-adjusted p-values are shown, with red colored dots for significant non-adjusted p-values. For all plots n = 27 for the remission group and n = 31 for the active group.

Species saturation curves of 16S rRNA and ITS sequencing can be found in Supplementary Fig. 1.

Assessing the alpha diversity, as measured by the Shannon index, did not reveal significant differences between patients with active disease and those in remission for both their bacteriome and mycobiome (Fig. 1a, b and Supplementary Figs. 2 and 3). However, when examining beta diversity by Bray-Curtis dissimilarity index, we observed a significant different dissimilarity between the two patient groups for bacteria (P = 0.0029) with higher similarity between the patients in remission compared to those with active disease (Fig. 1d). Patients in the remission group showed lower relative levels of Akkermansia and higher relative levels of Bifidobacteriaceae and Coriobacteriaceae (Fig. 1c). Of note, we find that Faecalibacterium prausnitzii is reduced in remission. In total, 160 bacterial ASVs were significantly different (Fig. 1f).

The mycobiomes exhibited considerable variation among all individuals and did not exhibit clustering based on disease state (Fig. 1e). A single Malassezia OTU was shown to be significantly different between active disease and remission (Fig. 1f), confirming a previous study⁵⁶.

Fecal metabolomes show more significant different metabolites than plasma and urine metabolomes

A variety of metabolites displayed significant differences between the two patient groups, with most of the differences observed in the fecal compartment (Fig. 1g). Fecal oxidized glutathione was higher in patients with active disease, confirming previously described increased oxidative stress in CD⁵⁷. While nicotinic acid, also known as vitamin B₃, and pyridoxal, also known as vitamin B₆ were higher in patients in remission, which may reflect changed microbial composition. The fecal compartment also showed significantly increased levels of amino acids cystine, tryptophan, histidine, threonine, phenylalanine, and branched chain amino acids valine, leucine and isoleucine in the active disease group, indicating reduced branched chain amino acid degradation and uptake, which is often observed in IBD^15,58.

In plasma, higher levels of creatine are observed in remission, while fecal creatine and creatinine levels are decreased in remission (Fig. 1g, h). The metabolomes of the urine compartment did not show any significant different metabolites between the two patient groups (Fig. 1i).

Proteomics of the fecal compartment shows differences in proteins related to inflammation and epithelial barrier functioning

The proteome of fecal water was analyzed using a database of human proteins. In the active disease group, immune system-related proteins such as lactoferrin (LTF), alpha-2-macroglobulin (A2M), transferrin (TF) and complement component 4a (C4A) are upregulated. In remission proteins playing role in epithelial barrier functioning, such as cadherin 1 (CDH1) or host-microbe interaction such as annexin A2 (ANXA2), zymogen granule protein 16 (ZG16), and selenium binding protein 1 (SELENBP1) (Fig. 1j) were upregulated. Interestingly, the only protein that was significantly different after stringent false discovery rate correction was lactoferrin, a well-known marker of active CD⁵⁹.

Multi-omics model can accurately predict active disease and remission: bacteriome and fecal metabolome are the most discriminative compartments

To identify which bacteria, fungi, proteins, or metabolites were the most discriminative between the two patient groups, we integrated the data from diverse omics layers. Our model was able to accurately discriminate patients with active disease from patients in remission (AUC ROC 0.80, Fig. 2b). From the top 50 discriminative features, based on their importance scores derived from permutation importance within our machine learning model, 40% belong to the bacteria, 22% to the fecal metabolites, 16% to the proteome, 12% to the plasma metabolites, 6% to the fungi and 4% to the metabolites in urine (Fig. 2a and Supplementary Data). The relative importance scores of the top 300 features can be found in the Supplementary Data.

[See PDF for image]

Fig. 2

Machine learning model for integration of—omics datasets reveals inter- and intra-compartmental correlations.

a Pie-chart showing that from the top 50 features, 40% belong to the bacteria, 22% to the fecal metabolites, 16% to the proteome, 12% to the plasma metabolites, 6% to the fungi, and 4% to the metabolites in urine. b Average receiver operating characteristic (ROC) curve of the machine learning model (AUC = 0.80). Interactome plots depicting (c) inter and intra compartmental correlations between top 10 features of each dataset (d) inter compartmental correlations between top 10 features of each dataset (e) inter and intra compartmental correlations between top 50 features of all data (f) inter compartmental correlations between top 50 features of all data. Correlations are based on Spearman correlation coefficients with the thickness of the lines representing the strength of the correlation and the color the direction of the correlation (blue for positive and red for negative). The node color resembles the dataset and the sizes the importance of the feature for differentiating between active disease and remission.

Multi-omics model shows both inter- and intra-compartmental correlations

To transition from individual features measured by different omics in different anatomical compartments to a more comprehensive understanding of pathways, we analyzed both intra and inter correlations between fecal bacteria, fungi, proteins and urinal, fecal and plasma metabolites by applying Pearson correlation between the top 10 features from each dataset (Fig. 2c, d, Supplementary Fig. 5) and the top 50 of all datasets together (Fig. 2e, f, Supplementary Fig. 6). For clarity we show plots where all interactions are depicted (Fig. 2c, e) and plots in which only the inter compartment interactions are depicted (Fig. 2d, f).

Especially in urine, the intra-compartmental correlations predominate. This may indicate a relative independence of this compartment from the other anatomical compartments sampled.

Increased levels of taurine correlate with FABP1 and bacterial ASVs belonging to Lachnospiraceae and Faecalibacterium prausnitzii

Taurine and fatty acid binding protein 1 (FABP1) (Fig. 2e), an abundant protein that binds bile salts with high affinity⁶⁰ are both increased in feces of patients with active disease. Taurine and FABP correlate in our multi-omics model (Fig. 2e, j) The model also shows positive correlations between taurine and Lachnospiraceae UCG004 and Faecalibacterium prausnitzii CM04-06 (Fig. 2c–f) that are increased in active disease.

Anti-fungal metabolite hydroxyphenyllactic acid correlates with intestinal fungi

Hydroxyphenyllactic acid does not correlate with other urine metabolites in the interactome but strongly and positively correlates with four different fungal OTUs (Fig. 3c, d). Due to common difficulties⁶¹ in exact taxonomic determination of fungi based on ITS sequencing we could not annotate these OTUs. However, all 4 fungal OTUs also strongly correlate with each other, suggesting they represent the same species or occupy the same niche. The relationship is likely explained because hydroxyphenyllactic acid is a compound produced by intestinal bacteria in the Lactobacillus genus that has anti-fungal activity^62,63.

[See PDF for image]

Fig. 3

Overview of alterations in the purine pathway during active Crohn’s Disease as observed in this study.

Inosine is converted to hypoxanthine by purine nucleoside phosphorylase (PNP), hypoxanthine is converted to xanthine, and xanthine is converted to uric acid by xanthine oxidase (XO), which is encoded by the xanthine dehydrogenase (XDH) gene, generating oxidative stress. XDH is upregulated in CD patients compared to healthy controls. We measured increased inosine in plasma and reduced xanthine in fecal water of active disease patients and identified inosine and hypoxanthine in plasma and uric acid as important in discriminating between active disease and remission with a machine learning model.

Lactoferrin negatively correlates with Collinsella species; in vitro experiments show that Collinsella aerofaciens benefits from lactoferrin supplementation

The inflammation-associated protein lactoferrin, depicted as LTF, negatively correlates with an ASV from the Collinsella genus (Fig. 2f). To address the effect of lactoferrin on the growth of Collinsella species, we grew Collinsella aerofaciens and Faecalibacterium prausnitzii, as control species, anaerobically on YCFA medium supplemented with 1 g/L human lactoferrin, resembling concentrations in patients with active inflammation⁶⁴. There was no effect on the growth rate for both species (Supplementary Fig. 7b). However, in contrast to what was expected based on the interactome, we observed a higher yield for Collinsella aerofaciens (P < 0.01) (Supplementary Fig. 7b). Lactoferrin had no effect on the yield of Faecalibacterium prausnitzii.

Pathway analysis reveals altered purine metabolism independently of thiopurine treatment

The metabolites identified as significant by single-compartment analysis, plus those identified as important in the multi-omics model, were used to perform pathway analysis using MetaboAnalyst. This revealed purine metabolism as the most significantly associated pathway (Supplementary Fig. 8).

Purine pathway metabolites, xanthine in feces and inosine in plasma, were significantly different between active disease and remission (Fig. 1g, h). In the multi-omics model, inosine and hypoxanthine in plasma and uric acid in urine were identified as important in distinguishing active disease from remission (Fig. 2e, f).

Since a large percentage of our patients receive azathioprine (a purine analogue) as therapy, we investigated if the observed differences in purine metabolites were caused by azathioprine use. When the patients were stratified according to azathioprine use, there are no differences in fecal and urine metabolomes (Fig. 4a, b) observed between the two groups and for the plasma metabolome a few metabolites are significant different, but these do not belong to the purine pathway (Fig. 4c). These results are not surprising since there is no significant difference in azathioprine use between the patients with active disease and in remission (Table 1). Some significant differences in ASVs are associated with azathioprine use, mainly Firmicutes belonging to the genera Ruminococcaceae, Lachnoclostridium, and Streptococcus (Supplementary Fig. 9)

[See PDF for image]

Fig. 4

Metabolomics in thiopurine users.

Volcano plot (a) fecal, (b) plasma and urine (c) metabolomics, a positive log2FoldChange corresponds with an increase in users while negative values correspond with an increase in non-users. Non-adjusted p-values are shown, with red colored dots for significant non-adjusted p-values. For all plots n = 27 for the remission group and n = 31 for the active group.

Confirmation of alterations in purine metabolism in Crohn’s disease patients by single-cell sequencing analysis

Using a publicly available dataset of intestinal samples from pediatric CD patients and controls⁵¹ we investigated whether pediatric patients with CD have a changed expression pattern of crucial enzymes in purine metabolism (Fig. 5). As a control, we confirmed the increased expression of the two different calprotectin subunits (S100A8, S100A9) in patients with CD compared to healthy controls.

[See PDF for image]

Fig. 5

Single-cell RNA-sequencing data analysis in public data.

A Dotplot visualization of the gene expression where size and color intensity represent the percentage cells with measurable expression and the median expression, respectively. B Arrowplot visualization of the differences in gene expression where the direction and color of the arrow indicates higher (red) or lower (green) expression relative to the control subjects. Full arrows represent genes that were found to be statistically significantly different in expression with the size representing the level of significance.

Xanthine dehydrogenase (XDH), which catalyzes the conversion of hypoxanthine to xanthine and subsequently to uric acid, is increased in patients compared to healthy controls. There was no difference in the expression of purine nucleoside phosphorylase (PNP) or laccase domain containing 1 (LACC1), encoding for the protein FAMIN (Fig. 5B). However, LACC1 is predominantly expressed in cell types that are increased during active disease such as macrophages^65,66 (Fig. 5A). Together, the single cell sequencing data seem to confirm the occurrence of changes in purine metabolism in CD.

Discussion

In this study, we compared CD patients with biochemically determined active disease with CD patients in remission to gain new insights in the metabolites, proteins, and microbes and the corresponding pathways that play a role in the shift from active disease to remission. To the best of our knowledge, we are the first to combine and integrate datasets from the fecal microbiome, proteome, metabolome, and the metabolomes of urine and plasma. We are convinced of the validity and strength of our individual omics datasets because many of the individual features we find different between active CD and remission are in line with those identified in previous studies and have a clear link with inflammation.

For example, we observed more dissimilarity between the bacteriomes of patients with active disease than between patients in remission. An observation that is described as the Anna Karenina principle; healthy microbiomes are more alike than diseased or stressed ones, paralleling Tolstoy’s “all happy families look alike; each unhappy family is unhappy in its own way”⁶⁷.

The finding that Faecalibacterium prausnitzii is reduced in remission is in contrast to many other studies where this species has been attributed a protective role. However, in an independent pediatric cohort from our center, we have observed that reduced Faecalibacterium prausnitzii was also associated with remission induced by nutritional therapy¹⁵.

We found a significant increase in a Malassezia OTU in active disease. Malassezia is a genus of fungi that has previously been described as important in CD, being linked to IBD-associated polymorphism in the gene for CARD9, a signaling adaptor activating pro- and anti-inflammatory cytokines. More specifically, Malassezia restricta was shown to elicit inflammatory responses through CARD9 signaling⁵⁶. Our findings confirm this link with inflammation.

Consistently, our metabolomics and proteomics analyses reveal connections to inflammation. We observed higher levels of creatine in plasma in remission, while fecal creatine and creatinine levels were decreased in remission. This might be the result of damaged epithelium in the gut or reduced uptake by creatine transporters on epithelial cells⁶⁸ and is observed in other metabolomic studies as well⁶⁹. Creatine regulates energy distribution within cells and thereby contributes to the maintenance of the mucosal barrier and protects from inflammation. Therefore, supplementation with creatine is proposed as supporting therapy for CD⁷⁰.

Not surprisingly, proteome analysis showed differences in proteins playing a role in the immune response, epithelial barrier functioning, and interactions with the microbiota and microbial-produced compounds.

Using a machine learning model, we identified the fecal metabolome and bacteriome as the most discriminative between active disease and remission. Because the microbial dataset is the most complex, containing the most individual features, of the 6 different datasets we obtained per patient, it could be argued that this is the reason why it is identified as most important. However, the fecal, plasma, and urine metabolomics datasets are of equal complexity. Yet, the fecal metabolome had a higher importance in the model distinguishing active disease from remission.

The interactome derived from our model shows both intra- and inter-compartmental correlations. Some intra-compartmental correlations are likely explained as they are in the same metabolic pathway, such as the purine metabolites inosine and hypoxanthine in plasma, acetylglutamic acid and glutamate, and creatine and creatinine in feces. Other correlations may be explained by taxonomic relatedness. Strong intra-compartment links are seen in the gut microbes, such as between the different Faecalibacterium, Bacteriodes, and Lachnospiraceae ASVs. All these correlations are positive, which might indicate similar species or strains occupying the same functional niche.

The inter-compartmental correlations within the interactome show the value of combining different datasets from different anatomical compartments. For example, the positive interaction between taurine and Lachnospiraceae UCG004 and Faecalibacterium prausnitzii CM04-06. These species are both increased during active disease and are likely able to deconjugate bile acids, explaining the increased levels of free taurine^71,72. Subsequently, this increase might lead to increased levels of FABP1, a protein that binds bile salts with high affinity, another positive interaction we have observed. Altogether, these correlations may point to altered taurine-conjugated bile salt metabolism in active CD. This further explains the changes in bile salt metabolism related to pediatric CD that have been previously reported by us and others¹⁵.

Another microbe-host interaction we observed by integrating the data was the negative correlation between an ASV annotated as Collinsella and lactoferrin. Lactoferrin-derived peptides have been shown to possess antibacterial properties⁷³. In vitro experiments showed that physiological levels of lactoferrin increased the yield of Collinsella aerofaciens while Faecalibacterium prausnitzii was not affected. Thus, lactoferrin has a positive effect on Collinsella species, not the negative effect expected. It might be that lactoferrin first needs to be degraded into lactoferricin peptides or that our negative correlation might have to do with the utilization of lactoferrin as iron donor⁷³ by Collinsella, which thereby might lower the concentration of lactoferrin. Further research is needed to elucidate this.

We also observed interactions between the mycobiome and urine levels of hydroxyphenyllactic acid, an antifungal compound produced by the Lactobacillaceae family. Inverse correlations between intestinal fungi and Lactobacillaceae have been observed before in a pediatric cohort with an increased risk of childhood atopy and asthma⁷⁴. The positive correlation between hydroxyphenyllactic acid and the fungal OTU is puzzling but may reflect an increased production of hydroxyphenyllactic acid as response to this species or occupation of the fungal niche by these species after the toxic action of hydroxyphenyllactic acid on other fungi. Because these fungal features and hydroxyphenyllactic acid are not among the 50 top important features that define our model separating active CD and remission, it remains uncertain how important this relationship is in CD.

The importance of purine metabolism in CD, especially the role of ATP, has been extensively documented^75,76. There is increasing evidence for the importance of purine metabolism, downstream from ATP, in host immunity. The enzyme FAMIN is a multi-functional purine metabolizing enzyme that is important in autoimmunity by controlling T cell function^77,78. Patients with a biallelic null mutation in FAMIN can have juvenile onset CD⁷⁸. Changes in purine metabolism have been associated with remission in dietary therapy for pediatric CD⁷⁹.

Recent papers also show the importance of a bacterial-host purine axis for immune system function. Several intestinal bacterial species have been shown to produce inosine, important for the modulation of immune therapy in mouse models⁸⁰. A recent paper shows that the ability of intestinal bacteria to anaerobically use purines as a carbon and energy source plays a crucial role in purine homeostasis. This catabolism of purines lowers the uric acid load and is important in atherosclerosis development⁸¹. In addition, purines from the intestinal microbiome have been shown to be important in mouse intestinal wound healing⁸². On top of intestinal bacteria, modulation of purine metabolism by intestinal fungi has been shown to be important in a mouse colitis model⁸³.

Our multi-omics analysis highlights the importance of purine metabolism in CD, it expands results from multiple studies and points to possible new therapeutic targets. An overview of the in this study identified alterations of this pathway can be found in Fig. 3. Importantly, we show that the changes in purine metabolism between patients with active disease and in remission are not a consequence of thiopurine use. Altered purine metabolism may be explained by the changes in expression levels of purine metabolizing enzymes shown in single cell analysis. The increased levels of XDH may also contribute to increased oxidative stress as the peroxisomal conversion of xanthine to uric acid generates hydrogen peroxide. This could help explain the higher amounts of oxidized glutathione we measured in feces.

We show that plasma inosine correlates with several ASVs from the Faecalibacterium genus and Ruminococcaceae family, which may point to a crucial role of this genus and family in purine metabolism. In line with our findings, a connection between Faecalibacterium prausnitzii and inosine regulatory networks has also been observed during IBD-related dysbiosis⁸⁴.

In this study, we only compare patients diagnosed with CD and all of them had relatively mild forms as they were recruited from patients that were presenting at the standard outpatient clinic. There is considerable heterogeneity of the patient population, which may obscure some differences but has the advantage that our conclusions are relatively robust.

Disease activity of the patients in our study was based on fecal calprotectin, an endoscopy would be required to definitively establish disease status. However, in our pediatric cohorts, endoscopies are typically reserved for cases where there is suspected therapy resistance or complications. The designation active disease and remission should be understood with these limitations in mind.

Our study identified pathways and relationships between features and anatomical compartments that are important in CD, which give direction for potential new research in therapeutics, but do not directly lead to a target for intervention.

Conclusion

We identified purine metabolism as important in the differentiation between active disease and remission in CD. Integrating various omics datasets enabled us to identify interactions between features from different anatomical compartments and find pathways that are important in the disease progression of CD. Such pathways might go undetected when analyzing individual omics datasets or single anatomical compartments due to the complexity of the disorder. This underscores the strength of a holistic approach in which multiple anatomical compartments are considered.

Acknowledgements

We would like to thank Hilde Herrema, Mark Davids, and Theo Hakvoort from the Microbiota Center Amsterdam for 16S and ITS sequencing and bioinformatic analysis and Bauke Schomakers and Jill Hermans for their help with the urine metabolomics. Stanley Brul acquired University funding for the project. This research was conducted without external funding.

Author contributions

N.K. and J.S. conceived the project and YJ., P.T.v.L., S.K. provided input to the initial design of the study. K.D., A.K., A.A.t.V. and M.B. were responsible for sample collection and patient inclusion. N.K., Y.J., P.T.v.L. and J.S. performed experimental work. P.T.v.L. performed the ITS analysis. N.K. performed the 16S analysis. Y.J. and S.K. performed the metabolomic analysis. N.K., W.R. and G.K. performed the proteomics analysis. K.C. and E.L. performed the multi-omics integration. A.L.Y. performed the single-cell analysis. N.K. and J.S. wrote the original draft. W.J.d.J. and S.B. critically assessed the project. All authors critically read and gave feedback to the initial version of the manuscript.

Peer review

Peer review information

Communications Medicine thanks the anonymous reviewers for their contribution to the peer review of this work. [Peer review reports are available].

Data availability

Raw sequencing and metabolomics data are available in the EMBL-EBI repositories, all accessible via https://www.ebi.ac.uk/. Metabolomics data can be found under MTBLS9877. 16S rRNA and ITS sequencing data can be found under project number PRJEB74164. Raw proteomics data can be found in the MassIVE repository under PXD062519. Source data for Figs. 1, 2 and 4 can be found in the Supplementary Data 1.

Code availability

The code for the Manifold Mixing for Stacked Regularization model³⁵ is provided at Zenodo⁸⁵.

Abbreviations

aPCDAI

abbreviated Pediatric Crohn’s Disease Activity Index

ASV

Amplicon Sequencing Variant

AUC ROC

Area Under the Receiver Operating Characteristics Curve

CD

Crohn’s Disease

FABP1

Fatty Acid Binding Protein 1

FAMIN

Fatty Acid Metabolism-Immunity Nexus

IBD

Inflammatory Bowel Disease

ITS

Internal transcribed spacer

LACC1

Laccase Domain Containing 1

OTU

Operational Taxonomic Unit

PCA

Principal Component Analysis

PCDAI

Pediatric Crohn’s Disease Activity Index

PCoA

Principle Coordinate Analysis

PNP

Purine Nucleoside Phosphorylase

TIMS-TOF

Trapped Ion Mobility Spectrometry Time Of Flight

UHPLC

Ultra High Performance Liquid Chromatography

XDH

Xanthine Dehydrogenase

YCFA

Yeast Extract Casitone Fatty Acid

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s43856-025-00984-7.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. Mitchel, EB; Rosh, JR. Pediatric management of Crohn’s disease. Gastroenterol. Clin. North Am.; 2022; 51, pp. 401-424. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35595422][DOI: https://dx.doi.org/10.1016/j.gtc.2021.12.013]

2. Conrad, MA; Rosh, JR. Pediatric inflammatory bowel disease. Pediatr. Clin. North Am.; 2017; 64, pp. 577-591. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28502439][DOI: https://dx.doi.org/10.1016/j.pcl.2017.01.005]

3. Fitzgerald R.S., Sanderson I.R., Claesson M.J. Paediatric inflammatory bowel disease and its relationship with the microbiome. Microb. Ecol. https://doi.org/10.1007/s00248-021-01697-9 (2021).

4. Ananthakrishnan, AN. Environmental risk factors for inflammatory bowel diseases: a review. Dig. Dis. Sci.; 2015; 60, pp. 290-298.1:CAS:528:DC%2BC2cXhsFelu7nJ [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25204669][DOI: https://dx.doi.org/10.1007/s10620-014-3350-9]

5. Loddo, I; Romano, C. Inflammatory bowel disease: genetics, epigenetics, and pathogenesis. Front. Immunol.; 2015; 6, 551. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/26579126][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4629465][DOI: https://dx.doi.org/10.3389/fimmu.2015.00551]

6. Diederen, K et al. Raised faecal calprotectin is associated with subsequent symptomatic relapse, in children and adolescents with inflammatory bowel disease in clinical remission. Aliment Pharmacol. Ther.; 2017; 45, pp. 951-960.1:CAS:528:DC%2BC2sXjs1Sqtrs%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28138990][DOI: https://dx.doi.org/10.1111/apt.13950]

7. Johne, B et al. Functional and clinical aspects of the myelomonocyte protein calprotectin. Mol. Pathol.; 1997; 50, pp. 113 LP-113123. [DOI: https://dx.doi.org/10.1136/mp.50.3.113]

8. Weinstein-Nakar I. et al. Associations among mucosal and transmural healing and fecal level of calprotectin in children with Crohn’s disease. https://doi.org/10.1016/j.cgh.2018.01.024 (2018).

9. Koninckx, CR et al. The use of fecal calprotectin testing in paediatric disorders: a position paper of the European Society for Paediatric Gastroenterology and Nutrition Gastroenterology Committee. J. Pediatr. Gastroenterol. Nutr.; 2021; 72, pp. 617-640.1:CAS:528:DC%2BB3MXhs12ktrvE [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33716293][DOI: https://dx.doi.org/10.1097/MPG.0000000000003046]

10. Loonen, HJ; Griffiths, AM; Merkus, MP; Derkx, HHF. A critical assessment of items on the Pediatric Crohn’s Disease Activity Index. J. Pediatr. Gastroenterol. Nutr.; 2003; 36, pp. 90-95.1:STN:280:DC%2BD38jjslaisg%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/12500002]

11. Shepanski, MA; Markowitz, JE; Mamula, P; Hurd, LB; Baldassano, RN. Is an abbreviated pediatric Crohn’s Disease Activity Index better than the original?. J. Pediatr. Gastroenterol. Nutr.; 2004; 39, pp. 68-72. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/15187784]

12. Carman N. et al. Clinical disease activity and endoscopic severity correlate poorly in children newly diagnosed with Crohn’s disease. https://doi.org/10.1016/j.gie.2018.09.025.

13. Hoekman, DR et al. Relationship of clinical symptoms with biomarkers of inflammation in pediatric inflammatory bowel disease. Eur. J. Pediatr.; 2016; 175, pp. 1335-1342.1:CAS:528:DC%2BC28XhsVOrtb7L [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27573259][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5031739][DOI: https://dx.doi.org/10.1007/s00431-016-2762-2]

14. Reznikov E. A., Suskind D. L. Current nutritional therapies in inflammatory bowel disease: improving clinical remission rates and sustainability of long-term dietary therapies. Nutrients15, https://doi.org/10.3390/nu15030668 (2023).

15. Diederen, K et al. Exclusive enteral nutrition mediates gut microbial and metabolic changes that are associated with remission in children with Crohn’s disease. Sci. Rep.; 2020; 10, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33144591][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7609694][DOI: https://dx.doi.org/10.1038/s41598-020-75306-z] 18879.

16. Verburgt, CM; Ghiboub, M; Benninga, MA; de Jonge, WJ; Van Limbergen, JE. Nutritional therapy strategies in pediatric Crohn’s disease. Nutrients; 2021; 13, pp. 1-20. [DOI: https://dx.doi.org/10.3390/nu13010212]

17. Kumar, M; Garand, M; Al Khodor, S. Integrating omics for a better understanding of Inflammatory Bowel Disease: a step towards personalized medicine. J. Transl. Med.; 2019; 17, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31836022][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6909475][DOI: https://dx.doi.org/10.1186/s12967-019-02174-1] 419.

18. Levine, A et al. ESPGHAN revised Porto criteria for the diagnosis of inflammatory bowel disease in children and adolescents. J. Pediatr. Gastroenterol. Nutr.; 2014; 58, pp. 795-806.1:CAS:528:DC%2BC2cXovVWgsbk%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24231644][DOI: https://dx.doi.org/10.1097/MPG.0000000000000239]

19. Diederen, K; Krom, H; Koole, JCD; Benninga, MA; Kindermann, A. Diet and anthropometrics of children with inflammatory bowel disease: a comparison with the general population. Inflamm. Bowel Dis.; 2018; 24, pp. 1632-1640. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29697827][DOI: https://dx.doi.org/10.1093/ibd/izy027]

20. Sherlock, ME et al. Phenotypic evolution of pediatric inflammatory bowel disease using the new Paris classification. Gastroenterology; 2011; 140, S-90. [DOI: https://dx.doi.org/10.1016/S0016-5085(11)60370-3]

21. Schomakers, BV et al. Polar metabolomics in human muscle biopsies using a liquid-liquid extraction and full-scan LC-MS. STAR Protoc.; 2022; 3, 1:CAS:528:DC%2BB38Xhs1SnsrrE [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35479116][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9035783][DOI: https://dx.doi.org/10.1016/j.xpro.2022.101302] 101302.

22. Hughes, CS et al. Ultrasensitive proteome analysis using paramagnetic bead technology. Mol. Syst. Biol.; 2014; 10, 757. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25358341][DOI: https://dx.doi.org/10.15252/msb.20145625]

23. Sielaff, M et al. Evaluation of FASP, SP3, and iST protocols for proteomic sample preparation in the low microgram range. J. Proteome Res.; 2017; 16, pp. 4060-4072.1:CAS:528:DC%2BC2sXhsFCrtb%2FP [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28948796][DOI: https://dx.doi.org/10.1021/acs.jproteome.7b00433]

24. R Core Team. R: A language and environment for statistical computing. https://www.r-project.org/ (2021).

25. Zhang, X et al. Proteome-wide identification of ubiquitin interactions using UbIA-MS. Nat. Protoc.; 2018; 13, pp. 530-550.1:CAS:528:DC%2BC1cXisl2ku74%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29446774][DOI: https://dx.doi.org/10.1038/nprot.2017.147]

26. Rahman S. et al. Dietarycurdlan enhances bifidobacteria and reduces intestinal inflammation in mice. Nutrients13, https://doi.org/10.3390/nu13041305 (2021).

27. Haak B. W. et al. Integrative transkingdom analysis of the gut microbiome in antibiotic perturbation and critical illness. mSystems6https://doi.org/10.1128/mSystems.01148-20 (2021).

28. Quast, C et al. The SILVA ribosomal RNA gene database project: improved data processing and web-based tools. Nucleic Acids Res.; 2013; 41, pp. D590-D596.1:CAS:528:DC%2BC38XhvV2ksb%2FN [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23193283][DOI: https://dx.doi.org/10.1093/nar/gks1219]

29. McMurdie P. J., Holmes S. Phyloseq: an Rpackage for reproducible interactive analysis and graphics of microbiome census data. PLoS ONE8, https://doi.org/10.1371/journal.pone.0061217 (2013).

30. Love, MI; Huber, W; Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol.; 2014; 15, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25516281][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4302049][DOI: https://dx.doi.org/10.1186/s13059-014-0550-8] 550.

31. Weißbecker, C; Schnabel, B; Heintz-Buschart, A. Dadasnake, a Snakemake implementation of DADA2 to process amplicon sequencing data for microbial ecology. Gigascience; 2020; 9, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33252655][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7702218][DOI: https://dx.doi.org/10.1093/gigascience/giaa135] giaa135.

32. Schloss, PD et al. Introducing mothur: open-source, platform-independent, community-supported software for describing and comparing microbial communities. Appl. Environ. Microbiol.; 2009; 75, pp. 7537-7541.1:CAS:528:DC%2BC3cXis1yltw%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/19801464][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2786419][DOI: https://dx.doi.org/10.1128/AEM.01541-09]

33. Firth, HV et al. DECIPHER: database of chromosomal imbalance and phenotype in humans using ensembl resources. Am. J. Hum. Genet.; 2009; 84, pp. 524-533.1:CAS:528:DC%2BD1MXmvVahsLk%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/19344873][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2667985][DOI: https://dx.doi.org/10.1016/j.ajhg.2009.03.010]

34. Kõljalg, U. et al. The taxon hypothesis paradigm—on the unambiguous detection and communication of taxa. Microorganisms8, https://doi.org/10.3390/microorganisms8121910 (2020).

35. Pereira, J., Stroes, E.S.G., Groen A.K., Zwinderman, A.H. & Levin, E. Manifold mixing for stacked regularization BT—Machine learning and knowledge discovery in databases. In: Cellier P., Driessens K., eds. Cham: Springer International Publishing; 444-452. https://doi.org/10.1007/978-3-030-43823-4_36 (2020).

36. Picard, M; Scott-Boyer, M-P; Bodein, A; Périn, O; Droit, A. Integration strategies of multi-omics data for machine learning analysis. Comput. Struct. Biotechnol. J.; 2021; 19, pp. 3735-3746.1:CAS:528:DC%2BB3MXisFegt7%2FJ [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34285775][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8258788][DOI: https://dx.doi.org/10.1016/j.csbj.2021.06.030]

37. Weiss, S et al. Normalization and microbial differential abundance strategies depend upon data characteristics. Microbiome; 2017; 5, pp. 1-18. [DOI: https://dx.doi.org/10.1186/s40168-017-0237-y]

38. Jović, A., Brkić, K., Bogunović, N. A review of feature selection methods with applications. In: 2015 38th International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO) 1200–1205 (IEEE, 2015). https://doi.org/10.1109/MIPRO.2015.7160458.

39. Geurts, P; Ernst, D; Wehenkel, L. Extremely randomized trees. Mach. Learn.; 2006; 63, pp. 3-42. [DOI: https://dx.doi.org/10.1007/s10994-006-6226-1]

40. Altmann, A; Toloşi, L; Sander, O; Lengauer, T. Permutation importance: a corrected feature importance measure. Bioinformatics.; 2010; 26, pp. 1340-1347.1:CAS:528:DC%2BC3cXlvVWqtrk%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/20385727][DOI: https://dx.doi.org/10.1093/bioinformatics/btq134]

41. Van Rossum G., Drake F. L. Python 3 Reference Manual. Scotts Valley, CA: CreateSpace; 2009.

42. Virtanen, P et al. {SciPy} 1.0: Fundamental algorithms for scientific computing in Python. Nat. Methods; 2020; 17, pp. 261-272.1:CAS:528:DC%2BB3cXislCjuro%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32015543][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7056644][DOI: https://dx.doi.org/10.1038/s41592-019-0686-2]

43. Pedregosa, F et al. Scikit-learn: machine learning in {P}ython. J. Mach. Learn. Res.; 2011; 12, pp. 2825-2830.

44. Harris, CR et al. Array programming with {NumPy}. Nature; 2020; 585, pp. 357-362.1:CAS:528:DC%2BB3cXitlWmsbbN [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32939066][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7759461][DOI: https://dx.doi.org/10.1038/s41586-020-2649-2]

45. Pang, Z et al. MetaboAnalyst 5.0: narrowing the gap between raw spectra and functional insights. Nucleic Acids Res.; 2021; 49, pp. W388-W396.1:CAS:528:DC%2BB3MXhvFWgtb3I [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34019663][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8265181][DOI: https://dx.doi.org/10.1093/nar/gkab382]

46. Ritchie, ME et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res.; 2015; 43, e47. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25605792][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4402510][DOI: https://dx.doi.org/10.1093/nar/gkv007]

47. Strimmer, K. A unified approach to false discovery rate estimation. BMC Bioinformatics; 2008; 9, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/18613966][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2475539][DOI: https://dx.doi.org/10.1186/1471-2105-9-303] 303.

48. Wickham, H et al. Welcome to the Tidyverse. J. Open Source Softw.; 2019; 4, 1686. [DOI: https://dx.doi.org/10.21105/joss.01686]

49. Wickham, H. Ggplot2: Elegant Graphics for Data Analysis. Springer-VerlagNew York. Vol. 35 http://link.springer.com/10.1007/978-0-387-98141-3 (2016).

50. Petzoldt, T. Growth rates: estimate growth rates from experimental data. https://cran.r-project.org/web/packages/growthrates/index.html. (2022).

51. Elmentaite, R et al. Single-cell sequencing of developing human gut reveals transcriptional links to childhood Crohn’s disease. Dev. Cell; 2020; 55, pp. 771-783.e5.1:CAS:528:DC%2BB3cXisFaiu7fE [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33290721][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7762816][DOI: https://dx.doi.org/10.1016/j.devcel.2020.11.010]

52. Zappia, L. & Lun, A. Zelkonverter: conversion between scRNA-seq objects. https://doi.org/10.18129/B9.bioc.zellkonverter (2023).

53. Amezquita, RA et al. Orchestrating single-cell analysis with bioconductor. Nat. Methods; 2020; 17, pp. 137-145.1:CAS:528:DC%2BC1MXit1yjsr3L [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31792435][DOI: https://dx.doi.org/10.1038/s41592-019-0654-x]

54. Hao Y. et al. Dictionary learning for integrative, multimodal and scalable single-cell analysis. Nat Biotechnol. https://doi.org/10.1038/s41587-023-01767-y (2023).

55. Zimmerman, KD; Espeland, MA; Langefeld, CD. A practical solution to pseudoreplication bias in single-cell studies. Nat. Commun.; 2021; 12, 1:CAS:528:DC%2BB3MXjtlejtr0%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33531494][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7854630][DOI: https://dx.doi.org/10.1038/s41467-021-21038-1] 738.

56. Limon, JJ et al. Malassezia is associated with Crohn’s disease and exacerbates colitis in mouse models. Cell Host Microbe.; 2019; 25, pp. 377-388.e6.1:CAS:528:DC%2BC1MXkt1aiu7g%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30850233][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6417942][DOI: https://dx.doi.org/10.1016/j.chom.2019.01.007]

57. Alemany-Cosme, E. et al. Oxidative stress in the pathogenesis of Crohn’s disease and the interconnection with immunological response, microbiota, external environmental factors, and epigenetics. Antioxidants10, https://doi.org/10.3390/antiox10010064 (2021).

58. Aldars-García L., Gisbert J. P. & Chaparro M. Metabolomics insights into inflammatory bowel disease: a comprehensive review. Pharmaceuticals14, https://doi.org/10.3390/ph14111190 (2021).

59. Bohra, A. et al. The utility of faecal calprotectin, lactoferrin and other faecal biomarkers in discriminating endoscopic activity in Crohn’s disease: a systematic review and meta-analysis. Biomedicines11https://doi.org/10.3390/biomedicines11051408 (2023).

60. Favretto, F et al. Bile salt recognition by human liver fatty acid binding protein. FEBS J.; 2015; 282, pp. 1271-1288.1:CAS:528:DC%2BC2MXlsFarsbc%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25639618][DOI: https://dx.doi.org/10.1111/febs.13218]

61. Cui, L; Morris, A; Ghedin, E. The human mycobiome in health and disease. Genome Med.; 2013; 5, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23899327][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3978422][DOI: https://dx.doi.org/10.1186/gm467] 63.

62. Siedler, S; Balti, R; Neves, AR. Bioprotective mechanisms of lactic acid bacteria against fungal spoilage of food. Curr. Opin. Biotechnol.; 2019; 56, pp. 138-146.1:CAS:528:DC%2BC1cXitlalur3O [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30504082][DOI: https://dx.doi.org/10.1016/j.copbio.2018.11.015]

63. Lavermicocca, P et al. Purification and characterization of novel antifungal compounds from the sourdough Lactobacillus plantarum strain 21B. Appl. Environ. Microbiol.; 2000; 66, pp. 4084-4090.1:CAS:528:DC%2BD3cXmsVajsbg%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/10966432][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC92262][DOI: https://dx.doi.org/10.1128/AEM.66.9.4084-4090.2000]

64. Kane S.V. et al. Fecal lactoferrin is a sensitive and specific marker in identifying intestinal inflammation. Off. J. Am. Coll. Gastroenterol.98, https://doi.org/10.1111/j.1572-0241.2003.07458.x (2003).

65. Yao, K et al. Increased numbers of macrophages in noninflamed gastroduodenal mucosa of patients with Crohn’s disease. Dig. Dis. Sci.; 1996; 41, pp. 2260-2267.1:STN:280:DyaK2s%2FpvFCjtw%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/8943982][DOI: https://dx.doi.org/10.1007/BF02071410]

66. Mahida, YR; Patel, S; Gionchetti, P; Vaux, D; Jewell, DP. Macrophage subpopulations in lamina propria of normal and inflamed colon and terminal ileum. Gut; 1989; 30, pp. 826 LP-826834. [DOI: https://dx.doi.org/10.1136/gut.30.6.826]

67. Ma, ZS. Testing the Anna Karenina principle in human microbiome-associated diseases. iScience; 2020; 23, 1:CAS:528:DC%2BB3cXnsVWhurw%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32305861][DOI: https://dx.doi.org/10.1016/j.isci.2020.101007] 101007.

68. Hall, CHT et al. Creatine transporter, reduced in colon tissues from patients with inflammatory bowel diseases, regulates energy balance in intestinal epithelial cells, epithelial integrity, and barrier function. Gastroenterology; 2020; 159, pp. 984-998.e1.1:CAS:528:DC%2BB3cXhvVKns7rO [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32433978][DOI: https://dx.doi.org/10.1053/j.gastro.2020.05.033]

69. Notararigo, S et al. Targeted 1H NMR metabolomics and immunological phenotyping of human fresh blood and serum samples discriminate between healthy individuals and inflammatory bowel disease patients treated with anti-TNF. J. Mol. Med.; 2021; 99, pp. 1251-1264.1:CAS:528:DC%2BB3MXhtlOmt7zP [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34021361][DOI: https://dx.doi.org/10.1007/s00109-021-02094-y]

70. Wallimann, T., Hall, C. H. T., Colgan, S. P. & Glover, L. E. Creatine supplementation for patients with inflammatory bowel diseases: a scientific rationale for a clinical trial. Nutrients13, https://doi.org/10.3390/nu13051429 (2021).

71. Vital, M; Rud, T; Rath, S; Pieper, DH; Schlüter, D. Diversity of bacteria exhibiting bile acid-inducible 7α-dehydroxylation genes in the human gut. Comput. Struct. Biotechnol. J.; 2019; 17, pp. 1016-1019.1:CAS:528:DC%2BC1MXhsFersb3O [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31428294][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6692061][DOI: https://dx.doi.org/10.1016/j.csbj.2019.07.012]

72. Song, Z et al. Taxonomic profiling and populational patterns of bacterial bile salt hydrolase (BSH) genes based on worldwide human gut microbiome. Microbiome; 2019; 7, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30674356][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6345003][DOI: https://dx.doi.org/10.1186/s40168-019-0628-3] 9.

73. Vega-Bautista A. et al. The impact of lactoferrin on the growth of intestinal inhabitant bacteria. Int. J. Mol. Sci. 20, https://doi.org/10.3390/ijms20194707 (2019).

74. Fujimura, KE et al. Neonatal gut microbiota associates with childhood multisensitized atopy and T cell differentiation. Nat. Med.; 2016; 22, pp. 1187-1191.1:CAS:528:DC%2BC28XhsV2rtb%2FL [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27618652][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5053876][DOI: https://dx.doi.org/10.1038/nm.4176]

75. Idzko, M; Ferrari, D; Eltzschig, HK. Nucleotide signalling during inflammation. Nature; 2014; 509, pp. 310-317.1:CAS:528:DC%2BC2cXotVyqtLc%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24828189][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4222675][DOI: https://dx.doi.org/10.1038/nature13085]

76. Eltzschig, HK; Sitkovsky, MV; Robson, SC. Purinergic signaling during inflammation. N. Engl. J. Med.; 2012; 367, pp. 2322-2333.1:CAS:528:DC%2BC38XhvVKmt77L [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23234515][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3675791][DOI: https://dx.doi.org/10.1056/NEJMra1205750]

77. Assadi, G et al. Functional analyses of the Crohn’s disease risk gene LACC1. PLoS ONE; 2016; 11, e0168276. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27959965][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5154582][DOI: https://dx.doi.org/10.1371/journal.pone.0168276]

78. Cader, MZ et al. FAMIN is a multifunctional purine enzyme enabling the purine nucleotide cycle. Cell; 2020; 180, pp. 278-295.e23.1:CAS:528:DC%2BB3cXnvFeltLo%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31978345][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6978800][DOI: https://dx.doi.org/10.1016/j.cell.2019.12.017]

79. Ghiboub, M et al. Metabolome changes with diet-induced remission in pediatric Crohn’s disease. Gastroenterology; 2022; 163, pp. 922-936.e15.1:CAS:528:DC%2BB38XisVGmur3I [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35679949][DOI: https://dx.doi.org/10.1053/j.gastro.2022.05.050]

80. Mager, LF et al. Microbiome-derived inosine modulates response to checkpoint inhibitor immunotherapy. Science; 2020; 369, pp. 1481-1489.1:CAS:528:DC%2BB3cXhvVOns7zO [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32792462][DOI: https://dx.doi.org/10.1126/science.abc3421]

81. Kasahara, K et al. Gut bacterial metabolism contributes to host global purine homeostasis. Cell Host Microbe; 2023; 31, pp. 1038-1053.e10.1:CAS:528:DC%2BB3sXht1WnsLfJ [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37279756][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10311284][DOI: https://dx.doi.org/10.1016/j.chom.2023.05.011]

82. Lee, JS et al. Microbiota-sourced purines support wound healing and mucous barrier function. iScience; 2020; 23, 1:CAS:528:DC%2BB3cXhtlWqt7bL [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32554188][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7303675][DOI: https://dx.doi.org/10.1016/j.isci.2020.101226] 101226.

83. Chiaro, TR et al. A member of the gut mycobiota modulates host purine metabolism exacerbating colitis in mice. Sci. Transl. Med.; 2017; 9, [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28275154][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5994919][DOI: https://dx.doi.org/10.1126/scitranslmed.aaf9044] eaaf9044.

84. Lloyd-Price, J et al. Multi-omics of the gut microbial ecosystem in inflammatory bowel diseases. Nature; 2019; 569, pp. 655-662.1:CAS:528:DC%2BC1MXhtVOgtL3O [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31142855][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6650278][DOI: https://dx.doi.org/10.1038/s41586-019-1237-9]

85. Horaizon, B. V. Multi-omics analysis via manifold mixing. https://doi.org/10.5281/zenodo.15680275 (2025).