Introduction

In captive common marmosets, gastrointestinal (GI) diseases are the most common and widespread clinical finding1,2. Inflammatory bowel disease (IBD) prevalence is reported to be as high as 28–60% in captive marmosets and presents with diarrhea, weight loss, enteritis, muscle atrophy, alopecia, hypoproteinemia, anemia, elevated liver enzymes, failure to thrive and mortality1,3. In addition to IBD, a novel GI disease has been described in young adult to adult marmosets characterized by duodenal dilation or stricture near the major duodenal papilla4,5,6. Clinical signs, such as diarrhea, weight loss, or poor weight gain, resemble IBD but increased vomiting is also observed. This syndrome was associated with hypoalbuminemia, hypoglobulinemia, hypoproteinemia, hypocalcemia (total), elevated alkaline phosphatase, anemia, and in some cases, leukocytosis5. Histologically, duodenal mucosal ulcerations with associated chronic-active granulocytic and lympho-histiocytic inflammation were observed. Thus far only two institutions have reported this disease characterized by duodenal dilation amongst captive marmosets6. This duodenal syndrome was found in 21.9% of necropsy cases in a Japanese institution4. Within our institution, we observed a 13% prevalence of stricture cases during this 2 year study. However, 91.3% of cases (21 of 23 cases) within our colony were observed in marmosets sourced from the former New England Primate Resource Center, yielding a 26% prevalence when only considering marmosets from this source (MITNE). In this study, we investigated the changes observed in the gut microbiome of MITNE marmosets that could be associated with the development of duodenal strictures.

The human GI tract harbors trillions of microorganisms from at least 400 species that compose the intestinal microbiota7,8. In healthy individuals, the microbiome influences many physiological functions such as extracting nutrients, maintaining the gut mucosal barrier, training immune cells and protecting against pathogens9. Dysbiosis occurs due to loss of beneficial microbes, expansion of pathobionts (opportunistic microbes), or reduction of microbial diversity. Dysbiosis has been associated with human diseases, including IBD, irritable bowel syndrome, obesity, psoriasis, rheumatoid arthritis, autism spectrum disorders, and Clostridioles difficile infection9,10. As the microbiome has been associated with human GI diseases, factors affecting the microbiome in non-human primates (NHP) are being explored, such as species, social structure, environment and diet11,12,13,14. Captivity and diets fed to captive marmosets have been associated with microbial diversity loss, shifts in the Firmicutes:Bacteroidetes ratio, and increased GI disease and mortality11,12,14. Dietary specialists, such as marmosets, are more susceptible to captivity-associated dietary changes14. Marmosets are exudivores that consume large amounts of indigestible oligosaccharides from tree gums15 and may harbor specific gut microbes dedicated to carbohydrate metabolism. Currently, few reports on the marmoset microbiome are available16,17,18,19,20,21. We recently published a longitudinal marmoset microbiome study examining both healthy marmosets and marmosets presenting clinically with IBD which surveyed 503 samples from 140 marmosets from four different sources22. Interestingly, even after months or years of husbandry at MIT, the most notable differences in the microbiome were based on the animals’ source, while marmoset age or sex had little to no effect in the microbiome22.

In this study, we compared the microbiome of 21 marmosets diagnosed with duodenal ulcer/strictures with the microbiome of 52 healthy, non-stricture marmosets imported from the same source [New England Primate Research Center (NEPRC)] and maintained at the MIT colony (MITNE). “Stricture” animals were defined as marmosets that had or developed strictures based on clinical and histological assessments while “non-stricture” individuals were not clinically diagnosed with strictures and were not receiving chronic drug treatments during the study period. Healthy, non-stricture MITNE marmosets were included in our prior publication comparing the microbiome of MIT marmosets imported from different sources22. The current study evaluated fecal and rectal swab samples collected over a 2 year period during physical examinations or necropsies. Serum chemistry and complete blood count (CBC) samples from both healthy marmosets and marmosets diagnosed with duodenal ulcer/strictures from the MITNE colony were routinely collected once a year during physical examinations or during necropsies within the same 2 year period23. We identified changes in both microbial communities and blood parameters that may serve as marmoset biomarkers for strictures and propose that marmosets may be useful animal models to study Clostridium-driven GI disorders, such as duodenal strictures.

Results

Effects of Duodenal strictures on microbiome of MITNE marmosets

Of the 23 stricture cases identified at MIT during the study period, 21 marmosets belonged to the MITNE cohort, which had an estimated prevalence of 26% for this disease. This study focused on comparing biomarkers over a 2 year period collected from MITNE animals that were clinically or histologically diagnosed with strictures against healthy, non-stricture marmosets from the same source22. Marmosets included in this study were housed in a single building, kept isolated from marmosets from other sources for the duration of the 2 year study, and provided standardized diet, husbandry and veterinary care. In our analysis, the gut microbiome of healthy MITNE marmosets was characterized by high abundance of Bacteroides, followed by Prevotella 9 and Anaerobiospirillum. We then compared the samples from the 52 non-stricture MITNE marmosets with samples from 21 MITNE marmosets diagnosed with strictures (Fig. 1, Table 1). The analysis identified 601 ASVs that could be collapsed into 128 genera with 47 genera accounting for 99% of total reads. Stricture marmosets had markedly different microbiomes compared to non-stricture animals. On average, a 32% decrease in Bacteroides was observed in stricture cases (35.8 ± 1.2% (mean ± standard error) abundance in non-stricture vs. 24.5 ± 2.0% in stricture). This decrease in Bacteroides, decreased the Bacteroides:Prevotella 9 ratio from 3.1 in non-stricture to 1.4 in stricture. Anaerobiospirillum, the second most abundant genus in non-stricture marmosets (13.2 ± 0.7%), decreased to 4.6 ± 1.2% in stricture cases. The decreased abundance in these two genera was compensated by a 50% increase in Megamonas that was observed in stricture cases (6.5 ± 0.4% in non-stricture vs. 9.4 ± 1.0% in stricture) (Fig. 1). Analysis of Composition of Microbiomes (ANCOM), a log-ratio based statistical method that accounts for the compositional nature of microbiome data in differential abundance analysis between groups, was used to compare the two cohorts at the genus level. At this taxonomic level, ANCOM highlighted the decrease in Anaerobiospirillum observed in stricture cases and identified Clostridium sensu stricto 1 as differentially expressed due to an increase in relative abundance observed in marmosets with stricture. Despite changes in microbial composition, no changes in alpha diversity were observed using multiple metrics (Chao1, Pielou’s Evenness, Observed OTUs and Shannon).

Figure 1
figure 1

(A) Microbiome composition of individual samples at the genus level and (B) pie charts with average bacterial abundances of stricture (n = 57) and non-stricture marmosets (n = 153) show dysbiosis associated with stricture characterized by decreased Bacteroides and Anaerobiospirillum and increased Megamonas.

Full size image
Table 1 Description of microbiome sample demographics of MITNE.
Full size table

Using Principal Coordinate Analysis (PCoA) of Weighted UniFrac distances, 54.86% of variance was accounted for by 3 axes, with separation between stricture and healthy, non-stricture animals observed along the 2nd axis (Fig. 2). Using PERMANOVA (PERmutational Multivariate Analysis Of VAriance), a significant difference between the centroids of the healthy sample cluster and the stricture sample cluster was observed (PERMANOVA, P < 0.001 with F-statistic of 15.96), but beta-dispersion was also significantly different between clusters, implying that the spread of each cluster was different and could account for the difference observed between the centroids (beta-dispersion, P < 0.001) (Fig. 2).

Figure 2
figure 2

Weighted UniFrac PCoA plot depicting the top three principal components (PC) with (A) PC1 vs. PC2 and (B) PC2 vs. PC3. Clustering of microbiome profiles based on stricture status is observed with PC2 defining separation of healthy marmosets and stricture cases. However, beta-dispersion of both healthy and stricture cohorts is significantly different, implying differences in the variance observed in the two cohorts.

Full size image

Identification of microbial and serum biomarkers of duodenal strictures using machine learning

To identify potential microbial biomarkers associated with duodenal strictures, the microbiome data was analyzed using four machine learning algorithms (random forest (RF), support vector machines (SVM), classification and regression trees (CART) and k-nearest neighbor (KNN)) to determine the most accurate models for stricture classification based on marmoset microbiome profiles. For each algorithm, a subset of the dataset was used to train the model (training set) by providing the microbiome profile and the classification of the sample as a stricture or non-stricture sample. The remaining data was reserved to test the model’s accuracy (testing set). After the model was generated using the training set, the microbiome data from the testing set was provided to the model without the sample’s classifications. The model generated predictions based on the testing set microbiome profiles, which were then compared to the actual sample classifications in the testing set to determine the model’s accuracy. The model generation process was carried out iteratively to sample multiple subsets of the data and determine the robustness of the algorithm. Two metrics, accuracy and kappa, are shown for each model in Fig. 3a. Accuracy measures the percentage of correctly classified instances by comparing the clinically diagnosed ground truth data of the testing set with the model’s predictions. While kappa also compares the agreement of samples classified by the machine learning model with the ground truth data, it differs from accuracy by accounting the hypothetical probability of random agreements. Kappa values greater than 0.40 reflect moderate or substantial agreement between the model and the ground truth. Comparison of the four classification models shows that RF analysis provided the highest accuracy and kappa values when classifying microbiome profiles into “stricture” or “non-stricture.” Focusing on the RF model, we then evaluated the stability of three metrics (accuracy, kappa and F1 scores) to determine the least number of ASVs that maximized the three metrics. F1 score evaluates the model’s utility using both precision and recall (or sensitivity). The original analysis using QIIME2 generated a list of 601 ASVs. However, the RF model was able to detect the most important ASVs for classifying samples as “stricture” or “non-stricture.” Using a minimum of 4 ASVs, the three metrics in the RF model begin to stabilize, but we selected a 9 ASV model that presented the highest levels of accuracy, F1 and kappa observed with this model (Fig. 3b). The receiver operating characteristic (ROC) curve and area under the curve (AUC) value were calculated for the RF model using the 9 ASVs (Fig. 3c). The ROC curve had an AUC value of 0.82 with an accuracy of 85%, a sensitivity of 100% and a specificity of 45%, demonstrating a strong performance in classifying strictures and non-strictures. To better understand the effects of the 9 ASVs on the system, we identified the bacteria associated by QIIME2 with the ASVs and plotted the relative abundance of each based on their stricture status and determined that 8 of 9 ASVs were indeed significantly different by ANCOM analysis at the ASV level (Fig. 3d). Of these 9 ASVs, 3 Anaerobiospirillum ASVs, as well as Bacteroides and Parabacteroides ASVs, decreased in stricture cases. Increases were observed in ASVs from Bifidobacterium, Clostridium sensu stricto 1, Oribacterium, and Megamonas.

Figure 3
figure 3

(A) Evaluation of 4 machine learning algorithms for classification of marmoset microbiomes based on stricture status. (B) Relationship between random forest model’s accuracy, F1 and Kappa on the number of ASVs used by the model to classify stricture and non-stricture samples with an 85% accuracy. (C) Area under the curve (AUC) of receiver operating characteristic (ROC) curves for random forest models using microbiome shows strong performance of models in classifying strictures and non-strictures. (D) Comparison of the relative abundance of the 9 ASVs identified in the random forest model in stricture and non-stricture samples. ANCOM analysis at the ASV level determined 8 of 9 ASVs to be significantly different between stricture and non-stricture. *= (W > 571) E) Evaluation of relative abundance of Clostridium sensu stricto 1 reads in duodenal biopsies demonstrates increased abundance in stricture cases compared to non-stricture cases.

Full size image

Next, we developed RF models using serum chemistry or CBC data to determine if “stricture” and “non-stricture” could be identified using blood analysis (Supp. Table 1). First, we evaluated the serum chemistry parameters needed to optimize accuracy, F1 and kappa, and determined that 4 serum chemistry parameters (total protein, lipase, gamma-glutamyl transferase (GGT) and amylase) classified “stricture” and “non-stricture” with 84.8% accuracy, a sensitivity of 76.5%, a specificity of 93.8% and AUC of 0.89 (Fig. 4a,b,g). Total protein and GGT decreased in stricture cases, while pancreatic markers, lipase and amylase, were increased in stricture animals (Fig. 4c). Using CBC data, the RF classifier identified 6 parameters (hematocrit (HCT), hemoglobin (HGB), red blood cell count (RBC), red cell distribution width (RDW), mean corpuscular hemoglobin (MCH) and lymphocyte percentage) that classified strictures with an accuracy of 82.8%, a sensitivity of 89.4%, a specificity of 75% and AUC of 0.83 (Fig. 4d–f). All variables, except RDW, decreased in strictures (Fig. 4g).

Figure 4
figure 4

(A) Four serum chemistry parameters required for the optimal combination of accuracy and variable number in serum chemistry RF model. (B) Variables of importance for RF model using serum chemistry data to classify strictures. (C) Four most important parameters include Total Protein, GGT, amylase and lipase. Amylase and lipase are elevated in stricture cases suggesting pancreatic disease. (D) Six CBC parameters required for optimal combination of accuracy and variable number in CBC RF model. (E) Variables of importance for RF model using complete blood counts (CBC) to classify strictures. (F) 6 CBC parameters include HCT, HGB, RBC, RDW, MCH and lymphocyte %. 5 of 6 parameters indicate anemia. (G) Area under the curve (AUC) of receiver operating characteristic (ROC) curves for random forest models using serum chemistry (blue) or complete blood (yellow) count show strong performance in classifying strictures and non-strictures.

Full size image

Identification of Clostridium species based on sequencing reads

As ANCOM and the RF model highlighted the role of Clostridium sensu stricto 1, we further investigated the species that are encompassed by the Clostridium sensu stricto 1 taxonomy, which included the following Clostridium species: C. tetani, C. botulinum, C. kluyveri, C. acetobutylicum, C. novyi, C. perfringens and C. beijerinckii. These species are generally considered pathogenic and indicative of less healthy and less diverse microbiota24. Using the representative sequences assigned to all Clostridium sensu stricto 1 ASVs, we determined that 232,156 (69%) Clostridium sensu stricto 1 reads shared > 99% identity over the 370 bp sequence with C. perfringens. Remaining reads matched with C. baratii (19%), C. colicanis (7%) and an unknown Clostridium species (6%). Importantly, ASV256, which increased sixfold in stricture samples, shared 100% identity with C. perfringens. We then sought to confirm the presence of C. perfringens by microbial culture followed by 16S rRNA Sanger sequencing of clinical isolates. Clostridium perfringens was isolated and confirmed by sequencing in 4 of 9 duodenum samples tested from marmosets with histologically confirmed strictures. The only other Clostridium isolate recovered from the 9 samples was identified as either C. baratii or C. sardiniense, a rare causative agent of botulism25.

Increased clostridium sensu stricto 1 abundance in the duodenum of stricture cases

As the microbiome analysis of the lower GI identified an increase of the C. perfringens ASV and we isolated the putative pathogen from duodenal tissue collected from the stricture site, we analyzed the microbiome using duodenal samples from stricture (n = 17) and non-stricture cases (n = 12). Clostridium sensu stricto 1 was observed at greater than 1% abundance in 76% of strictures (13/17) but only in 16% of non-stricture cases (2/12). In 8 stricture cases, Clostridium was the most abundant genus with abundances ranging from 37 to 87%. Interestingly, one non-stricture sample with 30% abundance of Clostridium sensu stricto 1 had duodenal pathology characterized by mild duodenal mucosal congestion (Fig. 3e).

Effects of GI disease on gene expression of the small intestine

We tested whether strictures significantly altered marmoset transcriptomic profiles using RNA sequencing (RNAseq) on samples from non-stricture (n = 3) or stricture (n = 3) marmosets. Marmosets with strictures presented with gross thickening, duodenal stricture or ulceration (0.5–1 cm aboral to the major duodenal papilla). Duodenal tissue evaluated was immediately distal to the lesion (“stricture”) or in an equivalent anatomic region in IBD animals (“non-stricture”). However, IBD animals that served as non-stricture controls presented with thickened intestines that were grossly observed, and a diagnosis of duodenitis was noted. Comparing stricture and non-stricture duodenums, we identified 1,183 differentially expressed genes (DEG) (FDR < 0.05) (Fig. 5a, Supp. Table 2). To perform Gene ontology (GO) analysis, marmoset genes with official names were matched to Homo sapiens genes to retrieve Entrez IDs with associated GO categories. Analysis of this gene subset identified 903 DEGs with GO annotations. The top 15 biological processes (BP) with significant enrichment are listed in Table 2 (complete list Supp. Table 3). Stricture samples enriched BP sets involved with intestinal absorption, and lipid metabolism, localization, and transport (Fig. 5b, Supp. Fig. 1a). Stricture upregulated genes encompassed cholesterol-associated genes including apoliproteins (APOB, APOA1 and APOA4), transport genes (ABCG5, ABCG8, GRAMD1B, and STARD3), metabolic genes (DGAT1, CYP11A1, and CYP27A1) and binding/absorption genes (SOAT2, NPC1L1 and SCARB1) (Supp. Table 2). Other lipid-associated genes upregulated by stricture included genes associated with fatty acid binding proteins (FABP1 and FABP2), peroxisomes (PPARA, ABCD1, ACAA1 and EPHX2), ketogenesis (HMGCS2) and lipid synthesis (GPAM, SREBF1, SCAP, and ACACB). Enriched cellular membrane GO sets shared these lipid-associated genes due to functional overlap (Supp. Table 3, Supp. Fig.1). Interestingly, immunity-associated genes were more highly expressed in non-stricture duodenums (Fig. 5b, Supp. Fig. 2), possibly due to enteritis observed in IBD marmosets. These genes included antimicrobial responses (LCN2, LYZ, MUC20), toll-like receptors (TLR2 and TLR4), superoxide-generating NADPH oxidase activity (NOX1 and DUOX2), killer cell lectin-like receptor genes (KLRB1, KLRC1, KLRD1, and KLRF1), and chemokine activity and receptor binding (CXCL1, CXCL10, TFF2 and PF4) (Supp. Table 2b). The transcriptional profile implies the activity of natural killer (NK) cells, neutrophils and MHC class I protein complex binding.

Figure 5
figure 5

(A) Differentially expressed genes (DEG) (FDR < 0.05) in the duodenum of non-stricture and stricture cases. (B) Gene ontology (GO) sets enriched in stricture cases show upregulation of lipid metabolism, transport and localization. Non-stricture cases have enrichment of immune processes, possibly due to underlying pathology in necropsied animals with non-stricture diseases.

Full size image
Table 2 Top gene ontology sets in stricture.
Full size table

Discussion

GI diseases are the most prevalent clinical disease in captive common marmosets1,2,26, but the role of the microbiome is largely unknown. Recent literature demonstrates that housing in captive environments affects NHP microbiome composition, reduces alpha diversity, and alters host responses to disease11,14,27. In captivity, NHP microbiomes lose distinctive, wild microbiota and become dominated by Prevotella and Bacteroides, the most abundant genera in the modern human gut microbiome8,11,28. In the largest marmoset microbiome study to date, our previous report supported the hypothesis that captivity humanizes the primate microbiome, as Bacteroides and Prevotella 9 were the most abundant genera with levels similar to those observed in human feces8,22,28. Within the MIT colony, NEPRC marmosets had the highest relative abundance of Bacteroides compared to the other marmoset sources22. MITNE marmosets had the highest Bacteroidaceae abundance (37%) and the lowest Prevotellaceae levels (17%), and were most susceptible to strictures, a novel GI disease in marmosets4,5. This duodenal syndrome was found in 21.9% of necropsy cases in a Japanese institution4, while MITNE marmosets had a 26% prevalence. Clinical signs include vomiting, bloating, weight loss and palpable thickening of the duodenum that can be visualized through radiography and ultrasound4,5. As strictures were most prevalent in the NEPRC-sourced colony, we compared microbiome samples from 21 NEPRC-sourced marmosets that developed strictures with samples from 52 non-stricture, NEPRC-sourced marmosets. While captivity increases susceptibility to GI disease in marmosets, the comparison of marmosets from a single-source and maintained within a single institution helps normalize the effects of stress and diet, which can affect the microbiome29. Stricture-associated dysbiosis featured shifts the relative abundance of Bacteroides, Anaerobiospirillum and Megamonas (Fig. 1), but commonly used analyses, such as alpha and beta diversity, showed no significant changes or inconclusive results, respectively.

In order to gain further insights into role of the microbiome in duodenal strictures, we used machine learning to identify ASVs of importance that could help generate testable hypotheses. Our analysis utilized four classifiers: RF, SVM, CART and KNN, which usually outperform traditional supervised classifiers30,31,32,33. Due to inherent differences in each algorithm, we benchmarked the methods to help identify the correct algorithm for classification of strictures and non-strictures given our dataset, and avoid the potential for bias and overfitting that exists when only a single algorithm is evaluated33,34. After evaluating the performance of the four methods, we found that random forest models had the best performance based on both accuracy and kappa metrics. In our RF model of the microbiome data, the model was optimized with 9 of the 601 ASVs generated in the QIIME2 workflow. Model stability observed with only a small portion of the data is desirable as selection of relevant features from noisy data, that is dimensionality reduction, is one of the main tasks in machine learning33,35,36,37. After reaching the optimal number of ASVs, the inclusion of other ASVs to the model adds redundant data, and eventually noisy data, without a significant improvement in model performance. As data from only 9 ASVs is required to correctly differentiate the stricture and non-stricture samples, these ASVs represent candidate biomarkers of the two states, which streamlines hypothesis generation and testing. Both our ANCOM analysis and RF model highlighted the importance of decreases in Anaerobiospirillum and increases in Clostridium sensu stricto 1 in stricture cases. The consensus between two different analyses led us to further investigate Anaerobiospirillum and Clostridium sensu stricto 1. While Anaerobiospirillum has been previously reported in healthy marmosets, dogs, and cats38,39, these bacteria may cause GI disease in humans38. However, Anaerobiospirillum was present in high abundance in our healthy marmosets, and reduced levels were seen in stricture cases.

Using both microbial culture and sequence analysis, we determined that C. perfringens was observed at higher levels in the duodenal lesions of diseased animals. C. perfringens is a known GI pathogen that can encode multiple toxins (alpha, beta, epsilon, iota, perfringolysin O, and enterotoxin)24. In marmosets and other NHP, C. perfringens can cause gas gangrene and gastric dilatation syndrome40,41,42. Of note, C. perfringens-induced gas gangrene was reported in the Japanese vivarium that first reported duodenal strictures in captive marmosets40. In the United Kingdom, C. perfringens is one of the top 5 causes of foodborne death43, and has been linked to diarrhea, Clostridial necrotizing enteritis (CNE), necrotizing enterocolitis (NEC), ulcerative colitis (UC) and enterotoxemia in humans and other mammals24,44. CNE is a necrotizing inflammation of the small intestine that can induce mild diarrhea or severe abdominal pain, vomiting and ulcers24. NEC predominantly affects infants due to intestinal immaturity or dysbiosis24,44. While these symptoms match the clinical presentation of duodenal strictures in marmosets, they are non-specific. However, both small and large intestinal strictures developed in 11–29.5% of NEC infants and could occur up to 20 months post-NEC diagnosis45,46. Based on the site of C. perfringens infection at the junction of the duodenum and the common bile duct, we hypothesize that bile acid (BA) deregulation due to dysbiosis or antibiotic treatment may have facilitated C. perfringens infection. Antibiotic usage in infants has been linked with increased NEC risk47, and antibiotics are commonly prescribed to treat NHP GI diseases. Furthermore, C. perfringens was overrepresented in dogs with chronic enteropathy, an IBD-like disease, and bacterial abundance was regulated by secondary BAs (deoxycholic acid and lithocholic acid) that are produced by gut bacteria48,49. Due to its status as a known pathogen and presence in the site of injury, we propose that C. perfringens is a potential causative agent of duodenal disease in marmosets.

In addition to the role of C. perfringens, our serum chemistry and clinical chemistry-based RF models were highly sensitive in accurately classifying strictures. Decreased total protein levels are often observed with GI disease and may indicate poor digestion/absorption. The importance of amylase and lipase in our stricture model is supported by clinical findings of cholecystitis and secondary pancreatitis5. Secondary pancreatitis, attributed to extension from duodenal ulcers, was observed in 15 of 17 cases scored5. In the CBC-based model, HCT, HGB, RBC, RDW, and MCH relate to red blood cell function and suggested anemia. Anemia, a common finding in marmosets with strictures and IBD3,5, is also a risk factor for NEC in humans50. Interestingly, transcriptomic analysis of strictures showed enrichment of lipid metabolism and intestinal absorption genes, which may reflect enterocyte damage and is consistent with lipidomic alterations induced by C. perfringens alpha-toxin, a phospholipase C51. Increased expression of FABP1 and FABP2 was observed. These genes encode for liver and intestinal fatty-acid binding proteins (LFABP and IFABP), respectively, and are often used as biomarkers of GI diseases, including NEC47. To our knowledge, correlations of gut FABP2 levels with serum IFABP levels have not been described, but we hypothesize that increased expression might be a compensatory mechanism triggered by enteritis. While increased inflammatory responses were not observed due to the lack of healthy control tissue, based on the C. perfringens infection, development of enteritis, anemia and strictures and deregulation of lipid metabolism, we believe marmosets could be developed as a model to investigate the mechanisms of bacterially-driven CNE/NEC.

A potential limitation of this study is potential imbalance in the dataset due to the scarcity of stricture samples relative to non-stricture samples. Imbalanced datasets are commonly encountered in machine learning applications as real-world classification problems, such as fraud detection, medical diagnosis, etc., are usually imbalanced52. While the 1:3 ratio of stricture to non-stricture samples is not generally considered an imbalanced dataset, the application of our benchmarking strategy to iteratively sample the dataset and assess performance metrics gives us confidence of the robust model performance. Another limitation of this study was the inability to ethically obtain age-matched duodenal samples from healthy, MITNE marmosets due to the early onset of the disease. In both analyses involving duodenal tissue, stricture samples were compared to duodenal tissue collected from sex-matched marmosets undergoing non-stricture-related necropsies. Additionally, we excluded non-stricture samples presenting with gross pathology due to other diseases, but we were not able to match the age and source. Previous microbiome studies have found minimal differences in the marmoset microbiome associated with age18,22. We have found that source strongly influence microbiome composition22, and while the microbiome may influence host responses, our interest in the RNAseq analysis was to elucidate the transcriptomic response to duodenal strictures. Future studies will focus on addressing small sample size concerns by banking MITNE duodenal tissues for further metagenomics and transcriptomic studies with MITNE stricture cases.

The common marmoset has emerged as useful NHP model for studying human disease as marmosets are small size, are easier to handle, are less costly to maintain than other NHP and mimic human disease53. Based on the presentation of bacterially-driven, intestinal inflammation in young adults presenting with no other illness, we propose that marmosets can be a viable model to investigate C. perfringens-associated enteritis. Better understanding of these disease profiles, the effects of diet and husbandry, and their inherent robustness to insults and disease will be helpful in promoting animal health, developing better models of human disease and understanding how to modulate microbial communities.

Materials and methods

Ethics statement

All research was conducted under an animal use protocol approved by the MIT Institutional Care and Use Committee (IACUC). The facility where this research was conducted is accredited by the AAALAC, International and adheres to principles stated in the Guide for the Care and Use of Laboratory Animals. Methods were carried out in accordance with the ARRIVE guidelines. Animals are cared for by a large staff of highly qualified veterinarians, veterinary technicians, and animal caretakers, who undergo substantial training to ensure only the highest quality animal care and use.

Animals

Common marmosets (Callithrix jacchus) were housed at the Massachusetts Institute of Technology in Cambridge, MA, and were originally imported from the New England Primate Research Center (NEPRC) in 2014. This source is referred to as MITNE. All animals were housed in pairs or family groups within one vivarium at MIT, an AAALAC International accredited facility. This study included 32 male and 41 female marmosets. Of the 52 healthy (non-progressor) animals, 28 were female and 24 were male, while in the stricture (progressor) cohort, 13 were female and 8 were male (Table 1). Samples from non-stricture marmosets were collected from marmosets ranging between 0.31 and 13.01 years of age, while samples from stricture animals were collected from marmosets aged 0.86–6.08 years of age. The animal holding room temperature was maintained at 74.0 + /−2°F with a relative humidity of 30–70%. The light cycle was maintained at a 12:12 h light:dark cycle. Marmosets were housed in cages composed of stainless-steel bars and polycarbonate perches with the following dimensions: 30″ W × 32″ D × 67″ H). Each cage had a nest box made of polycarbonate attached the outside of the cage. Other cage furniture present in the cages included hammocks, hanging toys, and manzanita wood branches. Foraging enrichment in the form of dried acacia gum-filled branches and forage board were provided weekly. Cages were spot-cleaned daily and removed for sanitization on a biweekly rotation.

All animals received a base chow diet of biscuits (Teklad New World Primate Diet 8794). Initially, biscuits were soaked in water for at least 20 min, but the practice was then changed to a pour-on/pour-off soak only. About halfway through the 2 year period encompassing this study, biscuit prep protocol reverted to the original practice of a 20 min soak to alleviate any concerns that soaking duration could be contributing to the development of duodenal ulcers. In addition to the base chow, a cafeteria-style supplemental offering of fruits (e.g. bananas, blueberries, mangoes, apples and grapes), vegetables (e.g. carrots, vegetable blend), acacia gum, and additional protein sources including hard-boiled eggs, mealworms, cottage cheese or ZuPreem (Premium Nutritional Products, Inc., Mission, KS).

On a semiannual basis, preventative health physical exams were performed on all colony animals. Rectal swabs and fecal samples were collected and screened for potentially pathogenic bacteria (including Salmonella spp., Shigella spp, beta-hemolytic E.coli, Klebsiella spp., and Campylobacter spp.) and parasites (including Enterobius spp., Entamoeba spp., Giardia spp., Taenia spp., and Cryptosporidium spp.). Additional fecal and rectal swab samples were collected between 2016 and 2018 for microbiome analysis. Intradermal testing for Mycobacterium tuberculosis was performed semiannually as well. All animals derived from progenitor stock were negative for squirrel monkey cytomegalovirus, Saimiriine herpesvirus 1, Saimiriine herpesvirus 2, and measles virus. Complete blood count and serum chemistry analysis were performed on an annual basis and during diagnostic workup of clinical cases. Hematology analysis was performed by the MIT DCM diagnostic laboratory using a HemaVet 950 veterinary hematology analyzer (Drew Scientific, Oxford, CT). Serum chemistry analysis was performed by Idexx Laboratories (Westbrook, ME). Serum chemistry and complete blood counts data were collected from the clinical records from the MIT colony. Investigators collecting samples were aware of health status, but investigators processing samples were blinded.

Bacterial culture methods

Duodenal tissue and duodenal contents collected from MIT common marmosets during necropsies performed by clinical veterinarians and veterinary pathologists were evaluated. Representative sections of major organs were collected, fixed in 10% neutral buffered formalin, embedded in paraffin, sectioned at 5 µm, and stained using hematoxylin and eosin (HE) for scoring by a boarded veterinary pathologist. Samples were flash frozen in vials containing Brucella broth in 20% glycerol and frozen at − 80 °C. Samples from stricture and non-stricture cases were selected. The tissues were thawed in an anaerobic atmosphere (10% CO2, 10% H2, 80% N2), and were homogenized with freeze medium with tissue grinders. The homogenate was divided into the following aliquots. For aerobic culture, the homogenates were plated onto chocolate agar, blood agar, MacConkey agar, and Brucella Broth medium containing 10% FCS. The plates were incubated at 37 °C in 5% CO2 for 24–48 h. For anaerobic culture, the homogenates were plated onto pre-reduced Brucella Blood Agar plates (BBL) and inoculated into thioglycollate broth. The cultures were incubated at 37 °C in an anaerobic chamber (Coy Lab Products) with mixed gas (10% CO2, 10% H2, 80% N2) for 48 h. For microaerobic culture to detect the growth of Helicobacter spp., the homogenates were plated onto selective antibiotic-impregnated plates (50 μg/ml amphotericin B, 100 μg/ml vancomycin, 3.3 μg/ml polymyxin B, 200 μg/ml bacitracin, and 10.7 μg/ml nalidixic acid)54 and Brucella Blood Agar plates after passing through 0.65 µm syringe filter. The plates were placed into a vented jar filled with mixed gas (10% CO2, 10% H2, 80% N2) and incubated at 37 °C for up to 3 weeks. The plates were checked every 2–3 days for growth. Aliquots of the homogenates were also used for DNA extraction using the High Pure PCR Template Preparation kit (Roche Molecular Biochemicals). Bacterial DNA was then subjected to 16S rRNA gene sequence analysis using conserved primers 9F 5′-GAGTTTGATYCTGGCTCAG-3′ and 1541R 5′-AAGGAGGTGWTCCARCC-3′ creating a 1.5 kb product from the 16S rRNA gene55. The amplicons were purified with a QIAquick PCR Purification kit (Qiagen) and directly sequenced using an ABI Prism BigDye terminator cycle sequencing ready reaction kit on a genetic analyser 3500 (Applied Biosystems).

16S microbiome profiling

Fecal DNA was extracted using the DNeasy PowerLyzer PowerSoil Kit, and DNA was amplified using universal primers of F515 (GTGYCAGCMGCCGCGGTAA) and R926 (CCGYCAATTYMTTTRAGTTT) to target the V4 and V5 regions of bacterial 16S rRNA fused to Illumina adaptors and barcode sequences as described previously56. Individual samples were barcoded and pooled to construct the sequencing library, followed by sequencing with an Illumina MiSeq instrument to generate pair-ended 300 × 300 reads. Sequencing quality was inspected using FastQC57. Reads were processed using QIIME 2–2018.6 within the MicrobiomeHelper v. 2.3.0 virtual box56,58. Briefly, primer sequences were trimmed using the cutadapt plugin59. Forward and reverse reads were truncated at 243 and 195 bases, respectively, prior to stitching and denoising reads into amplicon sequence variants (ASV) using DADA2. Samples with fewer than 7500 reads were excluded. ASVs present in fewer than 3 samples and with less than 24 counts were also excluded. Taxonomic classification was assigned using the custom 16S V4/V5 region classifier based on the SILVA 132 database (SSU Ref NR 99)60. Following initial quality control, 601 ASVs proceeded to further analysis in the fecal microbiome analysis. Phylogenetic trees, composition, alpha rarefaction, beta diversity metrics and ANCOM (Analysis of Composition of Microbiome)61 were evaluated using built-in QIIME2 functions62. Microsoft Excel and R (v 3.6.3 at http://www.R-project.org/) were used to perform statistical analyses and graphically represent data. Duodenal tissue samples flash frozen in liquid nitrogen were obtained from necropsies performed by clinical veterinarians and veterinary pathologists on 29 marmosets. Duodenal stricture samples were obtained from 7 males and 10 females ranging from 1.71 to 8.44 years of age. Non-stricture duodenal samples were obtained from 7 males and 5 females ranging from 1.82 to 10.4 years of age. Necropsy samples were collected from MITNE (n = 18) and two additional MIT sources (n = 11). Tissues were processed as described above for fecal samples to determine the relative abundance of Clostridium sensu stricto 1 in the duodenum.

Additionally, R libraries phyloseq63, ggplot2 (2.2.1)64, caret65, vegan66, pROC67, and gtools68 were used to model microbiome data. 6 samples (5 rectal swabs and 1 fecal sample) were excluded from microbiome analysis due to poor sampling characterized by low quantities of visible fecal matter and a microbiome dominated by a single species (e.g. Helicobacter) that was discordant from samples from the same individual19. We analyzed the Bacteroides/Prevotella abundance ratio by taking the ratio of the averaged Bacteroides abundance and the averaged Prevotella abundance.

Machine learning

Machine learning analysis was performed following the approach of our previous work using a strategy to benchmark classifiers to the identify the most suitable method for the each particular dataset33. Data from the microbiome, serum chemistries and complete blood counts were utilized to train classifiers. To minimize the stress caused by handling and sampling, banked samples collected during physical examinations were utilized. As testing needs varied for each exam, paired blood and microbiome samples were not available at every time point. Due to this limitation, we generated three independent models for the microbiome, serum chemistries and CBC data.

Data was normalized using min–max normalization. The data was then split using a single partition method and the classifiers were trained on 80% of the samples (training) and the discovered signatures were used to predict the populations on the remaining 20% of samples (testing) using the four machine learning approaches: support vector machines (SVM), random forest (RF), K-nearest neighbor (KNN), and Classification and Regression Trees (CART). A R script using the function in the Caret package utilized default parameters for training with cross-validation. The variable importance metric was calculated using the varImp function, which associated a specific value for each parameter. To evaluate the contribution of each parameter, the script ranked the parameters and calculated the variable importance starting with the ranked parameters with the highest score. This process was processed iteratively adding ranked parameters and recalculating the metrics with each subsequent addition until all ranked genes were evaluated. Metrics included accuracy (correct classification percentage), kappa value (inter-rater classification agreement), sensitivity, specificity, precision, recall, prevalence, and F1 score (harmonic average of the precision and recall). Based on the contribution of each parameter, we selected a K value of top parameters based on the following criteria: (1) the stability of the metrics (priority for accuracy, kappa, and F1) when the increment of ranked genes was done, and (2) minimum number K of parameters as possible. After the selection of the K value, ROC (Receiver-operating characteristic) curve and AUC (Area under the curve) value were calculated for each algorithm.

RNAseq

Tissues were collected from the duodenum from marmosets with either stricture or IBD during necropsies performed by clinical veterinarians and veterinary pathologists. The three stricture marmosets were female sourced from the MITNE and aged 1.9, 2.7, and 2.9 years old (average 2.5 ± 0.4). The stricture duodenal samples were distal of the site of stricture (n = 3) and exhibited chronic active duodenitis and duodenal ulcers. Given the early onset of this disease, it was not ethically justifiable to collect duodenal samples from healthy MITNE females. Non-stricture marmosets were selected from MITCL female marmosets, aged 6.5, 6.6, and 7.0 years old (average 6.7 ± 0.2), that were undergoing necropsy due to IBD. Non-stricture samples were collected from IBD animals from the same region of the duodenum, and presented with mild thickening based on gross observations (n = 3). Tissues were flash frozed in liquid nitrogen and stored at − 80 °C. RNA was extracted using TRIzol reagent according to manufacturer’s instructions (Thermo Fisher Scientific). Total RNA was shipped on dry ice to Arraystar, Inc. (Rockville, MD) for quality control, rRNA depletion and sequencing on an Illumina HiSeq4000. FASTA files and the NCBI RefSeq GTF files for Callithrix jacchus based on the March 2009 (WUBSC 3.2/calJac3) assembly were obtained from the UCSC Genome browser69. Raw sequencing reads were mapped to an index built from C. jacchus FASTA files using the buildindex function in Rsubread70. 29,575 feature counts were obtained via the featureCounts function from the bam files using annotated exons in the C. jacchus GTF files. Analysis was then performed using edgeR71,72. Lowly expressed exons were removed using a cutoff of 10 counts per million (CPM) and presence in at least 2 samples. Normalization was performed using the Trimmed Mean of M-values (TMM) method using the calcNormFactors function. Following removal of lowly expressed exons and normalization, a dataset with 19,254 feature counts was further analyzed. Multidimensional scaling (MDS) plots and heatmaps were used to evaluate grouping of biological samples. Data were fitted using the glmQLFit function that uses a generalized linear model (GLM) implementing a quasi-likelihood (QL) fitting method. Quasi-likelihood F-tests were performed to test for differential expression using the decideTestsDGE function in edgeR using the Benjamini–Hochberg correction and False Discovery Rate (FDR) adjusted P-values of 0.05. To retrieve Gene Ontology (GO) classifications, C. jacchus genes that matched Homo sapiens gene names were assigned both the C. jacchus and Homo sapiens Entrez IDs. GO analysis was performed using limma73, AnnotationDbi74, GO.db75, topGO76, mygene77 and org.Hs.eg.db.

Data were visualized using ggplot2, gplots, Rgraphviz78, colorspace79 and ggVennDiagram80.