Enhanced microbiota profiling in patients with quiescent Crohn’s disease through comparison with paired healthy first-degree relatives

Impact of environmental and host genetic variations on gut microbiota composition in patients with CD

The current study was based on the DamnIBD cohort (detailed in STAR Methods), in which every patient with IBD was paired with a HFDR. A total of 52 pairs of patients with CD with healthy controls were enrolled, including 27 patients with CD with active inflammation and their paired siblings (CD-A/sibling-A) and 25 patients with CD in remission and their paired siblings (CD-R/sibling-R). Patients and their paired HFDRs had been living together for a substantial period, and approximately half of the patients were still living together with their HFDRs. Additionally, an independent control group was enrolled, consisting 44 healthy subjects not related to the patients with CD (non-relative). Patients in the study groups (CD-A, sibling-A, CD-R, sibling-R, and non-relative) had similar ages, gender ratios, and body mass indices (Table S1). The average dietary intake of macro- and micronutrients showed no significant differences between patients with CD-R and their HFDRs, as well as between patients with CD-R and the non-relative controls, as determined by 24-h dietary recall (Table S2). On the other hand, the dietary intake of fiber, calcium, potassium, vitamin C, and beta carotene was lower in patients with CD-A compared to their HFDRs (Table S2), suggesting that different dietary habits could be a significant confounding factor in the study of the microbiota of CD-A.

From the 148 fecal microbiome samples, 86,805 reads of 16S rRNA gene were sequenced and assigned to 13,616 amplicon sequence variants (ASVs). After excluding rare ASVs found in only two or fewer samples, 2,492 ASVs remained and were classified into 223 genera across 89 families and 11 phyla (Figures 1A, S1A, and S1B). Similar patterns of microbial compositions were observed between the CD-A and the CD-R groups, as well as among the healthy control groups (sibling-A, sibling-R, and non-relative).

The alpha and beta diversities of the study groups were assessed at the ASV level. When evaluating ecological diversities within each sample using a phylogenetic distance metric, both CD-A and CD-R groups exhibited lower diversities compared to their respective paired controls, while no significant difference was observed between the CD-A and the CD-R groups (Figure 1B). Similar results were obtained using Shannon index and other alpha diversity metrics (Figures S1C and S1E). Ecological diversities within the study groups were measured by UniFrac-based principal coordinate analysis, showing that the beta diversities of the CD-A and the CD-R groups were substantially different from those of the sibling-A and the sibling-R groups, respectively (Figure 1C). No notable divergence was observed between the CD-A and the CD-R groups, nor between the sibling-A and the sibling-R groups.

Compositional changes were observed in the fecal microbiota at every taxonomic level in both the CD-A and the CD-R groups. At the phylum level, 8 and 2 differential phyla were observed in the CD-A and the CD-R groups, respectively, in comparison to the non-relative controls (Figure 1D). However, only Firmicutes and Proteobacteria were identified as differential phyla for both the CD-A and the CD-R groups, when compared to their paired HFDR control groups (sibling-A and sibling-R) (Figure 1D). At the genus level, 98 and 54 differential genera were observed in the CD-A and the CD-R groups, respectively, in comparison to the non-relative controls (Figures 1E and S2). Strikingly, when comparing with their paired HFDRs, the numbers of differential genera in the CD-A and the CD-R groups were drastically reduced to 27 and 3, respectively (Figures 1E and S2). Similar observations were made at the family level (Figure S2). Further analysis using 25 non-relative controls randomly selected from a total of 44 reaffirmed that utilizing HFDR controls resulted in fewer differential taxa compared to non-relative controls (Figure S3).

Thus, based on the differential taxa counts, the microbiota variances between patients with CD and their HFDRs appeared to be less pronounced than the variances observed between patients with CD and non-relative controls. Consistently, at the community level, the UniFrac distances between the CD and the paired HFDR groups were smaller than those between the CD and the non-relative groups, for both active and quiescent CD (Figure S4).

To further evaluate the differences in microbial compositions between the HFDR control and the non-relative control groups, we compared the metagenomic and metabolomics data between these two groups. Surprisingly, we found no significant differences in microbial genera, pathways, or metabolomics levels between these two groups (Table S3. Relative abundances of the bacterial genera in the HFDR and non-relative control groups, related to Figure 1, Table S4. Relative abundances of the bacterial pathways in the HFDR and non-relative control groups, related to Figure 1, Table S5. Fecal levels of the short-chain fatty acids in the HFDR and non-relative control groups, related to Figure 1). Consistently, the number of shared taxa between HFDRs and the CD-R groups did not differ significantly from the number of shared taxa between the non-relative and the CD-R groups, at the phylum, family, and genus levels (Figure S5). These findings suggest that the average microbial compositions and functions were similar between the HFDR and the non-relative control groups. The distinguishing factor could be that HFDR controls allow paired analysis for the identification of differential microbiota, unlike the non-relative controls.

Altered gut microbiota in patients with active and quiescent CD

The aforementioned results underscore the substantial impact of environmental and host genetic variations on the gut microbiota. In light of these results, paired HFDRs were employed as control subjects for further analyses of the microbiota in patients with CD.

At the phylum level, Firmicutes, Bacteroidota, Proteobacteria, and Fusobacteriota were the dominant phyla across all study groups. Both the CD-A and the CD-R groups exhibited decreased abundances in Firmicutes, along with increased abundances in Proteobacteria, when compared to their paired HFDRs (Figure S1A).

At the genus level, 27 differential genera were found between the CD-A and the sibling-A groups, whereas only 3 differential genera (Dorea, Fusicatenibacter, and Faecalibacterium) were identified between the CD-R and the sibling-R groups (Figures 2A and 2B). Among these differential genera, 25 were specific to CD-A, including Escherichia-Shigella (increased), an undefined genus of Enterobacteriaceae (increased), Prevotella (decreased), and Roseburia (decreased). Faecalibacterium (decreased) was the sole differential genus specific to CD-R. Dorea (decreased) and Fusicatenibacter (decreased) were common differential genera observed in both the CD-A and the CD-R groups (Figures 2C; Tables S6 and S7). According to the current hypothesis, the decreased abundances of Dorea, Fusicatenibacter, and Faecalibacterium in CD-R may potentially contribute to future relapses of intestinal inflammation. On the other hand, the decreased abundances of Dorea and Fusicatenibacter in CD-A indicate their possible involvement in sustaining intestinal inflammation.

A notable pattern of gradually altered abundances from healthy controls to CD-R and then CD-A was observed for the majority of the differential genera in the CD-A group, indicating associations of these genera with the inflammatory activities (Figure 3). However, this pattern was not observed for two of the three differential genera (Dorea and Faecalibacterium) in CD-R (Figure 3), indicating that the alterations in these genera were not a consequence of inflammation.

To gain further insight into the relationship between microbial composition and intestinal inflammation status, we examined possible associations between microbial genera and serum markers of inflammation, focusing on the 28 differential genera in CD-A and CD-R. The presence of significant correlations between microbial genera and serum inflammation markers was more frequently observed in CD-A compared to CD-R (Figure 4A). In CD-A, the differential genera Escherichia-Shigella correlated with four of the serum inflammation markers including hyper-sensitive C-reactive protein (hsCRP), C-reactive protein, platelet count, and erythrocyte sedimentation rate, followed by Veillonella correlating with three of the inflammation markers. In CD-R, no correlation was observed between any differential genera (Dorea, Faecalibacterium, and Fusicatenibacter) and the serum inflammation markers.

These results have significant implications for understanding the causal relationship between the microbiota and inflammation in CD.

Altered microbial functions in the CD-A and the CD-R groups

Functional analyses based on the 16S rRNA sequences revealed that the fecal microbiota of the CD-A were enriched in dozens of oxygen-dependent metabolic pathways, including fatty acid β-oxidation, sugar alcohol fermentation, and the tricarboxylic acid cycle, in comparison to their paired HFDR controls (Figure S6). This enrichment correlated with higher abundances in facultative anaerobes, such as Escherichia-Shigella. Furthermore, elevated abundances in lipopolysaccharide biosynthesis in CD-A (Figure S6) were consistent with an elevated representation of gram-negative bacteria, such as Proteobacteria. Conversely, patients with CD-A displayed decreased abundances in methanogenesis, a function typically associated with anaerobic bacteria (Figure S6). This finding aligns with the observed decrease in anaerobic bacteria within the CD-A group.

In contrast, much less differential pathways were identified in CD-R, which echoed the fact that less differential taxa were identified in CD-R compared to CD-A. Importantly, one of the top differential pathways, “pyruvate fermentation to acetate and lactate II,” was decreased in CD-R, as well as in CD-A (Figure S6; Tables S8 and S9). Two other pathways involved in acetate production, “bifidobacterium shunt” and “heterolactic fermentation,” were decreased in the CD-A group and showed a trend of down-regulation in the CD-R group (Figure S6). These data collectively suggest a reduced capability for acetate production within the gut of both the CD-R and the CD-A groups.

Correlations between fecal SCFA levels and gut microbes in the CD-A and the CD-R groups

The aforementioned results strongly indicate the relevance of SCFAs in the pathogenesis of CD. We thus examined the fecal levels of the SCFAs by targeted metabolomics analysis. The CD-A group exhibited lower levels of butyrate and valerate compared to their paired HFDRs. Additionally, a trend of lower levels was observed in other SCFAs, including acetate, propionate, iso-butyrate, iso-valerate, and caproate (Figure 4B, upper left panel). Similarly, the CD-R group also exhibited trends of lower SCFA levels, including propionate, iso-butyrate, butyrate, iso-valerate, valerate, and caproate (Figure 4B, upper right panel).

Next, we performed correlation analyses and found associations between the abundances of several differential genera and SCFA levels. As expected, the CD-R-specific differential genera, Faecalibacterium and Fusicatenibacter, showed correlations with multiple SCFAs (Figure 4B, lower panels). However, it was unexpected that Butyricicoccus, known for its butyrate production, did not correlate with butyrate levels, possibly due to the prevalence of other butyrate-producing bacteria like Faecalibacterium in the gut.²⁴ Interestingly, some of the differential genera (e.g., Faecalibacterium) were negatively correlated with acetate levels, likely due to the consumption of acetate by these bacteria.

Potential diagnostic value of the differential genera

We next evaluated the diagnostic potential of CD-R-specific microbial markers, namely Dorea, Fusicatenibacter, and Faecalibacterium. A classification model was constructed with these three genera and achieved an area under the receiver operating characteristic curve (AUROC) of 0.82 for distinguishing CD-R from sibling-R and 0.88 for distinguishing CD-R from non-relative controls (Figure 5).

Intriguingly, Faecalibacterium, Dorea, and Fusicatenibacter also demonstrated impressive performance in distinguishing CD-A from sibling-A and non-relative controls, with AUROC values of 0.85 and 0.93, respectively (Figure 5). However, this classification model did not perform well in differentiating between CD-R and CD-A (AUROC = 0.68, Figure 5), which aligns with the similar alteration trends observed in both the CD-R and the CD-A groups for these three genera.

Validation of microbial features in CD with a large external multi-center cohort

We utilized a large whole-genome shotgun sequencing dataset, iHMP (Integrated Human Microbiome Project),²² to validate our findings and test the diagnostic potential of CD-R-specific microbial features in a diverse population. This dataset was generated from a multi-center cohort of patients with CD with diverse cultural and geographical backgrounds, including 383 patients with CD (CD-A: 76, CD-R: 307) and 190 healthy controls (HC).

Several phyla displayed trends of altered abundances in CD-A and CD-R, but no statistical significance was achieved (Figure S7A; Table S10). At the family level, 10 differential families were identified between the CD-A group and the healthy control group (Table S11), among which Rikenelleaceae and Oscillospiraceae displayed decreased abundances in CD-A, consistent with the observations in the discovery cohort (Figure S7B). In CD-R, 6 differential families were observed compared to healthy controls, among which Ruminococcaceae displayed decreased abundances in CD-R (Figure S7B), aligning with the findings from the discovery cohort.

At the genus level (Table S12), 17 differential genera were identified between CD-A and healthy controls. Decreased abundances of SCFA-producing Alistipes, Roseburia, and Butyricicoccus were observed in CD-A, consistent with the discovery cohort (Figure 6A). On the other hand, 14 differential genera were identified between CD-R and healthy controls, among which Faecalibacterium and Fusicatenibacter displayed decreased abundances in CD-R (Figure 6A), in line with the findings from the discovery cohort. Dorea, showing decreased abundance in CD-R in the discovery cohort, exhibited a similar decreasing trend in the validation cohort. It is worth noting that SCFA-producing genera Agathobaculum and Eubacterium displayed decreased abundances in CD-R (Figure 6A). As observed in the discovery cohort, there was no gradual alteration in Faecalibacterium abundance from healthy to CD-R and then to CD-A (Figure S8). This supports the notion that the abundance of Faecalibacterium is not associated with the status of inflammation.

At the functional level, 25 pathways were differentially enriched between the CD-A and the healthy control groups (Figures 6B; Table S13). Among them are higher representations in oxygen-requiring metabolic pathways, including octane oxidation and oleate β-oxidation, while a lower representation in pathway anaerobic energy metabolism in the CD-A group reflects a higher abundance of facultative anaerobes in the CD-A group, consistent with the findings from the discovery cohort. In CD-R, 14 differential pathways were identified compared to the healthy control group (Figures 6C; Table S14). In line with the findings from the discovery cohort, pathways contributing to SCFA production, including anaerobic energy metabolism and bifidobacterium shunt, displayed decreased abundances in the CD-R group.

Next, the diagnostic potential of Dorea, Fusicatenibacter, and Faecalibacterium was evaluated with the validation cohort. Using a random forest classification model, the combination of these three biomarkers was able to distinguish CD-R from healthy controls, as well as CD-A from healthy controls, achieving AUROCs of 0.72 and 0.80, respectively (Figure 6D). As with the discovery cohort, these markers showed limited ability in distinguishing CD-A from CD-R. The diagnostic capabilities of differential KOs between CD-R and the healthy controls were also assessed with the validation cohort. We used 74 differential KOs associated with SCFA production to develop a random forest model, which achieved an AUROC of 0.75 for distinguishing CD-R from healthy controls and an AUROC of 0.71 for distinguishing CD-A from healthy controls (Figure 6E).

For further insights on the diagnostic potential of the microbial markers, we stratified the validation cohort according to race, ethnicity, geography, or diet (specific food consumption within the past week) and evaluated the diagnostic performances of the combined markers of Dorea, Fusicatenibacter, and Faecalibacterium across the subgroups. The white subgroup was the sole focus for race analysis due to underrepresentation in other racial groups. Similarly, the non-Hispanic subgroup was solely considered for ethnicity analysis given similar limitations. Stratification by race and ethnicity slightly improved the model’s performance (Figures S9A and S9B). For example, among the white subgroup, the combined markers achieved an AUROC of 0.75 (vs. 0.72 for the entire cohort) for distinguishing CD-R from healthy controls and 0.81 (vs. 0.80 for the entire cohort) for distinguishing CD-A from healthy controls. Geographical subgroup analysis (medical centers) (Figure S9C) indicated higher marker performance in cohorts from Massachusetts General Hospital and Cedars-Sinai Medical Center compared to Cincinnati Children’s Hospital (CCHMC), potentially linked to the younger age of CCHMC participants (average age of 12.13 years vs. over 15 years for other groups). Noteworthy observations were observed in dietary intake subgroup analysis (Figures S9D and S9E), with improved diagnostic performances in specific subgroups like no sweet, no vegetable, no fruit, and no soft drinks categories. Overall, these subgroup analyses suggest that demographic and lifestyle factors, particularly age and dietary habits, may have a large impact on the diagnostic performance of the microbial markers.

Disease specificity of the CD-R-specific microbial markers identified using HFDR controls

To test the disease specificity of the CD-R-specific microbial markers, we first examined the abundances of these three genera (Dorea, Faecalibacterium, and Fusicatenibacter) in five other diseases linked with microbial dysbiosis, using public datasets of ulcerative colitis (UC),²² colorectal cancer (CRC),²⁵ cardiometabolic diseases (CADs),²⁶ autism spectrum disorder (ASD),²⁷ and nonalcoholic fatty liver disease (NAFLD)²⁸ (Figure S10A), along with a validation cohort of CD.²² The results reveal distinct abundance patterns of these three genera in CD compared to the other conditions assessed. While all genera displayed decreased levels in CD, Fusicatenibacter exhibited increased abundance in UC, and Faecalibacterium remained unchanged in CRC. Notably, none of these three genera showed altered abundances in CADs, ASD, or NAFLD. In instances where changes were observed in the abundances of these genera in UC or CRC, they were less than 2-fold. Next, the three markers were used to construct random forest classification models to distinguish between the five aforementioned diseases and their respective control groups. By computing the average AUROC from ten iterations of 5-fold cross-validation for each disease and comparing them against a validation cohort of CD, we demonstrated that the performance of the markers was significantly better for CD than for the other five diseases (Figure S10B).

Superior performance of the CD-R-specific microbial markers identified using HFDR controls versus those identified using unrelated controls

Next, we compared the performance of the CD-R-specific microbial markers we identified using HFDR controls (i.e., Dorea, Fusicatenibacter, and Faecalibacterium) to those established using unrelated controls, including two microbial genera panels (marker panel 1²⁹ and 3³⁰) and one microbial family panel (marker panel 2²⁹). To ensure fair comparisons, random forest classification models constructed with these marker panels were tested on eight independent cohorts. Among these cohorts, iHMP and iHMP-pilot data^22,31 included patients with CD-A, patients with CD-R, and unrelated healthy controls, so that the models were tested to distinguish between CD-A and HC, between CD-R and HC, as well as between CD-A and CD-R. Except for one case where outside marker panel 1 and our panel (“this study”) showed similar classification power with iHMP (Figure 7A), our marker panel consistently outperformed the other marker panels in distinguishing between CD-R and HC, as well as between CD-A and HC, with both iHMP (Figure 7A) and iHMP-pilot data (Figure 7B). The classification power of the marker panels was also evaluated using six additional cohorts of patients with CD where information regarding the disease status (active or remission) of patients with CD was unavailable. In these cases, the models were solely tested to differentiate between CD and HC. Across all six cohorts, our marker panel consistently demonstrated satisfactory performance with AUROCs above 0.85 (Figures 7C–7H). Conversely, the outside marker panels showed either inconsistent performance (marker panels 1 and 2) or poor performance (marker panel 3) (Figures 7C–7H). These data demonstrate superior diagnostic potential of our CD-R-specific microbial markers identified using HFDR controls compared to previously published markers identified with unrelated controls.

Using HFDRs to control for genetic and environmental confounding factors in microbiome studies of CD

In this study, we took a novel approach to investigate the gut microbiota in patients with CD by utilizing paired HFDRs of patients with CD as controls. Our approach resulted in fewer differential taxa and fewer microbial markers between patients with CD and controls, compared to using non-relative healthy individuals as controls. Notably, we identified microbial markers Faecalibacterium, Dorea, and Fusicatenibacter with decreased abundances in the gut of CD-R. Although each of the microbial features has been previously reported (e.g., decreased abundance of Faecalibacterium),^13,32 as a panel, the collection of the markers we identified is novel. We hypothesize that, by implementing HFDR controls, we were able to pinpoint the most reliable markers for CD. This hypothesis was substantiated by validating our findings with a large multicenter cohort, which demonstrated that the markers identified in our study are universally applicable.

Our findings demonstrate the substantial impact of genetic and environmental factors on microbial composition. Consequently, this raises an important question about whether these factors play pivotal roles in CD pathogenesis or necessitate careful consideration as confounding variables.

Because healthy relatives of patients with CD are at high risk for CD,^33,34 several studies have investigated the gut microbiota of the HFDRs of patients with CD. Although with inconsistencies in differential taxa, dysbiosis has been reported with the HFDRs and is proposed to be a potential causal factor for the development of inflammation.^32,35,36,37 However, since HFDRs do not display intestinal inflammation, we reason that their microbial characteristics may not be implicated in the recurrence of intestinal inflammation. The innocence of the microbial features in HFDRs of patients with CD was recently corroborated by a prospective study.²³ Raygoza Garay et al. reported no significant difference in microbial taxonomic abundance between HFDRs who developed CD in five years and those who remained healthy during follow-up, suggesting that the average microbiota in HFDRs does not harbor the key features driving the development of intestinal inflammation.

Considering that the microbiota in HFDRs is not directly linked to inflammation, it becomes crucial to use HFDRs to control for genetic and environmental confounding factors when studying the microbiota of patients with CD. This strategy enables the exclusion of numerous altered taxa that are shared between patients with CD and their HFDRs. As a result, it facilitates more precise and accurate identification of specific differential taxa that are genuinely associated with intestinal inflammation. Counterintuitively, there were no significant differences in microbial composition and function observed between the HFDR and the non-relative control groups, highlighting the critical role of paired analysis enabled by the paired HFDR controls.

Evidence for the causal role of altered gut microbiota in CD-R

Using HFDRs as controls to minimize the influence of genetic and environmental confounding factors, 27 and 3 differential genera were identified in CD-A and CD-R, respectively. Several lines of evidence showed that the decreased abundances in the differential genera in CD-R, Faecalibacterium, Dorea, and Fusicatenibacter, were not a consequence of inflammation but may indeed contribute to disease relapse through decreased production of SCFAs.

Firstly, because CD-R is prone to relapses, the decreased abundance of these genera in CD-R is linked to ensuing gut inflammation, thereby indicating their potential involvement in CD relapse. Notably, in both the discovery and the validation cohorts, Faecalibacterium was not identified as a differential genus for CD-A, suggesting no correlation between ongoing inflammation and the abundance of Faecalibacterium.

Secondly, while many of the differential genera showed a progressively altered abundance pattern from healthy controls to CD-R, and then to CD-A, this pattern was absent in two of the differential genera in CD-R, Faecalibacterium and Dorea, suggesting that the decreased abundance of Faecalibacterium or Dorea in CD-R was not associated with intestinal inflammatory activities.

Thirdly, correlations between differential genera and serum inflammation markers were prevalent in CD-A but absent in CD-R. This further reinforces the lack of correlation between current inflammation and the differential genera in CD-R, effectively refuting the possibility that the alterations in Faecalibacterium, Dorea, and Fusicatenibacter are a consequence of inflammation.

Lastly, fecal SCFA levels provide an additional dimension of evidence for the potential causal role of microbiota in CD-R on the relapse of intestinal inflammation. CD-R displayed trends of decreased levels in all species of SCFAs except acetate. Importantly, Faecalibacterium,³⁸ Dorea,³⁹ and Fusicatenibacter⁴⁰ are known for SCFA production, and the abundances of these genera correlated with many of the SCFAs, suggesting that decreased levels of SCFAs may underlie the pathological impact of the dysbiosis. The anti-inflammatory role of butyrate has been established through studies involving human cells and rodent colitis models.⁴¹ In line with the anti-inflammatory role of butyrate in IBD, decreased butyrate production potential has been observed in mixed populations of quiescent and active IBD cases.^42,43 Additionally, the decreased abundance in butyrate-producing Roseburia is associated with higher risk scores for IBD in healthy individuals.²¹

Therefore, our data demonstrate associations of CD-R with decreased abundances in Faecalibacterium, Dorea, and Fusicatenibacter and that these alterations are not a consequence of inflammation. In addition, both microbial functional and metabolomic analyses concur in identifying decreased SCFA as a potential mechanism for disease relapse. Collectively, these observations provide multiple layers of evidence supporting a potential causal role of altered microbiota in the relapse of intestinal inflammation. Recently, Raygoza Garay et al. established a microbiome risk score (MRS) that predicts the likelihood of a healthy individual developing CD.²³ Their findings align with ours at the functional level, as one of the top genera contributing to the MRS is SCFA-producing Roseburia, and a trend of decreased abundance in Roseburia is linked to disease onset.

One apparent limitation of our study is that cross-sectional data only provide evidence for correlation. A higher level of evidence could be obtained through a prospective study investigating the link between microbial features in CD-R and the subsequent disease relapse. With just 25 patients with CD-R, our current study could not be subgrouped to conduct this analysis with adequate statistical power. To ascertain a definitive causal role of the microbiota in CD, intervention studies are required to establish a conclusive relationship. In support of our hypothesis, research on mouse colitis models has shown that oral administration of Faecalibacterium or its metabolites reduces the severity of colitis.^44,45,46 Nonetheless, further intervention studies involving patients with CD are essential to confirm the causal role of the microbiota in disease relapse.

Potential diagnostic and therapeutic value of the microbial features in CD-R

With our internal discovery cohort and a large multi-center validation cohort with diverse geographical and cultural backgrounds, we demonstrated that microbial features in CD-R have diagnostic value for distinguishing CD from healthy individuals. Moreover, our research highlights the specificity of these microbial markers for CD, as they exhibited limited discriminatory power for other diseases involving microbiota. This observation further underscores the association of these markers with the pathogenesis of CD. Previous studies have shown the successful application of classification models based on microbial features alone^47,48 or in combination with an inflammation marker⁴³ to distinguish patients with IBD from non-IBD controls. In this study, we took it a step further by presenting a microbiome-based classification model that achieved satisfactory performances in distinguishing both CD-R and CD-A from healthy individuals. This is particularly valuable considering the challenges in identifying CD-R using current blood and pathology tests, as well as the non-invasive nature of the microbial markers.

The CD-R-specific microbial markers we identified using HFDR controls demonstrated superior performance compared to published microbial markers identified with unrelated healthy controls. Our marker panels consistently demonstrated satisfactory classification power in distinguishing between patients with CD and healthy controls with multiple independent cohorts, while the performance of the outside marker panels was either inconsistent or poor. The consistent and superior performance of the markers we identified could be attributed to minimizing genetic and environmental influences during the identification of microbial markers using HFDR as controls. On the other hand, it is worth noting that a patient with CD and the paired HFDR control share no more than 50% similarity in genetics and are not exposed to the same environmental confounders; thus, the results we obtained could still be biased. This was made clear by the superior diagnostic performance of the microbial marker with older populations and with populations with certain dietary habits.

To fully appreciate the significance of the performance of our diagnostic models, it is important to consider several major challenges posted by the validation cohort. Firstly, the microbial features were initially identified in the discovery cohort using 16S rRNA sequencing data with open-reference taxonomy assignment, while the validation cohort was analyzed by whole-genome sequencing with closed-reference taxonomy assignment, a method leading to the exclusion of a significant number of microbes in the validation cohort. Secondly, the discovery cohort used paired HFDRs as healthy controls, whereas the validation cohort employed unrelated healthy subjects who might have been influenced by genetic and environmental factors. Lastly, the discovery cohort consisted of Chinese patients and healthy subjects, while the validation cohort comprised American and European patients and subjects. Despite these challenges, the classification model performed satisfactorily in the validation cohorts, comparable to its performance in the discovery cohort. The outstanding performance observed suggests that the differential genera we identified (Faecalibacterium, Dorea, and Fusicatenibacter) are universal features among patients with CD of distinct cultural and geographical origins. Lower abundances of Faecalibacterium have also been observed in CD-A⁴⁹ and CD-R^13,32 by other investigators. Sokol et al. reported a lower representation of Faecalibacterium in active IBD, but not in patients with IBD in remission, likely due to a small sample size.¹⁴

The differential genera used for our classification model consist of SCFA-producing bacteria. In line with this, the KO genes related to SCFA production displayed a similar diagnostic potential for distinguishing CD from healthy controls. The satisfactory performance of the SCFA-related microbial features in distinguishing CD from healthy controls underscores the prevalence of diminished SCFA production capacity in the microbiota of both CD-R and CD-A, in support of their causal role in disease relapse and the perpetuation of gut inflammation. Thus, the microbial features we have identified hold promise as potential therapeutic targets. Future endeavors targeting SCFA-producing bacteria may pave the way for novel and effective therapeutics for both CD-R and CD-A.

In summary, by using HFDRs as controls to minimize the impact of genetic and environmental confounding factors, we have identified distinct microbial features in CD-A and CD-R. Multiple lines of evidence showed that the decreased abundances in the differential genera in CD-R (Faecalibacterium, Dorea, and Fusicatenibacter) are not a consequence of inflammation but rather potentially play a causal role in disease relapse through decreased production of SCFAs. These microbial features within CD-R displayed outstanding potential as non-invasive markers for distinguishing CD-R and CD-A from healthy controls across diverse populations with distinct geographical and cultural backgrounds.

Limitations of the study

Our study has several limitations. First, the current study utilized HFDRs in order to minimize genetic and environmental confounders. Although our approach is a step in the right direction, future studies could benefit from a more robust study design, albeit with increased efforts. For example, employing healthy twins as controls could help address genetic confounders, and adopting a prospective design could reduce biases and control for environmental factors like diet, medication, and lifestyle choices. Second, our findings demonstrate that microbial markers for CD-R are independent of inflammation and are associated with fecal SCFAs. Although these results suggest that dysbiosis in CD-R may be a causal factor for disease relapse, further studies incorporating prospective designs, intervention trials, and mechanistic investigations are imperative to provide robust evidence for this hypothesis.

Key resources table

Resource availability

Experimental model and study participant details

Method details

Quantification and statistical analysis