All the serum samples were acquired from Serum Biobank at Mayo clinic with approval from institutional review board. CD patients were randomly selected, followed by age and gender matched healthy controls and UC patients. The samples (100 CD, 100 UC and 100 controls) were divided evenly into two non-overlapping discovery and validation sets randomly (Table 1). Disease status for study participants was assessed by clinicians at Mayo clinic.

We analyzed 1570 microbial proteins, of which 1173 proteins were from 50 different species of bacteria, 397 proteins were from 33 different species of viruses and the remaining proteins were autoantigens. These proteins were selected from our large collection of microbial antigens (DNASU.org) with reference to our anti-microbial antibody studies on other diseases (unpublished data). Microbial protein arrays were fabricated as described earlier[15,16]. Briefly, plasmids with genes of interest cloned in the pANT7_cGST expression vector were obtained from the DNASU plasmid repository, prepared, and printed into silicon nanowells using a piezoelectric dispensing system to produce microbial protein arrays. On the day of experiment, proteins were freshly expressed from printed plasmids using an in-vitro transcription and translation protein expression kit (Fisher Scientific) and captured by anti-GST antibody co-printed in each nanowell. After expression, microarrays were incubated with 1:100 diluted serum samples. We randomized the case and control serum samples while profiling on microarray to reduce bias. IgG and IgA anti-microbial protein antibodies were detected by Alexa-647 goat anti-human IgG (H+L) and Cy3 goat anti-human IgA (Jackson ImmunoResearch). After washing and drying, the microarrays were scanned in a Tecan PowerScanner and the raw fluorescence intensity data were extracted using the ArrayPro Analyzer Software. Raw fluorescence intensity of each protein on the microarray was divided by the median intensity of all the proteins on the microarray for normalization. The normalized value was termed as Median Normalized Intensity (MNI) and used for all analysis. Seropositivity of antibody for a particular antigen was defined as MNI ≥ 2 as we have done for our other studies[17,18].

Pairwise comparisons of numbers of IgG or IgA antibodies for each bacterial species among the 3 subject groups were performed using Chi-squared tests to assess statistical significance (Supplementary Figure 1). For each pairwise comparison, the Chi-squared P values were adjusted using the FDR (false discovery rate) method to reduce the likelihood of false positives. In addition to the multiple comparison adjustment at the antibody level, we performed adjustment at the species level.

For univariate analysis between two comparison groups, we calculated sensitivity for one group at the 96th percentile of the other group or the MNI of 2, whichever was larger. Antibodies with ≥ 14% sensitivity in the discovery set were selected as candidates for further validation. If an antibody had ≥ 14% sensitivity at 96% specificity in both discovery and validation sets, then it was considered as a “validated marker”. Venn diagram for the overlap of microbial antigen targets were plotted using Venny[19].

We used a three-stage approach to build our multi-antibody panels. In the first stage, we selected all candidate biomarkers that passed the criteria above, i.e., sensitivity was greater or equal than 14% at 96% specificity. Next, we applied the minimum redundancy maximum relevance algorithm to further select biomarkers that were possibly the most important and least correlated[20]. In the third stage, we fit a logistic regression model using the selected biomarkers from the first two stages and generated its receiver operating characteristic curve and AUC value to evaluate the model’s discriminatory performance between CD, UC, and healthy controls.

Pair-wise subgroup comparisons based on the Montreal classification were performed for the odds ratio (OR) of each antibody using the seropositivity threshold defined as the maximum of MNI 2 and the 75^th percentile of all samples. Chi-squared tests were used to test global significance between all groups with a slight modification by adding 0.5 to each cell of the table to avoid zero cell counts. P values from the Chi-squared method were adjusted for each pair of comparisons and for all candidate biomarkers. The number of antibodies with significant difference in prevalence among classifications were counted based on OR > 1 and OR < 1 for each pair of classification of CD behavior, CD location, UC extent and the surgery history of CD patients. The difference in total number of antibodies with significant difference between classification groups were computed using two sample proportion test. We did not perform a subgroup analysis for the UC patients based on the surgery history because most (84 out of 100) had no surgeries (Table 1).

Spearman’s rank correlation analysis was performed to assess the correlation between autoantibody and anti-microbial antibody reactivity for CD patients and healthy controls. The R “pheatmap” package was used to generate the heatmap for correlation coefficients.

The NCBI Taxonomy browser was used to find the taxonomical details of all the bacteria and viruses used in our study. The taxa were downloaded as phylip tree file and was used as an input in interactive tree of life software. Two phylogenetic trees were created for bacteria and viruses with different colors distinguishing the phylum.

For sequence homology analysis, a pair-wise BLAST analysis was carried out on the antigen protein sequences of validated antibodies for CD vs healthy control analysis. E-values were used to generate a heatmap using Python Seaborn package.

We studied IgG and IgA anti-microbial antibody profiles of 100 CD and 100 UC patients and 100 age-gender matched healthy controls (Table 1) against 1570 microbial antigens including 1173 antigens from 50 different bacteria and 397 antigens from 33 different viruses using our protein microarray platform (Figure 1, Supplementary Table 1). This study provided a representative overview of the anti-microbial antibody response in IBD patients (Supplementary Figure 1). The numbers of IgG antibodies against bacterial proteins from Bacteroidetes vulgatus (B. vulgatus) and Citrobacter koseri (C. koseri) were significantly higher in CD patients compared with those in healthy controls (Chi-square test, P < 0.01) (Supplementary Figure 1). On the contrary, the numbers of IgG antibodies against proteins from several bacteria, such as Streptococcus pneumoniae (S. pneumoniae), Haemophilus influenza (H. influenzae), Staphylococcus aureus (S. aureus), Helicobacter pylori (H. pylori) and Parvimonas micra (P. micra) were significantly lower in CD and UC patients compared with those in healthy controls (Chi-square test, P < 0.05) (Supplementary Figure 1). Overall, fewer IgA anti-microbial antibodies were found than IgG antibodies. The numbers of IgA antibodies against S. pneumoniae, H. influenzae, S. aureus, and H. pylori were significantly lower in UC patients compared with those in healthy controls (Chi-square test, P < 0.01). On the other hand, anti-viral IgG and IgA antibodies showed heterogenous prevalence with no clear trend of differences among CD, UC, healthy controls (data not shown). Therefore, we focused our analysis on anti-bacterial antibodies.

Open in New Tab   Full Size Figure   Download Figure

Figure 1 Phylogenetic tree of microbes studied with their corresponding number of proteins analyzed. A: Fifty species of bacteria with 1173 proteins were segregated into 6 phyla; B: Thirty-three species of viruses with 397 proteins were segregated into 10 phyla.

We compared prevalence for individual anti-microbial antibodies between CD patients and healthy controls. We randomly split samples evenly into discovery and the validation sets (Table 1). For antibodies with elevated prevalence in CD patients, 13 IgG antibodies passed the criteria (sensitivity ≥ 14% at 96% specificity) in both discovery and validation sets (Table 2). Anti-A4-Fla2_IgG, a well-studied anti-bacterial flagellin antibody in CD, had the best performance with 47% sensitivity at 96% specificity in the full sample set (Table 2). Beside the flagellins, we found antibodies to four novel target antigens from B. vulgatus (BVU_0562), P. mirabilis (PMI_RS06815), S. flexneri (SF_Lpp) and S. pneumoniae (SP_1992) (Table 2) with no significant sequence homology to flagellins (Figure 2A).

Open in New Tab   Full Size Figure   Download Figure

Figure 2 Sequence homology of target antigens of validated antibodies and overlap of antibodies among Crohn’s disease and ulcerative colitis. A: Heatmap showing sequence homology among target antigens for antibodies with validated performance of ≥ 14% sensitivity at 96% specificity comparing Crohn’s disease (CD) patients with healthy controls; B: CD_IgG, ulcerative colitis (UC)_IgG, CD_IgA and UC_IgG represent the overlap of anti-microbial antibodies of IgG and IgA isotypes in CD and UC patients with ≥ 14% sensitivity at 96% specificity against healthy controls in the discovery set. CD: Crohn’s disease; UC: Ulcerative colitis.

To our surprise, 12 validated IgG antibodies showed elevated prevalence in healthy controls relative to CD patients (Supplementary Table 2). Among these 12 antibodies, anti-bacterial antibodies performed better in differentiating CD patients from healthy controls than anti-viral antibodies (Supplementary Table 2). Antibody against SPy_2009, an anchoring protein located in the cell wall of Streptococcus pyogenes (S. pyogenes), had the highest sensitivity of 24% at 96% specificity in healthy controls relative to CD patients. Seven validated IgA antibodies showed higher prevalence in healthy controls relative to CD patients (Supplementary Table 4).

For anti-microbial antibodies with elevated prevalence in UC patients relative to healthy controls, 4 IgG antibodies passed the criteria in both discovery and validation sets (Table 2). Antibodies to A4-Fla2_IgG and a flagellin from C. koseri had a sensitivity of 18% and 15% respectively. For IgG antibodies with higher prevalence in healthy controls relative to UC patients, 32 antibodies got validated (Supplementary Table 3). Source microorganisms for the target antigens of these 32 antibodies were enriched for S. pneumoniae, S. aureus, and H. influenzae (2-sample proportion test, P < 0.05). 2.7% of the proteins on the microbial protein microarray were from S. pneumoniae while 18.7% of antigens for validated antibodies were from S. pneumoniae, 6.1% of the proteins on the microarrays were from S. aureus while 18.7% of antigens for validated antibodies were from S. aureus, and 1.4% of the proteins on the microarrays were from H. influenzae while 12.5% of antigens for validated antibodies were from H. influenzae. Nine validated IgA antibodies showed higher prevalence in healthy controls relative to UC patients (Supplementary Table 4).

Fewer anti-viral antibodies than anti-bacterial antibodies were validated comparing CD or UC patients with healthy controls (Table 2, Supplementary Tables 2 and 3). Anti-viral antibodies to Rhinovirus B14, Enterovirus C, Influenza A virus, Human metapneumovirus had higher prevalence in healthy controls compared with CD and UC patients (Supplementary Tables 2 and 3).

We found 46 IgG and 22 IgA validated anti-microbial antibodies with higher prevalence in CD patients compared to UC patients while 28 IgG and 9 IgA validated anti-microbial antibodies with higher prevalence in UC patients compared to CD patients. There was minimal overlap of the target antigens of these validated IgG and IgA antibodies (Figure 2B).

We built multi-antibody panels that could distinguish CD vs control, UC vs control, and CD vs UC with an area under the curve (AUC) of 0.81, 0.87, and 0.82 respectively. For CD vs control, antibodies against novel flagellins (HP_0115, CK_LafA, CK_LafA.1, VC_flaD, VC_flaB, VC_flaE, VC_flaA) had an AUC of 0.73, antibodies against non-flagellins (BVU_0562, SP_1992, PMI_RS06815, and SF_Lpp) had an AUC of 0.75 and the combined AUC of antibodies against novel flagellins and non-flagellins was 0.81 (Figure 3A). For UC vs control, a combination of seven antibodies, four against S. pneumoniae and one each against S. aureus, H. influenzae and B. vulgatus had an AUC of 0.87 (Figure 3B). For CD vs UC, combination of seven antibodies, two against H. pylori and one each against E. coli, S. pneumoniae, S. pyogenes, C. jejuni and L. bacterium A4 had an AUC of 0.82 (Figure 3C).

Open in New Tab   Full Size Figure   Download Figure

Figure 3 Receiver operating characteristic curves to discriminate Crohn’s disease, ulcerative colitis, and healthy controls. A: Receiver operating characteristic (ROC) curve for Crohn’s disease (CD) vs healthy controls. Area under the curve (AUC) values of novel anti-flagellin antibodies (HP_0115, CK_LafA, CK_LafA.1, VC_flaD, VC_flaB, VC_flaE, VC_flaA) and anti-non-flagellin antibodies (BVU_0562, SP_1992, PMI_RS06815, SF_Lpp) was 0.73 and 0.75, respectively. The AUC value obtained with a combination of novel anti-flagellin and anti-non-flagellin antibodies was 0.81; B: ROC curve for ulcerative colitis (UC) vs healthy controls. The AUC value obtained with a combination of 7 markers was 0.87; C: ROC curve for CD vs UC. The AUC value obtained with a combination of 7 markers was 0.82.

We investigated the association of CD behavior (B1, B2, B3), CD location (L1, L2, L3), and UC extent (E1, E2, E3) based on the Montreal classification with the anti-microbial antibody prevalence. We calculated the 4^th quartile odds ratio for each antibody between the two classification groups and compared the number of antibodies with significant odds ratio (P value < 0.05) in each group. We found B3 (penetrating) had the highest prevalence of antibodies followed by B2 (stricturing) and B1 (non-stricturing, non-penetrating) (Table 3). For CD location, we found L2 had the highest prevalence of antibodies followed by L3 (ileocolonic) and L1 (Table 3). For UC extent, E3 (extensive UC) had higher prevalence of antibodies compared to E2 (left sided UC). In addition to the Montreal classification, we also performed subgroup analysis based on the surgery history of CD patients. We found that patients who had surgery possessed higher prevalence of antibodies compared to those without surgery (Table 3).

We previously reported novel autoantibodies in CD patients using the same set of CD patients and healthy controls[21]. We profiled both IgG and IgA autoantibodies and anti-microbial antibodies in all 100 CD and 100 healthy controls. It is interesting to note the antibodies showing differences for autoantibodies were mostly IgA, but the anti-microbial antibodies were mostly IgG[21]. Anti-SNRPB_IgA had the highest sensitivity of 20% at 96% specificity among all autoantibodies compared with 47% sensitivity at 96% specificity for the best performing anti-microbial antibody, anti-A4-Fla2_IgG.

We compared the novel autoantibodies and validated anti-microbial antibody profiles to determine if correlation existed between their reactivity. Overall, we did not observe high correlation between autoantibodies and anti-microbial antibodies in CD patients (Figure 4). Anti-microbial antibodies formed two clusters, one with anti-flagellin antibodies, and the other with SF_Lpp_IgG and PMI_RS06815_IgG. Five autoantibodies, PRPH_IgA, SNAI1_IgA, PPP1R13L_IgA, SNRPB_IgA and PTTG1_IgA, formed a cluster. The remaining antibodies had relatively unique reactivity patterns.

Open in New Tab   Full Size Figure   Download Figure

Figure 4 Spearman’s rank correlation coefficient heatmap of anti-microbial antibodies and autoantibodies in Crohn’s disease patients. The names of anti-microbial antibodies are colored in blue while autoantibodies are colored in black.

Early and accurate diagnosis of inflammatory bowel disease (IBD) can benefit the clinical management of IBD patients. The clinical utility of available non-invasive biomarkers is limited.

The gut microbiota plays an important role in the pathogenesis of IBD. Antibody responses to microbial antigens can be exploited to identify better diagnosis markers and improve our understanding of IBD pathology.

This study aimed to compare anti-microbial antibody profiles in Crohn’s disease (CD) and ulcerative colitis (UC) patients and healthy controls.

We employed an innovative protein microarray platform and profiled antibodies against 1570 microbial antigens in 100 CD, 100 UC, and 100 healthy controls.

Antibodies to bacterial proteins were better in distinguishing IBD patients from healthy controls compared with antibodies to viral proteins. We identified a set of novel anti-microbial antibodies against the antigens of Bacteroidetes vulgatus (BVU_0562) and Streptococcus pneumoniae (SP_1992) elevated in CD patients relative to healthy controls. In addition, anti-microbial antibodies against the antigen of Streptococcus pyogenes (SPy_2009) were found to be elevated in healthy controls relative to UC patients. We constructed antibody panels that could distinguish CD vs control, UC vs control, and CD vs UC with an AUC of 0.81, 0.87 and 0.82 respectively. Patients with severe disease had higher prevalence of anti-microbial antibodies. There was minimal correlation among the occurrence of autoantibodies and anti-microbial antibodies in CD patients.

This study discovered novel anti-microbial antibodies with differential prevalence in CD and UC patients relative to healthy controls. In addition, this study revealed the potentially different roles of gut microbiota in CD and UC pathology.

Our study demonstrated the power of a microbiomics approach to identify biomarkers that could aid in the early and accurate detection of IBD non-invasively, and shed light into the role of various microbes in IBD etiology.