Metagenome-wide association study of gut microbiome revealed potential microbial marker set for diagnosis of pediatric myasthenia gravis

We recruited a discovery cohort of 53 pediatric MG patients and 46 HC that were matched for age and gender; this study was conducted at the Center of Treatment of Myasthenia Gravis Hebei Province, People’s Hospital of Shijiazhuang. Clinical, grading of all patients was class I (any ocular muscle weakness; may have weakness of eye closure; all other muscle strength is normal.), based on the Myasthenia Gravis Foundation of America (MGFA) clinical classification [18]. The diagnosis of pediatric MG was based on the clinical presentation of the patients and was confirmed with at least one diagnostic test, including positive antibodies (AChRAb), electromyography, or fatigue test and response to a therapeutic trial (Additional file 4: Table s-1). We further recruited 19 pediatric MG patients as the validation cohort based on the same criteria. Participants were excluded if they had one of these conditions: (1) age < 2 years and 10 months, or no age information; (2) antibiotic usage (except β-lactam) within 3 months of the study; (3) received any MG-related treatments (e.g., pyridostigmine, glucocorticoid, human immunoglobulin, or Chinese herbal medicine); (4) using of any drugs or unknown status; (5) with a family history or onset after birth; and (6) complicated with other diseases. The study was reviewed and approved by the ethics committee of the People’s Hospital of Shijiazhuang (NO. 66). Informed consent was obtained from all recruited patients’ parents and healthy individuals’ parents.

Approximately 2 g of a fresh fecal sample was collected in a Fecal collection tube (OMR-200 DNA Genotek) and stored at room temperature until DNA extraction (21–28 days). The instruction of OMR-200 DNA Genotek showed the samples could store at room temperature for 60 days (https://​www.​dnagenotek.​com/​us/​products/​collection-microbiome/​omnigene-gut/​OMR-200.​html). Samples were collected before patients were treated.

The DNA extraction and library construction are followed the instruction of TruSeq DNA Nano Reference Guide (1000000040135) (https://​sapac.​support.​illumina.​com/​downloads/​truseq-dna-nano-reference-guide-1000000040135.​html) and Hiseq 2500 System Guide (15035786) (https://​support.​illumina.​com/​downloads/​hiseq_​2500_​user_​guide_​15035786.​html) from Illumina. Sterile plastic cupDNA concentration was assessed using a fluorometer reader (Qubit Fluorometer, Invitrogen). DNA sample integrity and purity were confirmed with agarose gel electrophoresis (concentration of agarose gel: 1%, voltage: 150 V, electrophoresis time: 40 min). One microgram genomic DNA from each sample was randomly fragmented by Covaris (LE220), and sequences of an average size of 200–400 bp were selected by magnetic beads (Ampure XP). The fragments were end repaired and 3′-adenylated, and adaptors were ligated to their ends. This process was performed to amplify fragments with the adaptors from the previous step. PCR products were purified using magnetic beads. The double-stranded PCR products were heat denatured and circularized by the splint oligo sequence. The single-strand circular DNA (ssCir DNA) was formatted as the final library. The library was amplified with the phi29 DNA polymerase to make DNA nanoballs (DNBs), which were loaded into the patterned nanoarray. Paired-end 100/150 bp reads were generated via the combinatorial Probe-Anchor Synthesis (cPAS). After library construction, the samples were sequenced on the Illumina HiSeq 2500 platform.

We performed quality control of the sequencing data according to the following steps before analyses: (1) filtered low-quality reads and (2) removed contamination of the human genome sequence. We used fastp (Version: 0.21.1) [19] with its default parameters to screen out low-quality reads and sequence adapter sequences, aligned the reads to the human genome (hg38) with bowtie2 (Version: 2.3.5) [20], and used samtools (Version: 1.9) to screen out paired reads that did not align to the human genome. Clean data were used in subsequent analyses.

We used metaphlan2 (Version: 2.1.0) to map high-quality reads to the mpa_v20 marker gene database according to the metagenomics data analysis pipeline of The Huttenhower Lab; this enabled us to obtain microbial taxonomy abundance profiles at different taxonomic levels for each sample. We used the utils of metaphlan2 merge_metaphlan_tables.py to combine the results of all samples and utilized an in-house script to generate a combined profiling table for different taxonomic levels. Humann2 (Version: 2.0) was used to map high-quality reads to uniref90 and chocophlan (humann2 recommended databases) for profiling of gene family abundance and pathway abundance. We subsequently used the humann2 utils: humann2_join_table(s), humann2_renorm_table, and humann2_split_stratified_table to merge the abundance of all samples, normalize the abundance, and to stratify profiles with taxonomy annotations, respectively.

For the calculation of the Shannon index, we used in-house scripts to calculate the α-diversity of each sample using the taxonomy abundance data. Using the same input data, the vegdist method of the Vegan package (Version: 2.5-5) in R (with the parameter “method = dist_method”) was used to calculate β-diversity.

To explore relationships among samples, we performed principal component analysis (PCA) and principal coordinate analysis (PCoA) with the taxonomy abundance data using the Ade4 package (Version: 1.7-15) of R. We use the corr.test of the R package with “method = spearman, use = pairwise, adjust = BH” to calculate the Spearman’s rank correlation between the clinical phenotypes and significantly different microbial species and functions.

Significantly different microorganisms between MG patients and HC at all taxonomy levels were assigned as candidate microbial markers. Using the relative abundance of the candidate markers at each taxonomy level, we used a random forest model to build classifiers to distinguish between HC and MG patients with RandomForest package (Version: 4.6-14) in R. The samples we used in this step were labeled with the two-value disease situation. We applied the package in classification with default parameters, like the number of trees grow in the process is 500. Subsequently, a fivefold cross-validation method was used to evaluate the performance of the predictive model. We calculated the minimum error in the average cross-validation error curve and the average curve and used the standard deviation of this point as the cut-off point of the filter prediction model. The set of candidate markers, which contained the minimum number of candidate markers in all groups and had an error below the critical value, was selected as a marker set to build the diagnostic classifier. The probability of MG was then calculated based on this optimal set; the receiver operating characteristics (ROC) of the discovery and validation cohort were plotted using the pROC package (Version: 1.16.2).

Approximately 2 ml of fresh peripheral blood of MG patients and HC individuals were collected in the early morning before meals. Samples were collected before patients were treated. These samples were centrifuged at 3000 rpm for 10 min after being left at room temperature for 30 min; the serum was collected and stored at − 80 °C until further analysis. The SCFAs quantification protocol was referenced from Sandin et al., Vemuri et al., and Sivaprakasam et al. [21‐23].

The standard SCFAs solution was prepared as following: 1 mg/ml SCFAs was diluted with ultrapool water into 1 ml solution with concentrations of 1 μg/ml, 2 μg/ml, 5 μg/ml, 10 μg/ml, 50 μg/ml, 100 μg/ml, and 200 μg/ml, respectively. Take ≥ 50 μL fluid, add 400 μL saturated sodium chloride solution and 50 μL saturated hydrochloric acid sodium chloride solution with 3 mmol, vibrate it to dissolve completely, and use low-temperature ultrasound for 20 min. Then, the standard curve was drawn with the concentration of standard SCFAs (μg/ml) as the horizontal coordinate and the peak area as the vertical coordinate.

To prepare the testing solution, 50 g samples were mixed with 400 μL saturated sodium chloride solution and 50 μL with 3 mmol saturated hydrochloric acid sodium chloride solution. The fully dissolved mixed solution was carried on ice bath ultrasound for 20 min. After the ultrasound, 400 μL ice ether was added and the mixed solution was shaken for 10 min to fully extract. The solution was then centrifuged for 10 min at 12000 rpm at 4 °C. Collect the supernatant and add 50 mg anhydrous sodium sulfate to oscillate for 3 min and centrifuge for 5 min at 4500 rpm at 4 °C. The supernatant was used in the followed GC-MS analyses.

The analytical instrument used in testings was Thermo Trace1300-Thermo TSQ9000 GC/MS. SIM mode was used for data acquisition, and Tracefinder (Thermo Fisher Scientific, Waltham, MA, USA) was used for data processing.

The discovery cohort comprised 53 pediatric MG patients and 46 age- and gender-matched HC. The clinical information of all patients is given in Additional file 4: Table s-2. In the discovery cohort, the patients and healthy individuals showed no significant difference in clinical indices; there were significantly more female individuals than males in the cohort (Table 1, Wilcoxon test, BH-adjusted p < 0.05).

We initially focused on the metagenomic diversity of the MG patients and HC. After quality control filtering of the sequences (Additional file 4: Table s-3), we aligned high-quality reads to the mpa v20 maker gene database of MetaPhlAn2 [24] and obtained the species abundance data for the samples. We used this species abundance profiling data to conduct diversity analysis and detected no differences between patients and HC in alpha diversity (based on Shannon index results at different taxonomic levels) or beta diversity (Additional file 1: Figure s-1). Exploratory PCA and PCoA analyses using the profiling data of different taxonomic levels, from phylum to species, suggested no obvious separation or clustering (Additional file 2: Figure s-2).

We next applied differential tests for abundance profiling of MG and HC groups across six taxonomic levels. At the phylum level, Deinococcus–Thermus and Synergistetes were enriched in the HC group, while Cyanobacteria and Viruses_noname were enriched in the MG group (Wilcoxon test, BH adjusted p < 0.05) (Fig. 1A). At the genera level, six genera were enriched in the MG group and five genera were enriched in the HC group (Additional file 4: Table s-4).

At the species level, 16 species were enriched in the MG group and nine species were enriched in the HC group (Fig. 1B). Notably, Prevotella copri (P. copri), which was enriched in the MG patients, has been previously reported as an enriched species in the gut microbiomes of rheumatoid arthritis (RA) patients [25].

Besides bacteria, one virus was also enriched in the MG group. We detected human adenovirus in the microbiomes of 10 MG patients (total 53 MG patients). Nine of them were human adenovirus F and one was human adenovirus D. No healthy individuals contained viral sequences in the data. Since the database used in taxonomy annotation is a marker gene database, we aligned the clean data with the reference genome human adenovirus D (NC_010956.1) and human adenovirus F (NC_001454.1) data from NCBI and confirmed that the viral genome was present in the gut microbiome of these 10 MG patients, with over two average alignment depth of adenovirus reference genomes (Additional file 4: Table s-5). We detected no mapping reads in other samples.

We detected 12,803 significantly different stratified gene families between the discovery cohort MG patients and the HC group. Using Canopy-based algorithm clustering, 238 co-abundance gene groups (CAGs) were obtained [26]. For each CAG, if 80% of the included gene families were from the same species, the species annotation of this CAG was assigned to this species; otherwise, the annotation was unclassified. We constructed an interaction network for the CAGs based on Pearson correlation values (Fig. 2). The CAGs enriched in the HC group belonged to C. bartlettii, Bilophila wadsworthia, and Bacteroides dorei. The CAGs enriched in the MG group mainly belonged to P. copri and Bacteroides massiliensis.

We next explored whether the gut microbiome profiling data has the potential to distinguish pediatric MG patients from HC, by constructing a disease classifier using a random forest model and the data for the significantly different species (Additional file 3: Figure s-3).

In order to evaluate the accuracy of this classification model, an independent cohort consisting of 19 MG patients without HC samples was used as the validation cohort. To ensure the same proportion of the two types of samples in the training set and the test set, we randomly selected 12 HC samples from the discovery cohort and 19 MG patients in the validation cohort to form the test set, and the remaining 53 MG patients and 34 HC samples were used as the training set.

We used this training set to build a disease classifier using the 25 different species obtained in the above-mentioned differential tests; five species were selected as candidate microbial markers, and a classification model was established using a fivefold cross-validation method (Fig. 3A–C). The area under the receiver operating curve (AUC) of the corresponding ROC curve reached 0.94. All five marker species were enriched in the MG group. The bacterial species used for classification include Fusobacterium mortiferum (F. mortiferum), P. stercorea, P. copri, M. funiformis, and M. hypermegale. The highest cross-validation average accuracy was found when the markers contained P. stercorea and P. copri.

The test set achieved an AUC value of 0.84 (Fig. 3D, E). Detailed predicted probability of MG in the discovery cohorts and validation cohorts is listed in Additional file 4: Table s-6. These results suggest that MG-related microbial markers may have the potential for use in non-invasive diagnostic strategies for MG in children.

Based on the AChRAb levels in the blood, we divided the MG patients into AChRAb-positive (n = 34, AChRAb > 0.5 nmol/L) and AChRAb-negative (n = 19) groups (Additional file 4: Table s-1). There were no significant differences in demographic of these two (Additional file 4: Table s-7). We then assessed these two MG groups to explore relationships between AChRAb and the gut microbiomes of the MG patients (Fig. 4).

We first compared the taxonomic abundance data of AChRAb-negative and AChRAb-positive groups with HC at the species level and found that P. stercorea was the only significantly different species enriched in the AChRAb-negative patients. However, the abundance of 14 species was significantly different between HC and AChRAb-positive MG patients. The 14 species contained P. stercorea, which is a candidate microbial marker species identified in the previous experiment. Moreover, all 14 species were significantly different in abundance in the HC and MG cohorts.

We further compared the abundance of unstratified pathways in HC with the AChRAb-positive and AChRAb-negative MG patients. Of the 16 unstratified pathways with significant differences in abundance between HCs and AChRAb-negative MG patients (Additional file 4: Table s-8), four are related to SCFAs; “ANAGLYCOLYSIS-PWY: glycolysis III (from glucose)” and “PWY-6435: 4-hydroxybenzoate biosynthesis V” were enriched in the HC group, whereas “PWY-7003: glycerol degradation to butanol” and “PWY-6588: pyruvate fermentation to acetone” were enriched in the AChRAb-negative MG patients.

ANAGLYCOLYSIS-PWY: glycolysis III (from glucose) is the main pathway for degrading d-glucopyranose to pyruvate. PWY-6435: 4-hydroxybenzoate biosynthesis V has an intermediate step, with CoA as the enzyme to produce acetyl-CoA, while generation of acetate from pyruvate via acetyl-CoA is the most common pathway of enteric bacteria [27, 28]. The AChRAb-negative enriched pathways are consuming pyruvate. PWY-6588 is pyruvate fermentation pathway, which PWY-7003 will use pyruvate to produce butan-1-ol.

Comparison of HC and AChRAb-positive MG patients revealed 15 significantly different unstratified pathways (Additional file 4: Table s-9): Ten pathways were enriched in AChRAb-positive MG patients and five pathways were enriched in HCs. Four of these pathways were related to SCFAs: COA-PWY: coenzyme A biosynthesis I, OANTIGEN-PWY: O-antigen building blocks biosynthesis (E. coli), and PANTOSYN-PWY: pantothenate and coenzyme A biosynthesis I, which were all enriched in AChRAb-positive MG patients, and PWY-6435 was enriched in HCs. Both COA-PWY and PANTOSYN-PWY synthesize CoA, and OANTIGEN-PWY contains a catalysis step by acetyl-CoA and reduces CoA. Due to the competitive relationship, CoA inhibits the enzymatic reactions of the acetoacetyl-CoA transferase, which produces acetate [29, 30], while the HC-enriched pathway PWY-6435 converts CoA to acetyl-CoA.

To verify the relationship between MG gut microbiome and SCFAs, we tested the SCFAs quantity in blood samples from the discovery cohort (Additional file 4: Table s-10). This revealed significantly different levels of butyric acid and isobutyric acid between HCs and MG patients (Fig. 5A, B). We further calculated the Pearson correlation between species abundance (species level) and butyric acid and isobutyric acid content. This revealed 13 and 16 species exhibiting significant correlations with blood butyric acid and isobutyric acid content, respectively. The abundance of Lactobacillus sanfranciscensis and Prevotella nanceiensis is significantly positively correlated with the blood butyric (Fig. 5C, D), and Bacteroides dorei, Erysipelotrichaceae bacterium 3_1_53, and Eubacteriaceae bacterium ACC19a are significantly positively correlated species to blood isobutyric acid (Fig. 5E–G). The abundance of those species is different between HC and MG that we identified above (p before adjusted < 0.05). While Erysipelotrichaceae bacterium 3_1_53 is the only one enriched in MG, the other four species were enriched in HC.