Gut Microbiome–Wide Search for Bacterial Azoreductases Reveals Potentially Uncharacterized Azoreductases Encoded in the Human Gut Microbiome

Abstract

The human gut is home to trillions of microorganisms that are responsible for the modification of many orally administered drugs, leading to a wide range of therapeutic outcomes. Prodrugs bearing an azo bond are designed to treat inflammatory bowel disease and colorectal cancer via microbial azo reduction, allowing for topical application of therapeutic moieties to the diseased tissue in the intestines. Despite the inextricable link between microbial azo reduction and the efficacy of azo prodrugs, the prevalence, abundance, and distribution of azoreductases have not been systematically examined across the gut microbiome. Here, we curated and clustered amino acid sequences of experimentally confirmed bacterial azoreductases and conducted a hidden Markov model–driven homolog search for these enzymes across 4644 genome sequences present in the representative Unified Human Gastrointestinal Genomes collection. We identified 1958 putative azo-reducing species, corroborating previous findings that azo reduction appears to be a ubiquitous function of the gut microbiome. However, through a systematic comparison of predicted and confirmed azo-reducing strains, we hypothesize the presence of uncharacterized azoreductases in 25 prominent strains of the human gut microbiome. Finally, we confirmed the azo reduction of Acid Orange 7 by multiple strains of Fusobacterium nucleatum, Bacteroides fragilis, and Clostridium clostridioforme. Together, these results suggest the presence and activity of many uncharacterized azoreductases in the human gut microbiome and motivate future studies aimed at characterizing azoreductase genes in prominent members of the human gut microbiome.

SIGNIFICANCE STATEMENT

This work systematically examined the prevalence, abundance, and distribution of azoreductases across the healthy and inflammatory bowel disease human gut microbiome, revealing potentially uncharacterized azoreductase genes. It also confirmed the reduction of Acid Orange 7 by strains of Fusobacterium nucleatum, Bacteroides fragilis, and Clostridium clostridioforme.

Introduction

Orally administered drugs are an attractive, noninvasive mode of delivery of pharmaceuticals to the intestines. The human gut microbiome plays an important role in drug metabolism (Spanogiannopoulos et al., 2016) and is capable of activating (Peppercorn and Goldman, 1972; Morrison et al., 2012; Sousa et al., 2014), inactivating (Peters, 1978; Saha et al., 1983; Haiser et al., 2013), and even toxifying (Wallace et al., 2010) pharmaceutical drugs. Prodrugs containing an azo bond actually require bacterial azoreductase activity to release biologically active compounds (Peppercorn and Goldman, 1972). For conditions such as inflammatory bowel disease (IBD) and colorectal cancer (CRC), bacterial azoreductases have been used to deliver therapeutics such as 5-aminosalicylic acid (5-ASA), prednisolone (Ruiz et al., 2011), and celecoxib (Ruiz et al., 2011) topically to diseased intestinal tissues (Fig. 1). Following oral administration of sulfasalazine, bacterial azoreductases in the gut reduce azo bonds, liberating 5-ASA and allowing it to confer its anti-inflammatory properties (Mahida et al., 1991; Rachmilewitz et al., 1992; Weber et al., 2000) topically on inflamed intestinal tissue. Direct oral administration of 5-ASA is nonoptimal because the majority of the drug is absorbed in the small intestine and is sent through systemic circulation (Peppercorn and Goldman, 1973; Tozaki et al., 2002; Friend, 2005; Perrotta et al., 2015; Foppoli et al., 2019). Other examples of azo-bonded prodrugs are OPN501 and celecoxib-5-ASA. OPN501 is made of up prednisolone, 5-ASA, and an inert cyclization product, dihydroquinolone (DHQ). Upon azo reduction, 5-ASA is released and is able to act topically upon the target tissue. Following a spontaneous cyclization reaction, prednisolone and DHQ are released where prednisolone can act upon the target tissue (Ruiz et al., 2011). Celecoxib-5-ASA is made up of celecoxib, 5-ASA, and DHQ, which are all released upon azo reduction and cyclization in a similar mechanistic fashion to OPN501 (Ruiz et al., 2011).

Fig. 1.

Azo reduction by gut microbiota. (A) Pathways of azo-bonded drug activation by bacterial azoreductase activity. DHQ is produced via cyclization of an intermediate of OPN501 and celecoxib-5-ASA prodrug following intramolecular lactamization (Ruiz et al., 2011). (B) Description of downstream metabolites of bacterial azo reduction and the mechanisms of action in IBD and CRC. References for each molecular function described in this subfigure: 5-ASA (Mahida et al., 1991; Cominelli et al., 1992; Rachmilewitz et al., 1992), prednisolone (Cohen et al., 2000), celecoxib (Wu et al., 2003, ; Wu et al., 2010; Gustafsson et al., 2010), sulfapyridine (Nielsen, 1982), and DHQ (Ruiz et al., 2011). (C) Presence of azoreductase-containing bacteria is required for prodrug activation in the IBD gut and CRC gut (left). Many azo-reducing bacteria have been characterized; however, some species have shown experimental evidence of azo-reduction without the full characterization of the genes responsible for azo reduction (right).

A recent review article by Suzuki (2019) collected and curated experimentally confirmed azoreductases and described their preferred flavin cofactors and electron donors. They found that bacterial azoreductases qualitatively cluster into four main clades, which harbor a preference for either flavin mononucleotide (FMN) or flavin adenine dinucleotide as the flavin cofactor and either NADH or NADPH as the preferred electron donor. Clades I, II, and III are flavoproteins, whereas clade IV proteins are flavin free. Clade I azoreductases prefer NADPH as the electron donor, clade II prefers NADH as its electron donor, and clades III and IV generally use both. Of the 37 enzymes examined by Suzuki (2019), eight enzymes formed no distinct phylogenetic clade and featured differences in primary sequence length, flavin cofactor, and preferred electron donor. In addition to their relevance in drug delivery and efficacy, azoreductases are involved in nitroreduction (Brown 1981; Rafii et al., 1990; Rafii and Cerniglia, 1995; Liu et al., 2007; Mercier et al., 2013; Chalansonnet et al., 2017), quinone oxidoreduction (Liu et al., 2008; Leelakriangsak et al., 2008; Ryan et al., 2010a; Ryan et al., 2010b; Ryan et al., 2014), and azo dye reduction. Azo dyes such as Allura Red and Brilliant Black are commonly used in the food and textile industries, and waste from their production and usage pollutes the environment. This has led to a plethora of manuscripts characterizing the activity of azoreductases across the bacterial kingdom (Cerniglia et al., 1982; Zimmermann et al., 1982; Nakanishi et al., 2001; Suzuki et al., 2001; Blümel et al., 2002; Blümel and Stolz, 2003; Chen et al., 2004; Chen et al., 2005; Nachiyar and Rajakumar, 2005; Ooi et al., 2007; Matsumoto et al., 2010; Misal et al., 2011; Feng et al., 2012; Gonçalves et al., 2013; Lang et al., 2013; Misal et al., 2014; Eslami et al., 2016; Zhang et al., 2016), many of which exhibit nonnegligible sequence similarity to gut microbial azoreductases (Suzuki 2019).

There is a growing body of literature suggesting that azo reduction is a ubiquitous function of the human gut microbiome (Javdan et al., 2020), with many prominent bacterial strains showing significant reduction of sulfasalazine in vitro (Zimmermann et al., 2019). Javdan et al. (2020) showed that among 20 different individuals, sulfasalazine reduction was one of the only ubiquitous functions of the gut microbiome. Zimmermann et al. (2019) tested the reduction of sulfasalazine by 76 prominent strains of the gut microbiome and reported a significant [false discover rate (FDR) adjusted P value < 0.05] reduction of sulfasalazine by 62 of these strains. Interestingly, some of these experimentally confirmed azo-reducing strains reported by Zimmermann et al. (2019) have no prior evidence of azo-reduction and, thus, may encode novel or uncharacterized azoreductase genes (Fig. 1C). Identification of known azoreductases in newly reported sulfasalazine-reducing species can help narrow down strains to target for identification of uncharacterized azoreductases.

The prevalence, abundance, expression, and distribution of azoreductase enzymes in the human gut microbiome have implications for the efficacy of existing prodrugs mentioned above, as well as for the development of future azo prodrugs. Although azoreductases have been identified and characterized in many gut bacteria (Supplemental Table 1), the distribution of azoreductases has not been systematically explored across current gut bacterial reference genomes. To address this gap, we conducted a homolog search of known azoreductases across the Unified Human Gastrointestinal Genomes (UHGG) collection (Almeida et al., 2020) to identify putative azoreductases and azo-reducing species in the human gut microbiome. We then assessed the relative abundance and expression of known azoreductases in healthy, ulcerative colitis (UC), and Crohn disease (CD) participants of the Integrative Human Microbiome Project (HMP2) (Proctor and Huttenhower, 2019), the Prospective Registry of IBD Patients at MGH (PRISM) (Franzosa et al., 2019), and the Health Professionals Follow-Up Study (HPFS) (Abu-Ali et al., 2018). Finally, we tested the in vitro azo reduction of Acid Orange 7 by three strains of Fusobacterium nucleatum along with two strains of Bacteroides fragilis and two strains of Clostridium clostridioforme.

Materials and Methods

Description of Publicly Available Shotgun Metagenomic Sequencing Data

Shotgun metagenomic sequencing data obtained from the HMP2 [N (samples) = 703, N (individuals) = 103] (Proctor and Huttenhower, 2019), the PRISM [N (samples) = 218, N (individuals) = 218] (Franzosa et al., 2019), and the HPFS [N (samples) = 220, N (individuals) = 220] (Abu-Ali et al., 2018) are used throughout this work. Note that all samples referred to throughout this work are human stool samples.

Curation of Hidden Markov Models Representing Azoreductase Enzymes

We searched the literature for known and experimentally validated species of bacteria that have azoreductase activity. The list of gene sequences collected, along with the relevant metadata (organism, functional annotation, length, etc.) is available in Supplemental Table 1. Preliminary evidence for azoreductase gene sequence clustering is shown in Suzuki (2019), where sequences were aligned and phylogenetically compared. Next, after collecting 40 sequences of experimentally validated azoreductases, we generated a sequence similarity network using the Enzyme Function Initiative-Enzyme Similarity Tool (EFI-EST) (Gerlt et al., 2015) at a 35% amino acid sequence identity threshold for identifying similar clusters of azoreductase genes. This threshold corroborates the preliminary evidence for a diversity of azoreductase sequences put forth by Suzuki (2019). With the exception of clade IVb sequences, which reached 31% sequence identity, all other clusters of genes had at least 35% sequence identity. The groups shown in Fig. 2 were pressed into profile hidden Markov models (HMMs) using HMMER v​​ersion 3.1b2 (Finn et al., 2015). Other genes that did not fall into the clade I–clade IV clusters (arsH, yieF, mdaB, azo1, azoR, etc.) were pressed into singular HMMs and included in the homolog search.

Fig. 2.

Bacterial azoreductases cluster by cofactor and electron donor preferences. Following an extensive literature search for experimentally confirmed bacterial azoreductases, amino acid sequences of 40 azoreductase enzymes were collected and clustered with EFI-EST (Gerlt et al., 2015) at 35% sequence identity. Each node in the figure above is a single azoreductase gene, and the edges between nodes indicate at least 35% sequence identity between the two amino acid sequences. The colored clusters, clade I through clade IVa and IVb, are groups of azoreductases previously described by Suzuki (2019) as mechanistically similar groups based on cofactor and electron donor preferences. Clusters labeled with gene names (mdaB, yieF, etc.) represent homologous gene sequences found in two or more organisms. Each mechanistically characterized group of azoreductases were subsequently pressed into profile HMMs using HMMER v3.1b2 (Finn et al., 2015), which formed the basis of the homolog search. The group labeled “other azoreductases” contains sequences that did not fall into any cluster at the 35% identity threshold and were pressed into singular HMMs prior to the homolog search.

Search for Azoreductase Genes across Human Gut Microbial Genomes

We searched HMMs of known, experimentally validated azoreductases across 4644 nonredundant genomes contained in the UHGG collection (Almeida et al., 2020) using HMMER v3.1b2 (Finn et al., 2015). Alignments to queried HMMs with E-value < 1 × 10⁻¹⁰ and 60% coverage of the query sequence were labeled as putative azoreductase gene sequences. Putative azo-reducing bacterial species with experimentally confirmed azoreductase activity (Supplemental Table 2) were then labeled as “known” azo-reducing species and classified separately from the putative azo-reducing species. Putative azo-reducing species across the bacterial taxonomy were visualized using the iTOL web interface (Letunic and Bork, 2019), and prominent phyla of the gut microbiome were subsetted and presented in Fig. 3.

Fig. 3.

Azoreductases are widely distributed across gut bacterial taxonomy. Presence/absence of azoreductases across prominent phyla of the human gut microbiome. The taxonomic tree is obtained from the UHGG (Almeida et al., 2020) which is built on the Genome Taxonomy Database (Chaumeil et al., 2019). Phyla names are annotated on the left side. Phyla names followed by a capital letter, e.g., Firmicutes (A), indicate a novel phylum classified by the Genome Taxonomy Database toolkit. The bar chart in the center indicates the number of species contained in each genus shown in the tree. The size of the circles indicates the number of species that contain hits to the azoreductase genes specified. The color of the circles indicates the cofactor and preferred electron (e^-) donor of the enzyme.

Relative Abundance and Azoreductase Gene Abundance and Expression Estimation

Raw sequencing reads for samples from HMP2 (Proctor and Huttenhower, 2019), PRISM (Franzosa et al., 2019), and HPFS (Abu-Ali et al., 2018) were downloaded and extracted with the National Center for Biotechnology Information’s SRA toolkit v2.10.9 (http://ncbi.github.io/sra-tools/). Quality control and adapter trimming of the fastq sequence files were done with the Trim Galore wrapper v0.6.6 (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/). To remove potential human contaminants, quality trimmed reads were screened against the human genome (hg19) with Bowtie2 v2.4.2 (Langmead and Salzberg, 2012). Putative azoreductase sequences were extracted from UHGG genomes via custom shell and python scripts. Putative azoreductase gene sequences (HMP2, PRISM) and expression levels (HPFS) were quantified using salmon v1.4.0 (Patro et al., 2017) and were normalized and aggregated in R v4.1.1 and were subsequently visualized using the R package ggplot2 (Wickham, 2011) (Fig. 4). Taxonomy profiling of the cleaned metagenomic reads from HMP2 samples was performed using Kraken 2 v2.0.8-beta (Wood et al., 2019) to estimate the relative abundance of bacterial species present in each dataset. These relative abundances were then processed in R v4.1.1 and plotted using ggplot2 (Fig. 5). All computational and bioinformatic procedures are open source and are provided at https://github.com/dombraccia/Azoreductases.

Fig. 4.

Abundance and expression of azoreductase genes by human gut microbiota. (A and B) Visualization of shotgun metagenomic sequencing data from the Human Microbiome Project 2, also known as the HMP2 and the PRISM. We used salmon 1.4.0 (Patro et al., 2017) to quantify the abundance of azoreductase genes from hundreds of stool samples across healthy controls (non-IBD), UC, and CD participant cohorts. Raw DNA read alignment counts were normalized to counts per million (CPM), analogous to transcripts-per-million (TPM) normalization. Asterisks above each boxplot indicate statistical significance (*P < 0.05; **P < 0.01; ***P < 0.001, Wilcoxon rank sum test, all FDR adjusted using Benjamini-Hochberg method). (C) Visualization of high-throughput metatranscriptomic data obtained from the HPFS. We quantified the expression of bacterial azoreductases using salmon v1.4.1 (Patro et al., 2017) and normalized the raw read alignment statistics to transcripts per million (TPM). Please see Materials and Methods for a more detailed description of the computational and statistical methods employed.

Fig. 5.

Known and putative azo-reducing species are more abundant in the IBD gut. Relative abundance of known and putative azo-reducing species in HMP2 subjects with more than 20 total stool collections. (A) Relative abundance of known and putative azo-reducing species for qualifying participants across non-IBD, UC, and CD populations. Subjects with CD have significantly higher relative abundances of known+putative azo-reducing species than healthy subjects (non-IBD) per linear mixed-effects model ANOVA, F(1,10) = 17.09, P < 0.003. (B) Relative abundance of putative azo-reducing species over time for healthy (non-IBD), UC, and CD participants. Each line represents a single participant, and each point is the summed relative abundance of known+putative azo-reducing species at that collection point. The key on the right links relative abundance distributions for each subject with the same data point shown over collection numbers. Collections were taken approximately every 14 days.

Statistical Analysis of Relative Metagenomic Sequence Data from HMP2 and PRISM Datasets

Statistical analyses described in Fig. 4 were performed in R 4.1.1 with the Wilcoxon test method using default parameterization. Next, we compared the relative abundances of known and putative azo-reducing species for HMP2 subjects with 20 or more stool samples taken over the course of the study (Fig. 5). A linear mixed-effects model ANOVA was performed on this subset of HMP2 data to determine any statistically significant differences in relative abundance values across non-IBD, UC, and CD subjects. The R package lme4 (Bates et al., 2014) was used to fit the model to the data, and the package lmerTest (Kuznetsova et al., 2017) was used to perform the ANOVA on the model. All statistical analyses were performed in R version 4.1.1 and are provided at https://github.com/dombraccia/Azoreductases.

Acid Orange 7 Azo Reduction Assay

Biologic triplicates were grown in a 10 ml tube containing 10 ml of Yeast Casitone Fatty Acids (YCFA) broth for B. fragilis and C. clostridioforme strains and 10 ml of Brain Heart Infusion (BHI) broth for F. nucleatum strains. Each tube was inoculated with 10 μL of bacteria from glycerol stocks. The final concentrations of each substrate in the bacterial cultures were 50 μg/ml of FMN, 50 μg/ml of NADH, and 50 μmol/ml of Acid Orange 7. The bacterial cultures were left to grow in an anaerobic chamber for 72 hours, and Acid Orange 7 decolorization was measured once every 24 hours since inoculation. The decolorization of Acid Orange 7 was measured by aliquoting triplicates of 200 μL media aliquants to a 96-well plate for each bacterial culture, and absorbance of 550 nm light was measured using a spectrophotometer. The raw absorbance values for each biologic and technical replicate are reported in Supplemental Table 5.

Results

Primary Amino Acid Sequences of Bacterial Azoreductases Group by Mechanistic Preferences

To begin identifying putative azo-reducing species of the gut microbiome, we searched the literature for experimentally verified azoreductase enzymes, collected amino acid sequences, and metadata for these azoreductases (Supplemental Table 1) and compared their primary sequences using the EFI-EST (Gerlt et al., 2015). The resulting sequence similarity network captured mechanistic preferences such as flavin dependence and electron donor preference for each azoreductase reported by Suzuki (2019). We saw near-complete concordance between the clade I–IV azoreductases and the sequence similarity clusters at a 35% amino acid identity edge threshold, with the exception of clade IV, which was split into two separate clusters (Fig. 2). The gene families labeled arsH, mdaB, yieF, and azo1 clustered separately from clade I–IV azoreductases, and we consider each of these clusters as separate subfamilies of azoreductases. Multiple sequence alignments were generated for each cluster shown in Fig. 2 (with the exception of the “Other Azoreductases” group) using MUSCLE v3.8.425 (Edgar, 2004). HMMs were then trained on the multiple sequence alignments from the previous step using the HMMER v3.1b2 method hmmpress (Finn et al., 2015) and were queried against the UHGG collection (Almeida et al., 2020).

Homolog Search for Azoreductases Supports Evidence for Ubiquity of Azo Reduction by the Human Gut Microbiome

We searched for homologs of azoreductases across 4644 representative genomes in the UHGG collection (Almeida et al., 2020) using HMMs generated from sequences of experimentally validated azoreductase enzymes (Fig. 3). This collection contains 204,938 genome sequences of bacteria known to inhabit the human gut, of which 4644 are included in the representative collection (Supplemental Table 4). For the remainder of this work, we refer to species receiving statistically significant (E-value < 1 × 10⁻¹⁰) hits to azoreductase genes as “putative azo-reducing species” or “putative azo-reducing bacteria.” Of the 4644 genomes in the UHGG collection, there are 1443 (31.1%) with one putative azoreductase gene, 343 (7.4%) with two or more putative azoreductases, and 372 (8.0%) with three or more putative azoreductases, indicating the extensive potential of the gut microbiome to reduce azo bonds. Most notably, 364 genomes contain hits to the clade I profile, 452 contain hits to the clade II profile, 793 genomes contain hits to the clade III profile, 568 contain hits to the clade IVa profile, 410 contain hits to the clade IVb profile, 285 contain hits to the mdaB profile, and 477 contain hits to the yieF profile. Prominent phyla of the gut microbiome, such as Proteobacteria and Firmicutes, appear particularly rich with clade I, clade II, clade III, clade IVab, and flavin adenine dinucleotide utilizing azoreductases (purple columns in Fig. 3).

Systematic Evaluation of Predicted Azo-Reducing Species

We next sought to evaluate the results of our azoreductase homolog search with recent findings by Zimmermann et al. (2019) regarding sulfasalazine reduction. Zimmermann et al. (2019) tested the degradation of sulfasalazine by 76 prominent gut bacterial strains, 67 of which had corresponding reference genomes present in the UHGG collection. This provided an excellent source of data to compare our bioinformatic predictions against. We determined the sulfasalazine-reducing status as either sulfasalazine reducing (SR) or non–sulfasalazine reducing for each of the 67 strains based on the significant (FDR adjusted P value < 0.05) reduction of sulfasalazine in vitro reported by Zimmermann et al. (2019) (Table 1). We also determined the predicted reducing status as either a predicted reducer (PR) or a non–predicted reducer (Non-PR) for each strain based on the presence of a putative azoreductase identified from the homolog search step. For each strain, the sulfasalazine-reducing status and predicted reducing status were systematically compared to validate the results of the azoreductase homolog search (Table 1, columns 7–9). We correctly predicted the sulfasalazine-reducing status for 47.8% (32/67) of strains and we incorrectly predicted the sulfasalazine reducing status for 52.2% (35/67) of strains (Table 2). The vast majority (77.1%, 27/35) of incorrectly predicted strains are false negatives, meaning the strain does reduce sulfasalazine in vitro, but we did not identify an azoreductase in the homolog search step (Table 2). Interestingly, the majority of false positives (75%, 6/8) are members of the Proteobacteria phylum, which we previously noted to be particularly rich in azoreductase gene sequences (Fig. 3).

TABLE 1.

Systematic comparison with Zimmermann et al. (2019) sulfasalazine consumption results Complete results from the systematic comparison of predicted sulfasalazine-reducing bacteria to actual sulfasalazine-reducing bacteria. The columns labeled FC (fold change), FC_STD (standard deviation in fold change), P_FDR (FDR adjusted P value), Pct_Consumed (percent consumed), and Pct_Consumed_STD (standard deviation of the percent consumed) were all obtained directly from Zimmermann et al. (2019) (Supplemental Table 3). The SR_Status column contains values SR and non-SR, which were determined based on significant (p_FDR < 0.05) or nonsignificant (p_FDR ≥ 0.05) sulfasalazine reduction. The PR_Status column contains the values PR and non-PR, which were determined based on the presence or absence of one or more azoreductase homologs determined from the homolog search step. The final column, Result, contains the values TP (true positive), TN (true negative), FP (false positive), and FN (false negative). Correctly predicted SR strains have a result of TP and correctly predicted Non-SR strains have a result of TN whereas incorrectly predicted SR strains have a result of FP and incorrectly predicted Non-SR strains have a result of FN.

Strain_Name	FC	FC_STD	p_FDR	Pct_Consumed	Pct_Consumed_STD	SR_Status	PR_Status	Result
Akkermansia muciniphila ATCCBAA-835	−0.419	0.361	0.19	25.222	18.707	Non-SR	Non-PR	TN
Alistipes indistinctus DSM 22520	−9.01	0.11	0.003	99.806	0.015	SR	PR	TP
Anaerococcus hydrogenalis DSM7454	−8.826	0.198	0.008	99.78	0.03	SR	Non-PR	FN
Anaerotruncus colihominis DSM17241	−8.088	1.575	0.016	99.633	0.401	SR	Non-PR	FN
Bacteroides caccae ATCC43185	−9.247	0.262	0.008	99.835	0.03	SR	PR	TP
Bacteroides cellulosilyticus DSM14838	−1.502	0.308	0.002	64.691	7.532	SR	PR	TP
Bacteroides coprophilus DSM18228	−0.006	0.216	0.991	0.386	14.918	Non-SR	Non-PR	TN
Bacteroides dorei DSM17855	−4.631	0.748	0.013	95.965	2.091	SR	Non-PR	FN
Bacteroides eggerthii DSM20697	−9.59	0.104	0.002	99.87	0.009	SR	Non-PR	FN
Bacteroides finegoldii DSM17565	−0.79	0.296	0.042	42.176	11.877	SR	Non-PR	FN
B. fragilis 3397 T10	−0.974	0.348	0.009	49.084	12.282	SR	PR	TP
B. fragilis ATCC43859	−10.576	0.088	0.003	99.934	0.004	SR	PR	TP
B. fragilis DS-208	−9.327	0.57	0.006	99.844	0.062	SR	PR	TP
B. fragilis HMW610	−10.522	0.159	0.004	99.932	0.007	SR	PR	TP
B. fragilis HMW615	−10.398	0.256	0.013	99.926	0.013	SR	PR	TP
B. fragilis NCTC9343	−6.184	0.835	0.006	98.625	0.796	SR	PR	TP
B. fragilis T(B)9	−9.252	0.144	0.005	99.836	0.016	SR	PR	TP
Bacteroides intestinalis DSM17393	−1.296	0.398	0.005	59.267	11.246	SR	PR	TP
Bacteroides ovatus ATCC8483	−0.285	0.473	0.546	17.951	26.915	Non-SR	PR	FP
Bacteroides pectinophilus ATCC43243	−0.249	0.241	0.268	15.832	14.042	Non-SR	Non-PR	TN
Bacteroides stercoris ATCC43183	−0.588	0.326	0.053	33.46	15.033	Non-SR	Non-PR	TN
Bacteroides thetaiotaomicron 3731	−1.303	0.248	0.005	59.461	6.961	SR	PR	TP
Bacteroides thetaiotaomicron 7330	−1.032	0.237	0.003	51.082	8.044	SR	PR	TP
Bacteroides thetaiotaomicron VPI-5482	−1.252	0.155	0.006	58.006	4.511	SR	PR	TP
Bacteroides uniformis ATCC8492	−2.605	0.389	0.001	83.558	4.436	SR	Non-PR	FN
Bacteroides xylanisolvens DSM18836	−9.663	0.115	0.002	99.877	0.01	SR	PR	TP
Bifidobacterium adolescentis ATCC15703	−0.7	0.271	0.014	38.442	11.584	SR	Non-PR	FN
Bifidobacterium breve DSM20213	−9.241	0.182	0.008	99.835	0.021	SR	Non-PR	FN
Blautia hansenii DSM20583	−9.234	0.155	0.005	99.834	0.018	SR	Non-PR	FN
Bryantia formatexigens DSM14469	−0.722	0.432	0.113	39.359	18.171	Non-SR	PR	FP
Clostridium asparagiforme DSM15981	−9.184	0.203	0.01	99.828	0.024	SR	Non-PR	FN
Clostridium bolteae ATCCBAA-613	−7.113	0.611	0.005	99.277	0.306	SR	Non-PR	FN
Clostridium difficile 120	−6.039	0.703	0.013	98.479	0.741	SR	PR	TP
Clostridium ramosum DSM1402	−9.046	0.169	0.006	99.811	0.022	SR	PR	TP
Clostridium scindens ATCC35704	−9.064	0.324	0.004	99.813	0.042	SR	Non-PR	FN
Clostridium spiroforme DSM1552	−2.478	0.572	0.001	82.055	7.119	SR	Non-PR	FN
Clostridium sporogenes ATCC15579	−9.072	0.158	0.006	99.814	0.02	SR	PR	TP
Clostridium symbiosum ATCC14940	−9.256	0.209	0.009	99.836	0.024	SR	Non-PR	FN
Collinsella aerofaciens ATCC25986	−6.629	0.628	0.046	98.99	0.44	SR	PR	TP
Collinsella intestinalis DSM13280	−8.916	0.118	0.003	99.793	0.017	SR	Non-PR	FN
Coprococcus comes ATCC27758	−9.182	0.199	0.009	99.828	0.024	SR	Non-PR	FN
Dorea formicigenerans ATCC27755	−4.071	0.57	0.015	94.051	2.351	SR	Non-PR	FN
Edwardsiella tarda ATCC23685	−0.107	0.287	0.722	7.125	18.483	Non-SR	PR	FP
Eggerthella lenta ATCC25559	−0.435	0.217	0.038	26.042	11.107	SR	PR	TP
Enterobacter cancerogenus ATCC35316	−0.853	0.579	0.076	44.621	22.214	Non-SR	PR	FP
Enterococcus faecalis V583	−8.61	0.228	0.01	99.744	0.04	SR	PR	TP
Escherichia coli K-12	−0.714	0.331	0.038	39.057	13.995	SR	PR	TP
Eubacterium biforme DSM3989	−8.927	0.098	0.008	99.795	0.014	SR	Non-PR	FN
Eubacterium hallii DSM3353	−9.031	0.381	0.028	99.809	0.05	SR	Non-PR	FN
Eubacterium rectale ATCC33656	−9.002	0.322	0.022	99.805	0.044	SR	PR	TP
Eubacterium ventriosum ATCC27560	−4.653	0.827	0.051	96.024	2.278	Non-SR	Non-PR	TN
Odoribacter splanchnius	−7.892	0.744	0.004	99.579	0.217	SR	PR	TP
Parabacteroides distasonis ATCC8503	−1.007	0.298	0.022	50.253	10.272	SR	Non-PR	FN
Parabacteroides johnsonii DSM18315	−0.529	0.281	0.123	30.702	13.51	Non-SR	Non-PR	TN
Parabacteroides merdae ATCC43184	−0.749	0.208	0.006	40.508	8.593	SR	Non-PR	FN
Pretovella copri DSM18205	−8.693	0.41	0.015	99.758	0.069	SR	Non-PR	FN
Proteus penneri ATCC35198	−1.657	0.313	0.009	68.289	6.884	SR	PR	TP
Providencia alcalifaciens DSM30120	−0.379	0.247	0.108	23.094	13.183	Non-SR	PR	FP
Providencia rettgeri DSM1131	−0.205	0.204	0.25	13.245	12.273	Non-SR	PR	FP
Providencia stuartii ATCC25827	−0.072	0.246	0.807	4.854	16.212	Non-SR	PR	FP
Roseburia intestinalis L1-82	−8.876	0.156	0.006	99.787	0.023	SR	Non-PR	FN
Ruminococcus gnavus ATCC29149	−8.639	0.045	0	99.749	0.008	SR	Non-PR	FN
Ruminococcus lactaris ATCC29176	−9.522	0.145	0.005	99.864	0.014	SR	Non-PR	FN
Ruminococcus torques ATCC27756	−2.384	0.298	0.016	80.841	3.961	SR	PR	TP
Salmonella Typhimurium LT2	−0.673	0.712	0.252	37.287	30.936	Non-SR	PR	FP
Subdoligranulum variabile DSM15176	−8.854	0.189	0.008	99.784	0.028	SR	Non-PR	FN
Victivallis vadensis ATCC BAA-548	−1.844	0.392	0.004	72.146	7.574	SR	Non-PR	FN

TABLE 2.

Summarized results of systematic Zimmermann et al. (2019) comparison This table displays the summarized results of the systematic comparison of predicted sulfasalazine reducers to experimentally confirmed sulfasalazine reducers reported by Zimmermann et al. (2019).

	PR	Non-PR
SR	38.8% (26/67)a	40.3% (27/67)b
Non-SR	11.9% (8/67)c	9.0% (6/67)d

^aThe number of true positives.

^bThe number of false negatives.

^cThe number of true false positives.

^dThe number of true negatives.

Exploratory Analysis of Azoreductase Abundance and Expression Levels in the Human Gut Microbiome

After identifying putative azo-reducing species of the human gut microbiome, we next sought to examine the abundance and expression of putative azoreductases using publicly available metagenomic and metatranscriptomic datasets. We used shotgun metagenomic sequence data from the HMP2 (Proctor and Huttenhower, 2019) and the PRISM (Franzosa et al., 2019) to quantify azoreductase gene abundance. We also used high-throughput metatranscriptomic sequence data from the HPFS (Abu-Ali et al., 2018) to quantify the expression of azoreductases by the human gut microbiota (Fig. 4, A–C). Briefly, raw genomic and transcriptomic reads were filtered and processed using fastp (Chen et al., 2018), and azoreductases were quantified using salmon v1.4.0 (Patro et al., 2017). Please see Materials and Methods for more details on the computational and statistical procedures used.

Significant differences in azoreductase gene abundances between disease conditions are displayed in Fig. 4, A and B with asterisks. We find that clade I, clade II, clade III, clade IVa, clade IVb, mdaB, and yieF genes are considerably higher in abundance than ferB, azoR1_cperf, arsH, azoR_ropacus, azo1, and azoR1_llentus within all three disease conditions for both HMP2 (all FDR adjusted P < 2.2 × 10⁻¹⁶) and PRISM (all FDR adjusted P < 2.6 × 10⁻⁷) (Fig. 4, A and B). Clade IVa is higher in abundance than all other azoreductases across healthy, UC, and CD cohorts from HMP2 (Fig. 4A), but the same statistically significant difference was not observed in the PRISM study (Fig. 4B).

The expression of clade I, clade II, clade III, clade IVa, mdaB, and yieF azoreductases are significantly higher than those of clade IVb, ferB, azoR1_cperf, arsH, azoR_ropacus, azo1, and azoR1_llentus azoreductases in healthy individuals (minimum FDR adjusted P value < 1 × 10⁻⁷ between mdaB and azoR1_cperf) (Fig. 4C). Although clade IVb abundance levels are comparable to those of clades I, II, III, and IVa, the expression levels of clade IVb azoreductases are significantly lower in vivo than the expression clade I, II, III, and IVa azoreductases (all FDR adjusted P < 2.2 × 10⁻¹⁶).

The Relative Abundance of Putative Azo-Reducing Species Fluctuates over Time

We next sought to examine whether relative abundance levels of combined known and putative azo-reducing species are stable or fluctuate over time. The HMP2 dataset provides a unique opportunity to examine the stability of individuals’ gut microbiomes over time as there are 18 individuals across healthy, UC, and CD cohorts, with at least 20 stool samples taken once every 2 weeks over a 6-month period. To examine the stability of azo reduction in the human gut, we compared the relative abundance of known and putative azo-reducing species from these participants (Fig. 5). The median relative abundance of combined azo-reducing species ranges from 20.3 ± 3.58% to 33.9 ± 19.2% for non-IBD, 21.7 ± 7.6% to 49.0 ± 15.0% for UC, and 34.9 ± 6.39 to 62.3 ± 18.8 for CD subjects. Using linear mixed-effects model ANOVA, we found that combined azo-reducing species are significantly more abundant in CD subjects than in non-IBD subjects (P = 0.002) and are not significantly more abundant in UC subjects than in non-IBD subjects (P = 0.064) (Fig. 5A). Note that Fig. 5B shows the same relative abundance values displayed in Fig. 5A but over the course of the study, from collection 1 to collection 24.

Multiple Strains of F. Nucleatum, B. Fragilis, and C. Clostridioforme Reduce Acid Orange 7 in Vitro

Finally, we sought to test the azo reduction of Acid Orange 7 by three strains of the health-relevant (Castellarin et al., 2012; Kostic et al., 2012; Bashir et al., 2015; Abed et al., 2016) microbe F. nucleatum. As well, we tested the azo reduction of Acid Orange 7 by two positive control species, B. fragilis and C. clostridioforme. Acid Orange 7 is an azo-bonded dye commonly used in the food and textile industries (Bay et al., 2014), and the decolorization of azo-bonded dyes is commonly used to test azo reduction by bacteria in vitro (Feng et al., 2012). F. nucleatum CTI-06, F. nucleatum subsp. animalis D11, and F. nucleatum subsp. polymorphum were grown in BHI media, and Acid Orange 7 was added to the culture after 4 days of growth (Materials and Methods). We also tested the azo reduction of Acid Orange 7 by the known azo-reducing species B. fragilis and Clostridium clostridiforme. B. fragilis strains 3_1_12 and CL07T00C01 and C. clostridioforme strains 2_1_49_FAA and WAL-7855 were grown in YCFA broth and served as positive controls. We found that all strains examined in this assay significantly decolorized Acid Orange 7 in vitro (Fig. 6). To our knowledge, this is the first reporting of azo reduction by F. nucleatum CTI-06, F. nucleatum subsp. animalis D11, and F. nucleatum subsp. polymorphum.

Fig. 6.

Three putative azo-reducing strains of F. nucleatum degrade Acid Orange 7 in vitro. The absorbance of light at 550 nm (corresponding to the absorbance spectra of Acid Orange 7) was measured in cultures of F. nucleatum, B. fragilis, and C. clostridiforme isolate cultures. F. nucleatum strains were grown in BHI media and were compared with BHI-blank control mixture, whereas B. fragilis and C. clostridiforme strains were grown in YCFA media and were thus compared with a YCFA-blank. Each strain was grown and tested in biologic and technical triplicates. Each data point on the plot above is the average of three technical replicates from a single biologic replicate per strain. Please see Materials and Methods for more details regarding our experimental methodology. Asterisks indicate statistical significance calculated via two-sided t tests (*P < 0.05; **P < 0.01; ***P < 0.001).

Discussion

The presence of azo-reducing bacteria in the human gut is necessary for the effective delivery and activation of azo-bonded prodrugs. Although azoreductase activity has been identified in several prominent phyla of the human gut microbiota (Zimmermann et al., 2019) and appears to be ubiquitous across healthy individuals (Javdan et al., 2020), the prevalence, abundance, and distribution of azoreductases have not been systematically examined in the human gut microbiome of healthy individuals nor in individuals living with IBD. In this work, we curated and compiled known azoreductase genes (Fig. 2), searched for azoreductase gene families across a nonredundant set of 4644 human gut bacterial genomes (Almeida et al., 2020), and identified 1958 putative azo-reducing species (Fig. 3). The systematic comparison of our search results to recent experimental evidence of sulfasalazine reduction by prominent gut bacteria (Table 1, Table 2) indicates a disconnect between the current state of azoreductase annotation and experimental evidence of sulfasalazine reduction. Interestingly, the majority (77.1%, 27/35) of incorrectly predicted sulfasalazine-reducing strains are false negatives, meaning these strains did not return a significant hit to an azoreductase gene from the homolog search step but do, in fact, reduce sulfasalazine in vitro. This inconsistency between annotated azoreductases and experimental evidence of azo reduction suggests that many prominent bacterial strains of the human gut microbiome may encode and express previously uncharacterized azoreductase genes. These genes likely serve other endogenous roles such as nitro reduction (Liu et al., 2007; Chalansonnet et al., 2017) and quinone oxidoreduction (Leelakriangsak et al., 2008; Liu et al., 2008; Ryan et al., 2010a; Ryan et al. 2010b; Ryan et al., 2014), with the azo reduction being a side mechanism that these enzymes crossfunctionally participate in.

We next sought to report the relative abundance and expression of azoreductases in the human gut microbiome for healthy controls and IBD patients. Our analysis of 1558 metagenomic samples from 326 individuals across healthy, UC, and CD patient cohorts showed that clade I, II, III, IVa, IVb, mdaB, and yieF azoreductases are significantly more abundant in the gut microbiome compared with the other azoreductases examined in this study (Fig. 4, A and B). We also examined the expression of azoreductases by the human gut microbiota and found that, with the exception of clade IVb, expression levels of azoreductases roughly match with their corresponding genomic abundance (Fig. 4C). The incongruence of clade IVb abundance and expression levels suggests that, when feasible, shotgun metagenomic sequencing of stool samples should be performed in parallel with metatranscriptomic sequencing to better understand the functional landscape of the gut microbiome and the relative contributions of different azoreductases to overall azo reduction. We also sought to examine the relative abundance of known and putative azo-reducing bacteria in healthy, UC, and CD patients over time. We found that the relative abundance of known and putative azo-reducing bacteria is significantly (P = 0.002) higher in individuals with CD and is modestly (P = 0.06) higher in individuals with UC compared with healthy controls (Fig. 5). This bodes well for the future of azo-bonded prodrug development because these therapies are intended to treat individuals afflicted with UC and CD. However, the cumulative relative abundance of known and putative azo-reducing bacteria fluctuates over time (Fig. 5B), and future studies should explore whether there exists some minimum necessary abundance of azo-reducing species for adequate prodrug metabolism and activation.

Finally, we tested the reduction of the azo-bonded dye Acid Orange 7 by three strains of F. nucleatum alongside positive control strains of B. fragilis and C. clostridioforme (Fig. 6). F. nucleatum is positively correlated with colorectal cancer (Marchesi et al., 2011; Kostic et al., 2012), is present in and on cancerous tissue (Castellarin et al., 2012), and possibly contributes to the etiology of the disease (McCoy et al., 2013; Rubinstein et al., 2013; Han, 2015). We found that F. nucleatum CTI-06, F. nucleatum subsp. animalis D11, and F. nucleatum subsp. polymorphum all significantly reduce Acid Orange 7 in vitro, indicating the encoding and activity of azoreductases in these strains of F. nucleatum. The F. nucleatum reference strain present in UHGG received significant hits to the mdaB (E-value = 4.20 × 10⁻¹³) and yieF (E-value = 1.40 × 10⁻²³) HMMs and, thus, represents an accurately predicted azo-reducing bacteria (Supplemental Table 3). The identification and characterization of these, and possibly other, F. nucleatum azoreductases could lead to the eventual development of an azo-bonded colorectal cancer therapeutic designed specifically to activate in the presence of F. nucleatum on the surface of colonic tumors.

B. fragilis is a known reducer of azo dyes including Acid Orange 7 (this study), Amaranth, Orange II, and Tartrazine (Bragger et al., 1997), as well as of the quinone menadione (Ito et al., 2020). Additionally, many B. fragilis strains have been shown to be potent reducers of sulfasalazine in vitro (Zimmermann et al., 2019). Although we did identify a significant (E-value < 1 × 10⁻⁴⁵) hit to the clade IVa HMM in the two B. fragilis reference strains present in UHGG (Supplemental Table 3), there may be other B. fragilis genes or operons that exhibit azoreductase activity. Ito et al. (2020) described two NADH:quinone oxidoreductase operons, NQR and NUO, and one NADH:quinone oxidoreductase gene, ndh2, capable of reducing the quinone menadione. Recall that bacterial quinone oxidoreductases are often crossreactive with azo compounds and have even been proposed to be a part of the same FMN-dependent superfamily of NAD(P)H utilizing oxidoreductase enzymes (Ryan et al., 2014). Future studies are required to confirm or deny that NQR, NUO, and ndh2 are hitherto uncharacterized azoreductases contributing to the complete azo reduction of sulfasalazine by B. fragilis shown in Zimmerman et al. (2019).

Of the seven bacterial strains tested for reduction of Acid Orange 7, the two C. clostridioforme strains exhibited by far the most effective reduction of Acid Orange 7 (Fig. 6). Although C. clostridioforme is a known azo dye reducer (Raffi and Cerniglia, 1990; Nakamura et al., 2002; Xu et al., 2010), neither of the two reference strains present in UHGG recruited significant alignments to known azoreductase gene families curated in the homolog search step of this work (Supplemental Table 3). This could be the result of either 1) strain-level variation between the reference strains and those tested with Acid Orange 7 in this study or 2) the presence and activity of one or more uncharacterized azoreductases in C. clostridioforme. In either case, further research including a comparative genomics analysis and gene knockout experiment on various strains of C. clostridioforme could lead to an improved understanding of gut microbial azo reduction.

This study has two primary limitations. 1) The E-value and percent of alignment thresholds for determining a putative azoreductase in the homolog search step are not absolute but rather are designed to strike a balance between identifying spurious homologs and missing the identification of true azoreductase homologs. This is an inherent limitation of studies requiring hard cutoffs for homolog classification and, thus, is very difficult to avoid. 2) Bacterial azoreductases exhibit different substrate specificities (Bin et al., 2004; Deller et al., 2006; Sugiura et al., 2006; Joshi et al., 2008; Ryan et al., 2010a, ; Mendes et al., 2011; Lang et al., 2013) and, thus, have varying affinities for different azo prodrugs as well as azo dyes. Though we show a significant reduction of Acid Orange 7 by three strains of F. nucleatum in this work, future experiments showing the reduction of azo drugs such as sulfasalazine would further bolster the hypothesis that F. nucleatum encodes and expresses one or more uncharacterized azoreductases.

In conclusion, we show that known azoreductases are widely distributed in the human gut microbiome and that there are likely many more uncharacterized azoreductases encoded and expressed in the human gut microbiome. These results both 1) bolster previous findings suggesting the ubiquity of azo-reduction in the gut microbiome (Javdan et al., 2020) and 2) suggest the presence and activity of many hitherto uncharacterized azoreductases in the human gut microbiome. The list of false negative strains identified in our systematic comparison analysis can serve as a resource for future studies focused on identifying azoreductases encoded by the human gut microbiome (Table 1). Overall, this work describes the abundance and distribution of known azoreductases in the human gut microbiome and motivates the need for future studies focused on annotating hitherto uncharacterized azoreductases encoded in the human gut microbiome. Further validation and annotation of putative azoreductases encoded by prominent members of the gut flora such as B. fragilis, F. nucleatum, and C. clostridioforme, are important for functional characterization of azo reduction by the human gut microbiome and for the future of azo prodrug development.

Abbreviations

5-ASA

5-aminosalicylic acid

BHI

Brain Heart Infusion

CD

Crohn disease

CRC

colorectal cancer

DHQ

dihydroquinolone

EFI-EST

Enzyme Function Initiative-Enzyme Similarity Tool

FDR

false discover rate

FMN

flavin mononucleotide

FN

false negative

FP

false positive

HMM

hidden Markov model

HMP2

Integrative Human Microbiome Project

HPFS

Health Professionals Follow-Up Study

IBD

inflammatory bowel disease

PR

predicted reducer

PRISM

Prospective Registry of IBD Patients at MGH

SR

sulfasalazine reducing

TN

true negative

TP

true positive

UC

ulcerative colitis

UHGG

Unified Human Gastrointestinal Genomes

YCFA

Yeast Casitone Fatty Acid

Authorship Contributions

Participated in research design: Braccia, Minabou Ndjite, Jiang, Pop, Hall.

Conducted experiments: Minabou Ndjite, Weiss, Levy, Abeysinghe.

Performed data analysis: Braccia, Minabou Ndjite.

Wrote or contributed to the writing of the manuscript: Braccia, Minabou Ndjite, Weiss, Levy, Abeysinghe, Jiang, Pop, Hall.