Abstract
16S rRNA amplicon sequencing provides a relatively inexpensive culture-independent method for studying the microbial world. Although thousands of such studies have examined diverse habitats, it is difficult for researchers to use this vast trove of experiments when analyzing their findings and interpret them in a broader context. To bridge this gap, we introduce dbBact, an open wiki-like bacterial knowledge base. dbBact combines information from hundreds of studies across diverse habitats, creating a collaborative central repository where 16S rRNA amplicon sequence variants (ASVs) are manually extracted from each study and assigned multiple ontology-based terms. Using the >900 studies of dbBact, covering more than 1,400,000 associations between 345,000 ASVs and 6,500 ontology terms, we show how the dbBact statistical and programmatic pipeline can augment standard microbiome analysis. We use multiple examples to demonstrate how dbBact leads to formulating novel hypotheses regarding inter-host similarities, intra-host sources of bacteria, and commonalities across different diseases, and helps detect environmental sources and identify contaminants.
Introduction
Bacteria play an important role in the Earth’s ecosystem, having a total biomass higher than that of all vertebrates and fish, second only to plants (1). The introduction of 16S rRNA amplicon sequencing as a means for molecular identification has enabled a culture-independent view of such ecosystems (2). Combined with massively parallel sequencing technologies and DNA barcodes (3), 16S rRNA sequencing provides relatively cheap and accurate microbial profiling. This, in turn, has led to a huge surge in the number of 16S rRNA studies examining microbial populations in habitats ranging across oceans (4), soil (5), plants (6), animals (7), and large cross-sectional human studies (8–10).
A severe limitation when combining insights from multiple microbiome studies is the complexity of the underlying bacterial populations, ranging from tens of different bacteria in a single saliva sample (11), to thousands in a single soil sample (12). In addition, although the total number of different bacteria is large (e.g., ∼300000 unique 16S rRNA sequences of length 90bp appear in the Earth Microbiome Project (12)), the number of taxonomic names for describing these bacteria is much smaller (about 3500-4000 unique genera and 20000 unique species names appear in NCBITAX (13) and in the Encyclopedia of Life (14)). Moreover, grouping bacteria in higher taxonomic levels may not always maintain the basic habitat properties of many bacteria (12). Therefore, reaching cross-study biological insights should preferably be based on 16S rRNA amplicon sequences rather than on taxonomy.
Recently, amplicon sequence variants (ASVs) derived using denoising methods such as Deblur (15), DADA2 (16) and UNOISE2 (17) have been introduced as an alternative to OTU picking for identifying bacteria in a given sample. Such denoising methods provide an objective identification of each bacterial sequence in the sample (i.e., independent of external databases or additional bacteria/samples in the experiment), as well as high sequence resolution (a single nucleotide difference in the sequenced region is identified as a different ASV). Therefore, ASVs may serve as cross-study identifiers for bacteria, i.e., a bacterium in different studies will result in the same ASV, even when the studies are processed separately and denoised using different methods (15,18).
In this paper, we introduce dbBact, a knowledge base for reaching cross-experiment biological insights. dbBact is based on manually collecting genotype-phenotype associations between ASVs and relevant conditions. For clarity, “reserved” dbBact words appear in italics: surveyed studies are referred to as experiments, stored ASVs are sequences, phenotypes are ontology-based terms, and genotype-phenotype associations are called annotations.
Currently, dbBact contains more than 900 experiments in various habitats, covering more than 1,400,000 associations between 345,000 sequences and 6500 ontology terms. For retrieval, dbBact provides two query types: a single sequence/FASTA file query, asking what is known about these sequences (Figure 1), and a query contrasting two FASTA files, searching for dbBact terms significantly enriched in either of the groups, analogously to gene ontology enrichment analysis (19,20) (Figure 2). By examining the ontology terms associated with each sequence, users can gain insights regarding the biology associated with ASVs of interest.
Users add new entries in a wiki-like way, by uploading study results. a. For example, analyzing data from Ijaz et al. (37), we identified 189 ASVs that are more abundant in fecal samples of Scottish children with Crohn’s disease compared to healthy controls (see Methods section). b. These ASVs are uploaded as a FASTA file. c. Associations between ASVs and phenotypes are called annotations, which are created by assigning a set of ontology terms and predicates that characterize the context. The 189 sequences were annotated as “DIFFERENTIAL,” i.e., more abundant in children with Crohn’s disease (“HIGHER IN” terms), compared to healthy controls (“LOWER IN” terms). The general background terms common to both groups, i.e., “homo sapiens,” “feces” and “glasgow” are designated by “SOURCE.” Terms may be selected from several ontologies (e.g., DOID (33), ENVO (38,39), GAZ (40), UBERON (41), EFO (42), and NCBI Taxonomy (13)), allowing easy and precise annotations. d. Uploading annotations may be performed either through the dbBact website, dedicated clients (i.e., Calour (36)) or by REST-API. For clarity, the following nomenclature holds throughout the manuscript where “reserved” words appear in italics (e.g., experiment, sequence, annotation, term), predicates appear in all caps (e.g., “HIGHER IN,” “LOWER IN,” “SOURCE”), and specific term names follow the ontology convention of being lower case (e.g., “homo sapiens”).
a. Uploading a FASTA file of sequences results in a list of the most relevant annotations containing these sequences, and a “word cloud” of best matching terms. In this example, a V4 sequence of Clostridium XIVa, which is highly abundant in fecal samples of chronic fatigue syndrome patients (CSF) (Giloteaux et al., 2016), was submitted. Panel a1 provides representative annotations containing the query sequence (the full list of ∼150 annotations appears in Supplementary File 3). dbBact found this sequence to be higher in the disease group than in healthy controls in several studies (ulcerative colitis, irritable bowel disease, and lupus), and in antibiotic-treated mice supplemented with probiotics (last annotation arising from (43)). Panel a2 displays the word cloud summarizing the terms associated with the query sequence, where size corresponds to a term’s F1 score, while color designates the associated predicate (blue for “SOURCE”/”HIGHER IN” terms, and red color preceded by a minus sign corresponds to “LOWER IN” terms). Hence, this Clostridium XIVa query sequence is associated with human feces in dysbiosis states of “crohn’s disease,” “ulcerative colitis,” “diarrhea,” and “c. difficile infection” (a full list of F1 scores per term appears in Supplementary File S7). b. By contrasting two groups of sequences, dbBact identifies enriched terms characterizing each group. For example, 137 and 56 sequences were submitted, corresponding to differentially abundant sequences higher in fecal samples from domestic dogs and wolves living in zoos, respectively (data from (44)). Bar lengths show the normalized rank-mean difference for the top significantly enriched terms in the dog and wolf sequences (green and red bars, respectively). Term enrichment is based on a non-parametric rank mean test with FDR<0.1 using dsFDR (see the term enrichment analysis section in Methods). The numbers in the bar of each term correspond to the number of dbBact experiments in which the term differs significantly between the two sequence groups (numerator) and the total of dbBact experiments containing the term (denominator). Sequences that were more abundant in the wolf group are enriched in terms related to wolf, meat diet, and cheetah (Acinonyx jubatus).
dbBact differs from other microbial databases in several aspects: (a) Manual annotation: dbBact phenotype-genotype associations are extracted using manual analysis, in contrast to microbial data repositories, such as SRA/EBI, Qiita (21), HMP DACC (10,22), MGnify (23), FoodmicrobioNet (24), and redBIOM (25), which provide raw experimental data and metadata. In dbBact, the human expert understands the experimental setting and identifies abundant bacteria in different study groups, detects contaminants, etc., and these associations are uploaded. (b) Unlimited scope: dbBact accepts studies across all habitats, unlike databases that are highly limited in scope, e.g., human-disease-focused databases (MicroPhenoDB (26), gutMDisorder (27), Disbiome database (28), Peryton (29), and BugSigDB (30)), or other context-specific databases (database of the healthy mouse microbiome (31) or the sponge microbiome project (32)). (c) Volume and potential growth: to date, the number of studies in dbBact is ∼40% higher than in Qiita, and the number of ASVs is comparable to those in the Earth Microbiome Project. New studies are continually added by the dbBact team. Additionally, as a wiki-like database, we encourage the microbiome community to contribute to dbBact. (d) Structured genotype/phenotype search: observations are uploaded at the ASV level, allowing queries of specific sequences and subsequences. In addition, as phenotypes are designated by terms derived from multiple ontologies, subsequent querying allows for “cross-sectioning” of the data. For example, sequences associated with Crohn’s disease and ulcerative colitis both originate from the DOID ontology (33), and will be recalled when querying their “parent” term, “inflammatory bowel disease.” (e) Harmonizing studies performed using different variable regions: as uploaded studies may be sequenced using different 16S rRNA variable regions, stored sequences are “linked” through the SILVA database of full-length 16S rRNA genes (34), facilitating cross-region queries. For example, when submitting a query sequence from V1-V2, dbBact seeks the matching full-length 16S rRNA genes in SILVA, then extracts their V4 region, and subsequently retrieves relevant annotations. (f) Data analysis: dbBact provides a set of statistical tools for analyzing new studies and for generating novel biological hypotheses.
In the following sections, we present the current scope and comprehensiveness of dbBact, as well as demonstrate, using multiple examples, how dbBact may be incorporated into standard microbiome analysis pipelines, thus providing novel hypotheses.
dbBact can be accessed using its website (http://dbbact.org), plugins for Qiime2 (35) and Calour (36), and programmatically using the dbBact REST-API interface.
Results
dbBact: Scope and comprehensiveness
dbBact release 2022.07.01 contains approximately 345,000 unique bacterial amplicon sequences, an amount that is on par with the 300,000 sequences observed by the Earth Microbiome Project (12). Sequences arise from over 900 unique experiments, i.e., studies from which observations were added (Figure 3a). Over 7000 dbBact annotations associate these sequences with various phenotypes using ontology derived terms. As each annotation typically includes many sequences, this results in over 1,400,000 unique genotype-phenotype associations.
a. Scope of dbBact release 2021.05 (used for the analysis presented in this paper). b. The number of experiments for representative disease categories based on the DOID ontology. c. Scatter plot of the total number of annotations and experiments in which each dbBact term appears. d. Histogram of the number of experiments in which each dbBact sequence appears. e. Knowledge base comprehensiveness. The fraction of “COMMON” sequences from each experiment that have been annotated in additional experiments is shown for various terms. The number of experiments containing the term is designated above each bar. f. Comprehensiveness in a source tracking task. Sequences from eight sample types from Hägglund et al. were blindly submitted to dbBact. Their word clouds clearly display the sources of the samples (shown by the matched cartoon). Term sizes correspond to the F1 score of each term, combined for all sequences present in > 0.3 of the samples (for each sample type).
General statistics of dbBact
The >900 dbBact experiments cover a wide range of habitats (Figure S1a), geographic regions (Figure S1b), plant and animal hosts (Figure S1c), human body sites (Figure S1d), and human diseases (Figure 3b). For example, 149 experiments cover diseases, of which 86 are of an anatomical entity (e.g., Crohn’s disease or ulcerative colitis), and seven more are defined as metabolic diseases. The most abundant dbBact terms are “united states of america,” “homo sapiens,” and “feces,” each appearing in over 1000 annotations arising from more than 150 different experiments. Most of the other terms appear in less than twenty experiments (Figure 3c). The most prevalent bacterial sequence is E. coli, appearing in over 900 annotations from over 300 experiments (Figure 3d). Although this could reflect the universality of E. coli in various habitats, it may also be due to potential contaminations (45), a reason that may also explain the high prevalence of Staphylococcus (appearing in over 200 dbBact experiments). The number of experiments per sequence follows a power law distribution, with a majority of sequences appearing in a single experiment, yet over 80,000 sequences were observed in more than one experiment and 7000 sequences appeared in at least ten experiments (Figure 3d).
dbBact allows the upload of sequences from several commonly used regions (V1-V2, V3-V4 or V4; see Table S2 for a list of primers). Upon upload, sequences from different regions are “linked” through their full-length 16S rRNA sequence in the SILVA database (34) (see Inter-region querying section in Methods). When submitting query sequences from one region, dbBact retrieves all annotations containing the corresponding sequences across all regions (including, naturally, the region from which the query was provided). To demonstrate the usefulness of such “linking,” Figure S2 provides several examples of V1-V2 and V3-V4 sequence queries that are successfully characterized based solely on “linked” V4 sequences.
Comprehensiveness of dbBact
Intra-dbBact estimates
To estimate the comprehensiveness of dbBact, we tested how many bacterial sequences typical of a specific environment (e.g., human feces) have annotations arising from more than one experiment. We selected sequences having an annotation of type “COMMON” (i.e., present in more than half of the samples in an experiment) for each of several terms, and measured the fraction of these sequences that have annotations from another dbBact experiment (Figure 3e). For example, there are 97 experiments having a “COMMON IN homo sapiens feces” annotation. Iterating over each of these annotations, about 98% of the associated sequences appear in more than one experiment. Hence, fecal bacteria are already well covered by dbBact. A similar level of “coverage” occurs for several other human-related terms and for dogs, where almost all sequences were observed in more than a single experiment. Regarding the terms “cow,” “soil,” “root,” and “leaf,” about 80% of the sequences appear in more than a single experiment, whereas the “coverage” of “green turtle” is much lower, indicating that additional experiments are required to capture its full bacterial diversity.
Out-of-sample comprehensiveness
As another example of comprehensiveness, dbBact was tested in a source tracking task, i.e., identifying the host or niche of a sample based on its bacterial composition. Hägglund et al. collected samples from either sewage influent or from freshwater, as well as feces sampled from several animals (rabbit, cat, wild boar, dog, cow and deer), aiming to find unique bacterial footprints of each source (46). We used all sequences present in more than 1/3 of the samples from each group as queries to dbBact resulting in word clouds describing each sample group (Figure 3f). In almost all cases, the notable terms in each word cloud were indicative of the sources of the samples, e.g., sus scrofa for the wild boar fecal sample, allowing accurate source tracking. The only exception was cat fecal samples, which were detected as a combination of cat, dog, and human, probably because of the small number of cat fecal samples present in the current dbBact release.
The advantage of sequence-based associations
Results of 16S rRNA profiling experiments comprise a list of ASVs found in each sample and their abundances. Corroboration of these results with other microbiome studies is typically performed by searching published studies mentioning the taxonomy of these sequences. In many cases, however, such text-based mining may be limited because of constraints in taxonomic assignment. First, taxonomy is far from being full, e.g., species-level assignment is missing for about 80% of 16S rRNA sequences in Greengenes (47), and about 35% of the Greengenes sequences lack a genus assignment (48). Second, in many cases the same assigned taxonomy may be associated with vastly different phenotypes. As observed by the Earth Microbiome Project, bacteria of the same genus may be present in vastly different habitats, whereas specific sequences are associated with a certain habitat (12). This phenomenon underscores the importance of sequence-based association as provided by dbBact. For example, both sequences in Figure 4a belong to the genus Blautia, hence taxonomy-based associations may conclude that they play similar “roles” and are associated with the same phenotype. But querying dbBact with each of these two sequences results in a strikingly different picture, which we refer to as a “good” and “bad” Blautia. The “good” Blautia is more abundant in healthy controls than in patients of type 1 diabetes (T1D), Crohns’ disease (CD), inflammatory bowel disease (IBD), diarrhea, and kidney stones (Figure 4a), whereas the “bad” Blautia is more prevalent in patients suffering from IBD, CD and ulcerative colitis (Figure 4b).
a. Two sequences of the genus Blautia, that differ by nine bases over the 150bp Illumina read of the 16S rRNA V4 region are associated with opposite phenotypes, as discovered by dbBact. The two word clouds and annotations for each sequence, display “opposite” associations with disease. The left sequence is more prevalent in healthy subjects (“good” Blautia), whereas the other is highly abundant in a series of disease-related annotations. Such differences can be traced through dbBact, but are completely missed by a taxonomy-based analysis. b. The number of disease-related annotations for the two Blautia sequences across dbBact displays an opposite trend of being low and high in disease, for the “good” and “bad” Blautia, respectively. The total number of annotations in dbBact 2021.05 associated with the “good” and “bad” Blautia sequences is 377 and 124, respectively.
Collecting all “disease” related dbBact annotations shows that the “bad” Blautia is “HIGHER IN” in the disease group (compared to controls) in 8/9 disease annotations associated with it, whereas the “good” Blautia is “LOWER IN” in the disease group (compared to controls) in 22/24 disease annotations (Figure 4c). Therefore, sequence-based analysis provides a solid genotype-to-phenotype association compared to taxonomy-based associations.
dbBact provides a pan-microbiome view: Detailed example
dbBact may add another layer to data analysis in microbiome studies by identifying commonalities between different conditions and diseases, generating novel biological hypotheses. To demonstrate such a pan-microbiome analysis, we use data from a study comparing subjects consuming an American diet to a calorie restricted diet (49), and demonstrate the use of dbBact term enrichment. Fecal samples from two groups of lean individuals (BMI<25) who followed either an American diet (AMER) or a caloric restriction diet (CR) were selected. Standard analysis with FDR set to 0.1 (see “standard analysis” section in Methods) identified 28 and 141 bacterial sequences significantly more abundant in AMER and CR cohorts, respectively (Figure 3a). For clarity, we refer to these groups of sequences as S-AMER and S-CR, respectively. Figure 3b shows the internal “transformation” performed by dbBact from a heatmap of bacterial abundances to a heatmap of association scores of terms for each sequence (columns in Figure 3a-b are aligned and correspond to the same sequences). For example, the term “high BMI,” appears in almost all S-CR sequences, while it is almost absent in S-AMER sequences. These association scores in Figure 3b are then used as input to a non-parametric differential abundance test (see Methods section “statistical analysis in dbBact”), identifying terms significantly enriched in each of the two sequence groups (Figure 3c). Results indicate that sequences in the S-CR group are associated with terms related to low BMI (“low bmi,” “LOWER IN high bmi”) and with rural/undeveloped habitats (“LOWER IN united states of america,” “small village”) (see Supplementary File 4 for the full list of enriched terms). By contrast, bacteria from the S-AMER group have a significantly higher number of annotations related to high BMI (“LOWER IN low bmi,” “high bmi”) and urban/modernized habitats (“state of oklahoma,” “LOWER IN rural community,” “LOWER IN small village”).
Thus, although participants from both diet groups were lean, certain aspects of the underlying microbiome were associated with high and low BMI bacteria, for AMER and CR, respectively. Additionally, bacteria enriched in CR vs. AMER tend to be associated with rural/undeveloped habitats, which may indicate an adaptation of some bacteria found in rural communities to a low-calorie/higher vegetable diet content.
To further confirm the relationship between diet and BMI, we collected all sequences across dbBact having a “high bmi” annotation, resulting in 319 sequences. The overlap between these sequences and the S-AMER and S-CR groups is shown in Figure 3d (left). Although 93% (26/28) of S-AMER sequences overlap with “high bmi”-associated sequences, the overlap of S-CR sequences is 11% (15/141), i.e., a much larger fraction of S-AMER sequences is associated with high BMI. An analogous Venn diagram for the term “low bmi” displays an overlap of 54% (15/28) and 78% (110/141) of S-AMER and S-CRON bacteria with “low bmi”-associated bacteria in dbBact, respectively (Figure 3d left). As participants from both CR and AMER groups were lean (BMI<25), one may hypothesize that the effect of BMI on the microbial composition, observed in various studies, is due to dietary differences rather than the high BMI phenotype.
Remark regarding spurious/irrelevant terms
dbBact release 2022.07.01 contains annotations of approximately 6500 unique terms, some appearing only in a few experiments. As a result, word clouds and bar plots may often include seemingly odd terms. For example, the term “state of oklahoma” in Figure 5d is significantly enriched in S-AMER, a fact that seems implausible. This term appears only in two dbBact experiments, one of which compared a rural community in Peru to an urban community in Oklahoma (50). Hence, annotations from this experiment mentioned the term “state of oklahoma” together with more relevant terms (e.g., “rural community”) which, in turn, caused its inclusion. As dbBact continues to grow, such “transient” irrelevant inductions are expected to diminish.
a. Heatmap displaying bacterial abundances across fecal samples (rows) of low BMI individuals (BMI<25) practicing either a caloric restriction diet (CR, n=33) or an American diet (AMER, n=66), over a set of sequences (columns) that are significantly higher in either group. A differential abundance test (rank-mean test with dsFDR=0.1 multiple hypothesis correction) identified 136 bacteria higher in the CR group (S-CR) and 27 bacteria higher in the AMER group (S-AMER). b. dbBact terms (rows) enriched in the sequences appearing in panel a (columns in panels a and b are aligned). Heatmap values indicate the term score for each bacterium. Terms were identified using a non-parametric rank mean difference test with dsFDR=0.1 (top 6 terms for each direction are shown; see Supplementary File 4 for full list of enriched terms). c. Summary of the top enriched terms in the CR and AMER diets (green and red bars, respectively). Bar length and numbers are as in Figure 2. d. Venn diagrams of dbBact annotations related to the terms “low bmi” (right) and “high bmi” (left). Green and red circles indicate the number of sequences associated with the term in the CR and AMER diets, respectively; the blue circle indicates the number of such sequences across dbBact as a whole. The intersections of “low bmi” bacteria with the CR group are significantly higher (p=7E-5, using two-sided Fisher’s exact test), confirming the association. Similarly, the intersection of “high bmi” annotated sequences across dbBact with the AMER group is significantly higher than that with the CR group (p=3E-17, using two-sided Fisher’s exact test).
Integrating dbBact into microbiome analysis pipelines allows generating novel biological hypotheses
To demonstrate how dbBact may be incorporated into microbiome analysis pipelines, we chose 16 dbBact experiments, excluded them from dbBact, then analyzed their results in the same way researchers would look at their own studies. In each of these studies, dbBact provided novel hypotheses that did not appear in the original paper and could not be formulated by standard methods.
The experiments presented here were chosen almost arbitrarily to provide examples from different habitats and niches: the human host, animals, and environmental samples (Figure 6a). dbBact-derived hypotheses may be divided into several “types,” as follows.
a. Summary of biological hypotheses derived from dbBact-based analysis of published studies. Details of each analysis are given in the corresponding Supplementary Results section. Row colors correspond to hypothesis “type” (inter-host similarities – green; intra-host similarities – blue; inter-disease similarities – gray; environmental sources – brown; contamination detection – red). b-e. Analysis results related to conclusions shown in panel a. b. dbBact term word cloud for sequences found in sea otter oral samples shows resemblance to dogs’ and cats’ samples. c. Venn diagram showing number of dbBact sequences associated with the term “monkey” across dbBact (blue), and their intersection with sequences found in individuals from the American Gut study, who consume a high (green) and low (red) number of fruits per week. Sequences in the high-fruit consumption group are significantly more associated with the term “monkey” (Fisher’s exact test p-value < 0.00001). d. dbBact term enrichment comparing water samples collected in Hunts Point and Soundview Park, along the Bronx River in New York. Sequences higher in Hunts Point (green, located upstream) show significant fresh-water-related term enrichment (dsFDR=0.1). e. Term-based PCA of fecal samples of one individual collected daily for one year. The first principal component is the “feces-skin” axis, where higher values correspond to “skin” (see Methods for details). The values of a subset of samples, shown in magenta, is high, indicating possible skin-derived contamination in these fecal samples.
Detecting inter-host similarities
dbBact can identify unexpected similarities in microbial populations across hosts. (i) For example, when examining the oral microbiome of wild sea otters (51), dbBact indicates a high similarity to the microbiome of the oral cavity of dogs and cats (Figure 6b and Figure S3). (ii) In another example, fecal bacteria of Himalayan Griffons (52) are found to be similar to those of another carrion feeder, the California Condor (Figure S4). (iii) Such inter-host similarities are also observed for disease-related bacteria. Examining bacteria in colitis in horses (53), dbBact detects an enrichment of human-associated bacteria, indicating a possible colonization by bacteria that are less host-specific (Figure S5). (iv) Another recent meta-analysis of various human diseases identified shared disease-related bacteria in multiple diseases (54). When examining non-human-related annotations, dbBact finds these bacteria to be enriched in non-primate, homeothermic animals (mouse, horse, rat, chicken). By contrast, health-related bacteria found in this study are enriched in monkey-associated terms (Figure S6). This may indicate the disappearance of host-specific bacteria in multiple diseases, together with the appearance of more generalist bacteria. (v) A similar enrichment in monkey-associated bacteria and rural-community related terms is observed in individuals from the American gut project (55) who report high consumption of fruits, compared to those reporting low consumption (Figure 6c and Figure S7).
Detecting intra-host similarities
dbBact can identify similarities within hosts. (i) For example, Scheithauer et al. (56) profiled the bacteria detected in the IgA-positive and IgA-negative fractions of fecal samples. dbBact-based analysis shows that the IgA-positive fraction is enriched in oral related terms, indicating a possible contribution of oral IgA to bacterial antibody coating (Figure S8). (ii) In another study (57), dbBact finds an enrichment in dentition-related terms in an oral rinse of adolescents with braces compared to an enrichment in soft-tissue-associated bacteria in those that do not wear braces (Figure S9). (iii) Such soft-tissue-associated bacteria are also observed when analyzing Yeoh et al. data (58) of tonsilitis patients (Figure S10).
Detecting inter-disease similarity
(i) Zhu et al. compared the fecal microbiome of acute pancreatitis patients with that of healthy controls (59). dbBact-based analysis hints at a common gut response between pancreatitis and diarrhea, and Crohn’s disease, i.e., a phenomenon of general dysbiosis formerly suggested by Duvallet et al. (60) (Figure S11). (ii) Giloteaux et al. (61) compared fecal samples of chronic fatigue syndrome patients with those of healthy controls. dbBact finds shared sequences between these patients and individuals who do little physical activity (Figure S12).
Detecting environmental sources
dbBact can detect the sources of bacterial communities. (i) For example, Lavrinienko et al. collected skin swabs of bank voles inside the uninhabited Chernobyl exclusion zone and outside the contaminated region in the outskirts of Kyiv (62). dbBact-based analysis shows an overrepresentation of soil- and plant-related bacteria inside the exclusion zone, while skin bacteria of bank voles near Kyiv were enriched in human and farm animal terms. This leads to the hypothesis that the difference between the two sample groups is due to contact with humans and farm animals rather than to exposure to radioactivity (Figure S13). (ii) Similarly, Risely et al. (63) observed strong diurnal oscillations in the microbiome composition of South African wild meerkats’ fecal samples. dbBact-based analysis indicates that this effect is driven by a large number of soil/rhizosphere-related bacteria appearing in the afternoon fecal samples (Figure S14). (iii) dbBact analysis of air samples taken by Gat et al. (64) during clear days in Israel shows human farming as a source of air bacteria, compared to samples taken during a dust storm, which display desert and soil-associated bacteria (Figure S15). Hence, fecal bacteria from human and farm animals are airborne during ambient weather conditions, whereas dust storms bring over desert and soil associated bacteria. (iv) Finally, analysis of river water samples in two locations near the Bronx River estuary (65) shows that the difference in bacterial communities in these two locations is partially explained by ocean vs. freshwater bacteria, probably related to the salinity levels in the two sample locations (Figure 6d and Figure S16).
Contamination detection
dbBact allows the straightforward detection of potential contaminants. Each bacterium in a study may be assigned its best fitting dbBact term, thus “awkward” bacteria may be detected and discarded from downstream analysis. (i) Caporaso et al. (66) followed the oral, skin, and fecal microbiome of an individual using daily samples for a year. dbBact-based analysis detected a group of skin-associated sequences in a subset of fecal samples, indicating a potential contamination (Figure 6e and Figure S17). (ii) Similarly, in a dataset of infant nasopharyngeal samples (67), we observed a cluster of mouse-associated sequences (Figure S18), which may be attributed to a contamination or to low biomass kit-related bacteria. As these mouse-associated sequences are evenly spread across the sample types, they did not introduce a systemic bias in the authors’ results. But removal of the sequences before downstream analysis reduces inter-sample noise and increases the statistical power (Figure S18c,d).
Discussion
dbBact integrates 16S rRNA microbiome studies into a collaborative, coherent body of knowledge that facilitates pan-microbiome analysis of new studies using a rigorous statistical and algorithmic framework.
An important advantage of dbBact, compared to standard meta-analysis methods, is that the latter may suffer from the “streetlight effect” (68). For example, when examining the effect of fruit consumption in the American Gut experiment (Fig 6a), one might consider including other dietrelated studies in the meta-analysis. But this would miss the link between high fruit consumption and primate-associated bacteria. dbBact retrieves annotations from a wide range of sample types and habitats, providing additional and potentially unexpected insights into the biological roles of bacteria.
Terms in dbBact annotations are based on ontologies, providing a common language for phenotype description. The tree structure of ontologies facilitates the discovery of commonalities between bacteria in studies conducted under similar, albeit not identical, conditions. For example, data from Crohn’s disease and ulcerative colitis experiments may be combined based on their ontological “parent” term “inflammatory bowel disease.” Moreover, many “cross-sectional” questions may be asked and possibly answered using dbBact. For instance, what terms are similar with respect to their bacteria (e.g., are dogs more similar to cats or to wolves?), or are there connections between phylogeny and specific phenotypes (e.g., does genus X appear only in host Y or in geographic location Z?).
Apart from putting forth novel hypotheses, dbBact makes possible the detection of sources of bacterial groups. We recommend querying dbBact as a first step in any microbiome analysis (e.g., using the interactive heatmap of the dbBact-Calour module). Identifying relevant bacterial groups and their dbBact annotations fosters an initial understanding of biological processes, supporting better downstream analysis. dbBact also enables associating bacteria with a “candidate reagent contaminant” annotation. We have encountered numerous cases where examining bacteria in a study detected contaminations, e.g., bacteria having mostly “mus musculus” annotations, although samples were of human origin. Removing these sequences prior to downstream analysis can remove biases and increase the statistical power of the analysis.
The current coverage of dbBact is high in a large number of habitats, but many other habitats are still poorly covered (e.g., Figure 3e). Therefore, terms appearing in a small number of annotations may lead to dubious conclusions. For example, dbBact contains a single experiment originating from a scrubland environment. This experiment profiled the leaf microbiome of ivy plants, hence, querying a set of ivy leaf-related bacteria may result in the enrichment of both “ivy” and “scrubland” terms. Therefore, to avoid incorrect conclusions, users are advised to further examine the set of experiments associated with each enriched term. As the number and diversity of experiments in dbBact increase, such spurious terms are expected to be suppressed. Tens of microbiome studies are published weekly, but the dbBact team can process only a fraction. We expect the microbiome community to contribute to dbBact and help increase the number and diversity of uploaded studies.
dbBact may also be used in shotgun metagenomics studies. Whenever 16S rRNA sequences are inferred from shotgun data they may be submitted as queries or uploaded to dbBact. The “linking” mechanism for harmonizing studies from different variable regions enables shotgun and amplicon studies to be integrated into one coherent knowledge base. Similarly, studies using long read technologies (or synthetic long reads) also provide full-length 16S rRNA sequences, and thus can be integrated into dbBact in the same manner.
In sum, dbBact introduces a new “layer” of data analysis in microbiome studies. We believe that the scope and ontology-based structure of dbBact provides new means for studying core factors affecting bacterial communities, possibly answering questions that could not have otherwise been asked.
Methods
As dbBact is constantly growing in scope, and to facilitate reproducibility, the dbBact infrastructure described in this section and all analyses presented in the paper carried out using dbBact release 2022.07.01, available for download as part of the weekly snapshots at https://dbbact.org/download.
Implementation
Database
The dbBact database is stored as a SQL relational database (PostgreSQL 9.5.10). The database schema and detailed table descriptions are provided in Supplementary File 1 and Figures S5-6.
Ontologies
Table S1 presents ontologies available in dbBact release 2022.07.01. dbBact supports the addition of ontologies to allow more accurate annotations. When users provide terms that do not appear in any of these ontologies, a new term is automatically added to the generic dbBact ontology.
ASV sequences
Primers and trimming
dbBact uses exact prefix search for sequence identification, and therefore all sequences in dbBact are primer trimmed and originate from one of the supported 16S rRNA forward primers. For dbBact release 2022.07.01, the supported forward primers are V1-27F (AGAGTTTGATCMTGGCTCAGxxx), V3-341F (CCTACGGGNGGCWGCAGxxx), V4-515F (GTGCCAGCMGCCGCGGTAAxxx), where “xxx” denotes the beginning of the ASV sequence stored in dbBact. Although additional primers can be added to dbBact, the vast majority of 16S rRNA studies uses one of the three primers described. The minimum length of sequences uploaded to dbBact is 100bp. Upon upload, sequences are stored at their full length rather than being truncated to a fixed length. When submitting a query sequence, exact sequence matches are searched using length=min(query_sequence_length, database_sequence_length).
Taxonomy assignment
A python script runs daily to add taxonomy assignments to uploaded dbBact ASV sequences using RDP version 2.12 (69). Although taxonomy is not used in dbBact for analysis, it may be used for querying dbBact (e.g., retrieving annotations associated with bacteria of the genus Streptococcus).
Inter-region querying
dbBact supports the harmonization of microbiome studies performed using different protocols by inter-region linking. When submitting a sequence, dbBact uses the SILVA database of full length 16S rRNA genes (SILVA version 132, (34)) to identify sequences whose “footprint” in other variable regions matches the query. First, the SILVA sequences containing the query sequence are detected. Second, all dbBact sequences that match these SILVA sequences, in any region, are retrieved (i.e., Query(S)={T: T ∈ dbBact, ∃R ∈ SILVA so that ss(S, R) and ss(T, R)} where ss stands for “subsequence”). Querying is performed using the “wholeSeqIDsTable” table in the dbBact implementation. To enable fast queries, a daily script is run on new dbBact sequences, linking all sequences sharing a SILVA sequence. Such linking is performed only when querying dbBact, therefore new versions of SILVA or other full length 16S rRNA databases may be seamlessly applied. Currently, dbBact supports linking the V1, V3, and V4 forward primer reads; additional primers may be incorporated if needed.
Queries of different sequence length
Sequences uploaded to dbBact may vary in length depending on the sequencing platform and sequenced region. When adding new annotations, dbBact stores the full-length sequence of each ASV. For example, when two experiments provide information about the same bacterium using 150bp and 200bp reads, respectively, dbBact stores these sequences as separate entries and links each annotation to the corresponding sequence. Yet, when submitting a query using either sequence, dbBact retrieves annotations using exact match on the shortest common sequence, hence also retrieving annotations related to the other sequence.
dbBact interfaces
REST-API server
The dbBact REST-API server (http://api.dbbact.org) is implemented in Python 3.6, using Flask version 0.12/Gunicorn v19.9 to handle web queries, and psycopg2 version 2.7.1 for handling Postgres data queries. Full API documentation is available at http://api.dbbact.org/docs. Examples using the REST-API for querying are available at: https://github.com/amnona/dbbact-examples. The REST-API enables access to all dbBact functions. Querying dbBact or adding anonymous annotations does not require registration. Registration by username/password enables editing annotations submitted by the same user.
dbBact website
The dbBact website (http://dbbact.org) enables dbBact annotation retrieval based on ASVs, taxonomy, or ontology terms. Additionally, the website provides word-cloud generation and term enrichment analysis. The source code for the website as well as deployment instructions are available on the dbBact-website github page (https://github.com/amnona/dbbact-website).
dbBact-Calour interface
dbBact is integrated into the Calour microbiome analysis program (https://github.com/biocore/calour), using the dbBact-Calour module (https://github.com/amnona/dbbact-calour). Using this interface, users can both query dbBact regarding bacterial sequences, and add new annotations. The dbBact-Calour module provides dbBact annotation retrieval from the interactive Calour heatmap display, showing all annotations associated with the selected sequence. Additionally, the module enables GUI-based creation of new dbBact annotations for selected sequences, and performs term enrichment analysis, term-based PCA and word cloud generation. A Jupyter notebook tutorial is available at: http://biocore.github.io/calour/notebooks/microbiome_databases.html.
The module also works with EZCalour, the full GUI version of Calour, (https://github.com/amnona/EZCalour). A tutorial for dbBact enrichment analysis using EZCalour is available at: https://github.com/amnona/EZCalour/blob/master/using-ezcalour.pdf.
Qiime2 plugin
The q2-dbBact plugin (https://github.com/amnona/q2-dbbact) enables dbBact annotation-based analysis using the qiime2 framework (35). The interface provides term enrichment analysis for the output of various qiime2 differential abundance plugins (ANCOM (70), Songbird (71), ALDEx2 (72), DACOMP (73), or a rank-mean method). Additionally, the plugin supports dbBact term word cloud and interactive heatmap generation.
Data availability
Table 1 details the locations of different dbBact sites
Locations of dbBact sites
Standard analysis: Default dbBact preprocessing of an experiment
Although dbBact is a wiki-style knowledge base, the vast majority of annotations in dbBact release 2022.07.01 was added by the dbBact team. Studies were selected from published microbiome papers, and annotations were added following the re-processing of the experimental data, using a “standard” manual analysis pipeline as follows:
The raw data of each scientific paper (i.e., per-sample FASTA files and corresponding metadata) were downloaded using the provided accession (e.g., by SRA/ENA accession or Qiita (21) study ID). When data or metadata were not available, the authors were contacted and provided the missing data directly. When primer sequences were part of the reads, they were removed using a custom script (https://github.com/amnona/GetData). Subsequently, the Deblur pipeline (15) was applied to the reads of each sample (Deblur script version 1.1.0, using default parameters, https://github.com/biocore/deblur), resulting in a denoised biom table.
This biom table, together with the per-sample metadata, were manually re-analyzed using Calour (36), to add annotations capturing biological conclusions arising from the study. Three types of predicates were sought:
(i) “DIFFERENTIAL:” To detect sets of sequences associated with relevant conditions (e.g., sick vs. healthy), sequences significantly enriched between two conditions were identified using a non-parametric permutation based rank-mean test, followed by multiple hypothesis correction using dsFDR (74) (usually set to 0.1). The test was performed by the calour.diff_abundance() function. Correlations with continuous metadata fields (e.g., BMI) were detected with a permutation-based Spearman test with dsFDR correction using calour.correlation(). In both cases, the set of sequences higher or lower in one condition than in the other were then annotated as “DIFFERENTIAL,” i.e., “HIGHER IN” condition 1 and “LOWER IN” condition 2.
(ii) “COMMON”/“DOMINANT:” For each study, sequences present in more than half of the relevant samples were annotated as “COMMON.” “DOMINANT” sequences were identified as sequences whose mean frequency in the relevant samples was higher than 0.01. In studies containing samples from multiple sources (i.e., fecal and saliva samples, or samples from individuals from several countries or disease vs. healthy), a “COMMON”/“DOMINANT” annotation was added separately to each source subset.
(iii) “CONTAMINANT:” When a study contained a set of negative control (blank) samples, ASVs showing higher frequency in these controls than in the non-blank samples were manually annotated as a possible “CONTAMINANT.”
Examples of the different predicates appear in Table S3.
Remark
Using the abovementioned pipeline is not a prerequisite for adding new annotations, and any denoising method followed by statistical analyses can be applied by users contributing to dbBact.
Statistical analysis in dbBact
Word cloud generation
Calculating a term’s F1 score: Given a set of query sequences S, each dbBact term t is assigned an F1 score, corresponding to the harmonic mean of precision and recall. For each sequence s in S, we calculate the the fraction of s’s annotations that contain the term t. The average of these values across S provides the precision of t on S. Similarly, for a given sequence s and a term t, recall is calculated as the fraction of t’s annotations that contain s. To suppress terms that appear in a small number of experiments, the total number of t’s annotations (i.e., the denominator) is artificially increased by 1. The average of these values across S provides the recall of t on S.
Displaying a word cloud. The word cloud size of each term is proportional to its F1 score. If the term appears in “LOWER IN” annotations, its color is orange, otherwise it is blue. The brightness of each term represents the number of experiments containing the term, indicating the reliability of the term (white for a single experiment; ranging to dark blue/orange for >=10 experiments).
Word clouds were generated with the dbBact-Calour module for Calour (https://github.com/amnona/dbbact-calour) using the draw_wordcloud() function.
Term enrichment analysis
Given two sets of sequences, S1 and S2, we search for terms significantly enriched in either group using the following steps:
(a) Calculating an annotation-based score per sequence and term. Each annotation a in dbBact is assigned a “weight” w(a) according to its predicate. The predicates “COMMON,” “CONTAMINATION,” and “DIFFERENTIAL” are assigned a weight of 1, and the predicates “DOMINANT” and “OTHER” are assigned a weight of 2 and 0.5, respectively. These weights are applied to calculate a score Score(s, t) for each term t in dbBact and a sequence s in either S1 or S2. The score sums the weights of all annotations involving s and t. When t appears as “LOWER IN” in the predicate “DIFFERENTIAL,” a new term “not t” is created and assigned a weight of 1.
(b) Calculating effect size of a term. For a term t, Score(s, t) is calculated for all sequences in S1 and S2, and the effect size of t is defined as
where |S| corresponds to the number of sequences in the set S.
(c) Finding significant terms. Each term is assigned a p-value by comparing its scores over 1000 random permutations of the combined S1 and S2 sequences to sets of size |S1| and |S2|. Subsequently, a dsFDR multiple hypothesis correction (with a threshold of 0.1) is applied to detect significant terms.
(d) Calculating the significance of a term across experiments. Until this stage, we measured the enrichment of a term based on all dbBact experiments combined. To estimate whether such significance occurs across multiple experiments, or whether it is driven by a single or a few experiments, steps (a)-(c) were also repeated using each individual experiment that contains t. The total number of experiments containing each term and the fraction in which the term was significant appear in each figure.
The abovementioned analysis is performed using the dbbact-calour module enrichment() function.
Venn diagrams
Given two sets of sequences, S1 and S2, and a term t, we plot a Venn diagram indicating the number of sequences associated with t across all dbBact annotations, and the overlap of these sequences with S1 and S2.
Venn diagrams were generated using the dbbact-calour module plot_term_venn_all() function.
dbBact term-based principal component analysis
Each sample x is represented by a vector of sequence frequencies (abbreviated “sf”), x(S) = (sf1, sf2, …, sfn) across the n sequences that appear in all study samples (sf equals zero in case a sequence does not appear in a specific sample). We then transform x(S) into a term-based representation of x, i.e., x(T) = (ts1, ts2, …, tsm′), where each ts is a term-score described below, calculated across all m dbBact terms. The ts score for the term t is given by 
precision(i, t), where precision(i, t) is the fraction of annotations associated with sequence i that also contain the term t. Finally, once x(T) is calculated over all samples in a study, we perform principal component analysis of this space. Each principal axis is defined by its weights, where the highest weights (in absolute value) are used for providing biological meaning. The abovementioned analysis is performed using the dbbact-calour module plot_term_pcoa() function.
Processing of datasets
All datasets discussed in the paper were processed using the following pipeline: raw reads were downloaded and denoised using the Deblur pipeline (15) with default parameters; the resulting denoised biom table was loaded into Calour (36), and differentially abundant bacteria were identified using a permutation-based non-parametric rank mean test with dsFDR multiple hypothesis correction (74) set to 0.1 (using the calour diff_abundance() method). In the case of the American Gut Project dataset, multiple samples originating from the same individual were aggregated to a single sample using mean frequency for each ASV. The groups of high and low fruit consumption were controlled for confounders by stratifying samples in both groups based on the AGP metadata fields: age category (“AGE_CAT”), sex and BMI category (“BMI_CAT”), and randomly dropping samples to equalize the number of samples from each stratum prior to differential abundance testing. dbBact term word clouds were generated by applying the above-described word cloud approach using the dbbact-calour module draw_wordcloud() method on all sequences present in at least 30% of the samples. Term enrichment was performed using the above-described term enrichment approach using the dbbact-calour module enrichment() method with default parameters (dsFDR=0.1). dbBact term PCAs were generated using the dbbact-calour module plot_term_pcoa() function. Accession numbers for each dataset used are available in Supplementary File 2. Jupyter notebooks used for the creation of each figure are available at https://github.com/amnona/dbbact-paper.
Acknowledgements
We wish to thank Tzipi Brown and Rotem Hadar from the Sheba Microbiome Center, Zhenjiang (Zech) Xu, Jon Sanders, Qiyun Zhu, Tomasz Kosciolek, Stefan Janssen and Jeremiah Minich for fruitful discussions, feedback, and suggestions. N.S. is funded by the Ministry of Science, Technology & Space, Israel (Grant 3-16033).
Footnotes
We have added 16 examples reanalyzing published datasets, showing how dbBact leads to new biological insights.
