HydDB: A web tool for hydrogenase classification and analysis

H2 metabolism is proposed to be the most ancient and diverse mechanism of energy-conservation. The metalloenzymes mediating this metabolism, hydrogenases, are encoded by over 60 microbial phyla and are present in all major ecosystems. We developed a classification system and web tool, HydDB, for the structural and functional analysis of these enzymes. We show that hydrogenase function can be predicted by primary sequence alone using an expanded classification scheme (comprising 29 [NiFe], 8 [FeFe], and 1 [Fe] hydrogenase classes) that defines 11 new classes with distinct biological functions. Using this scheme, we built a web tool that rapidly and reliably classifies hydrogenase primary sequences using a combination of k-nearest neighbors’ algorithms and CDD referencing. Demonstrating its capacity, the tool reliably predicted hydrogenase content and function in 12 newly-sequenced bacteria, archaea, and eukaryotes. HydDB provides the capacity to browse the amino acid sequences of 3248 annotated hydrogenase catalytic subunits and also contains a detailed repository of physiological, biochemical, and structural information about the 38 hydrogenase classes defined here. The database and classifier are freely and publicly available at http://services.birc.au.dk/hyddb/

Microorganisms conserve energy by metabolizing H 2 . Oxidation of this high-energy fuel yields electrons that can be used for respiration and carbon-fixation. This diffusible gas is also produced in diverse fermentation and anaerobic respiratory processes 1 . H 2 metabolism contributes to the growth and survival of microorganisms across the three domains of life, including chemotrophs and phototrophs, lithotrophs and heterotrophs, aerobes and anaerobes, mesophiles and extremophiles alike 1,2 . On the ecosystem scale, H 2 supports microbial communities in most terrestrial, aquatic, and host-associated ecosystems 1,3 . It is also proposed that H 2 was the primordial electron donor 4,5 . In biological systems, metalloenzymes known as hydrogenases are responsible for oxidizing and evolving H 2 1,6 . Our recent survey showed there is a far greater number and diversity of hydrogenases than previously thought 2 . It is predicted that over 55 microbial phyla and over a third of all microorganisms harbor hydrogenases 2,7 . Better understanding H 2 metabolism and the enzymes that mediate it also has wider implications, particularly in relation to human health and disease 3,8 , biogeochemical cycling 9 , and renewable energy 10,11 .
There are three types of hydrogenase, the [NiFe], [FeFe], and [Fe] hydrogenases, that are distinguished by their metal composition. Whereas the [Fe]-hydrogenases are a small methanogenic-specific family 12 , the [NiFe] and [FeFe] classes are widely distributed and functionally diverse. They can be classified through a hierarchical system into different groups and subgroups/subtypes with distinct biochemical features (e.g. directionality, affinity, redox partners, and localization) and physiological roles (i.e. respiration, fermentation, bifurcation, sensing) 1,6 . It is necessary to define the subgroup or subtype of the hydrogenase to predict hydrogenase function. For example, while Group 2a and 2b [NiFe]-hydrogenases share > 35% sequence identity, they have distinct roles as respiratory uptake hydrogenases and H 2 sensors respectively 13,14 . Likewise, discrimination between Group A1 and Group A3 [FeFe]-hydrogenases is necessary to distinguish fermentative and bifurcating enzymes 2,15 . Building on previous work 16,17 , we recently created a comprehensive hydrogenase classification scheme predictive of biological function 2 . This scheme was primarily based on the topology of phylogenetic trees built from the amino acid sequences of hydrogenase catalytic subunits/domains. It also factored in genetic organization, metal-binding motifs, and functional information. This analysis identified 22 subgroups (within four groups) of [NiFe]-hydrogenases and six subtypes (within three groups) of [FeFe]-hydrogenases, each proposed to have unique physiological roles and contexts 2 .
In this work, we build on these findings to develop the first web database for the classification and analysis of hydrogenases. We developed an expanded classification scheme that captures the full sequence diversity of hydrogenase enzymes and predicts their biological function. Using this information, we developed a classification tool based on the k-nearest neighbors' (k-NN) method. HydDB is a user-friendly, high-throughput, and functionally-predictive tool for hydrogenase classification that operates with precision exceeding 99.8%.

Results and Discussion
A sequence-based classification scheme for hydrogenases. We initially developed a classification scheme to enable prediction of hydrogenase function by primary sequence alone. To do this, we visualized the relationships between all hydrogenases in sequence similarity networks (SSN) 18 , in which nodes represent individual proteins and the distances between them reflect BLAST E-values. As reflected by our analysis of other protein superfamilies 19,20 , SSNs allow robust inference of sequence-structure-function relationships for large datasets without the problems associated with phylogenetic trees (e.g. long-branch attraction). Consistent with previous phylogenetic analyses 2,16,17 , this analysis showed the hydrogenase sequences clustered into eight major groups (Groups 1 to 4 [NiFe]-hydrogenases, Groups A to C [FeFe]-hydrogenases, [Fe]-hydrogenases), six of which separate into multiple functionally-distinct subgroups or subtypes at narrower logE filters ( Fig. 1; Figure S1). The SSNs demonstrated that all [NiFe]-hydrogenase subgroups defined through phylogenetic trees in our previous work 2 separated into distinct clusters, which is consistent with our evolutionary model that such hydrogenases diverged from a common ancestor to adopt multiple distinct functions 2 . The only exception were the Group A [FeFe]hydrogenases, which, as previously-reported 2,17 , cannot be classified by sequence alone as they have principally diversified through changes in domain architecture and quaternary structure. It remains necessary to analyze the organization of the genes encoding these enzymes to determine their specific function, e.g. whether they serve fermentative or electron-bifurcating roles.
The SSN analysis revealed that several branches that clustered together on the phylogenetic tree analysis 2 in fact separate into several well-resolved subclades (Fig. 1). We determined whether this was significant by analyzing the taxonomic distribution, genetic organization, metal-binding sites, and reported biochemical or functional characteristics of the differentiated subclades. On this basis, we concluded that 11 of the new subclades identified are likely to have unique physiological roles. We therefore refine and expand the hydrogenase classification to reflect the hydrogenases are more diverse in both primary sequence and predicted function than accounted for by even the latest classification scheme 2  Three lineages originally classified as Group 1a [NiFe]-hydrogenases were reclassified as new subgroups, namely those affiliated with Coriobacteria (Group 1i), Archaeoglobi (Group 1j), and Methanosarcinales (Group 1i). Cellular and molecular studies show these enzymes all support anaerobic respiration of H 2 , but differ in the membrane carriers (methanophenazine, menaquinone) and terminal electron acceptors (heterodisulfide, sulfate, nitrate) that they couple to 21,22 . The previously-proposed 4b and 4d subgroups 2 were dissolved, as the SSN analysis confirmed they were polyphyletic. These sequences are reclassified here into five new subgroups: the formateand carbon monoxide-respiring Mrp-linked complexes (Group 4b) 23 , the ferredoxin-coupled Mrp-linked complexes (Group 4d) 24 , the well-described methanogenic Eha (Group 4h) and Ehb (Group 4i) supercomplexes 25 , and a more loosely clustered class of unknown function (Group 4g). Enzymes within these subgroups, with the exception of the uncharacterized 4g enzymes, sustain well-described specialist functions in the energetics of various archaea [23][24][25] . Three crenarchaeotal hydrogenases were also classified as their own family (Group 2e); these enzymes enable certain crenarchaeotes to grow aerobically on O 2 26,27 and hence may represent a unique lineage of aerobic uptake hydrogenases currently underrepresented in genome databases. The Group C [FeFe]-hydrogenases were also separated into three main subtypes given they separate into distinct clusters even at relatively broad logE values ( Fig. 1); these subtypes are each transcribed with different regulatory elements and are likely to have distinct regulatory roles 2,17,28 (Table 1).

HydDB reliably predicts hydrogenase class using the k-NN method and CDD referencing.
Using this information, we built a web tool to classify hydrogenases. Hydrogenase classification is determined through a three-step process following input of the catalytic subunit sequence. Two checks are initially performed to confirm if the inputted sequence is likely to encode a hydrogenase catalytic subunit/domain. The Conserved Domain Database (CDD) 29 is referenced to confirm that the inputted sequence has a hydrogenase catalytic domain, i.e. "Complex1_49kDa superfamily" (cl21493) (for NiFe-hydrogenases), "Fe_hyd_lg_C superfamily" (cl14953) (for FeFe-hydrogenases), and "HMD" (pfam03201) (for Fe-hydrogenases). A homology check is also performed that computes the BLAST E-value between the inputted sequence and its closest homolog in HydDB. HydDB classifies any inputted sequence that lacks hydrogenase conserved domains or has low homology scores (E-value > 10 −5 ) as a non-hydrogenase (Table S1).
In the final step, the sequence is classified through the k-NN method that determines the most similar sequences listed in the HydDB reference database. To determine the optimal k for the dataset, we performed a 5-fold cross-validation for k = 1… 10 and computed the precision for each k. The results are shown in Fig. 2. The classifier predicted the classes of the 3248 hydrogenase sequences with 99.8% precision and high robustness when performing a 5-fold cross-validation (as described in the Methods section) for k = 4. The six sequences where there were discrepancies between the SSN and k-NN predictions are shown in Table S2. The classifier has also been trained to detect and exclude protein families that are homologous to hydrogenases but do not metabolize H 2 (Nuo, Ehr, NARF, HmdII 1,2 ) using reference sequences of these proteins (Table S1).
Sequences of the [FeFe] Group A can be classified into functionally-distinct subtypes (A1, A2, A3, A4) based on genetic organization 2 . The classifier can classify such hydrogenases if the protein sequence immediately downstream from the catalytic subunit sequence is provided. The classifier references the CDD to search for conserved domains in the downstream protein sequence. A sequence is classified as [FeFe] Group A2 if one of the domains "GltA", "GltD", "glutamate synthase small subunit" or "putative oxidoreductase", but not "NuoF", is found in the sequence. Sequences are classified as [FeFe] Group A3 if the domain "NuoF" is found and [FeFe] Group A4 if the domain "HycB" is present. If none of the domains are found, the sequence is classified as A1. These classification rules were determined by collecting 69 downstream protein sequences. The sequences were then submitted to the CDD and the domains which most often occurred in each subtype were extracted.
In addition to its precision, the classifier is superior to other approaches due to its usability. It is accessible as a free web service at http://services.birc.au.dk/hyddb/ HydDB allows the users to paste or upload sequences of hydrogenase catalytic subunit sequences in FASTA format and run the classification ( Figure S2). When analysis has completed, results are presented in a table that can be downloaded as a CSV file ( Figure S3). This provides an efficient and user-friendly way to classify hydrogenases, in contrast to the previous standard which requires visualization of phylogenetic trees derived from multiple sequence alignments 30 .  Figure S1 shows the further delineation of the encircled [NiFe] hydrogenase classes.  (Fig. 1). HydDB contains detailed information on each of these classes, including their taxonomic distribution, genetic organization, biochemistry, and structures, as well a list of primary references. HydDB infers the physiological roles of H 2 metabolism. As summarized in Table 1, hydrogenase class is strongly correlated with physiological role. As a result, the classifier is capable of predicting both the class and function of a sequenced hydrogenase. To demonstrate this capacity, we used HydDB to analyze the hydrogenases present in 12 newly-sequenced bacteria, archaea, and eukaryotes of major ecological significance. The classifier correctly classified all 24 hydrogenases identified in the sequenced genomes, as validated with SSNs (Table 2). On the basis of these classifications, the physiological roles of H 2 metabolism were predicted ( Table 2). For five of the organisms, these predictions are confirmed or supported by previously published data 27,[31][32][33][34] . Other predictions are in line with metabolic models derived from metagenome surveying [35][36][37] . In some cases, the capacity for organisms to metabolize H 2 was not tested or inferred in previous studies despite the presence of hydrogenases in the sequenced genomes 32,38-40 . While HydDB serves as a reliable initial predictor of hydrogenase class and function, further analysis is recommended to verify predictions. Hydrogenase sequences only provide organisms with the genetic capacity to metabolise H 2 ; their function is ultimately modulated by their expression and integration within the cell 1,41 . In addition, some classifications are likely to be overgeneralized due to lack of functional and biochemical characterization of certain lineages and sublineages. For example, it is not clear if two distant members of the Group 1h [NiFe]-hydrogenases (Robiginitalea biformata, Sulfolobus islandicus) perform the same H 2 -scavenging functions as the core group 9 . Likewise, it seems probable that the Group 3a [NiFe]-hydrogenases of Thermococci and Aquificae use a distinct electron donor to the main class 42 . Prominent cautions are included in the enzyme pages in cases such as these. HydDB will be updated when literature is published that influences functional assignments.
HydDB contains interfaces for hydrogenase browsing and analyzing. In addition to its classification function, HydDB is designed to be a definitive repository for hydrogenase retrieval and analysis. The database presently contains entries for 3248 hydrogenases, including their NCBI accession numbers, amino acid sequences, hydrogenase classes, taxonomic affiliations, and predicted behavior ( Figure S4). To enable easy exploration of the data set, the database also provides access to an interface for searching, filtering, and sorting the data, as well as the capacity to download the results in CSV or FASTA format. There are individual pages for the 38 hydrogenase classes defined here (Table 1), including descriptions of their physiological role, genetic organization, taxonomic distribution, and biochemical features. This is supplemented with a compendium of structural information about the hydrogenases, which is integrated with the Protein Databank (PDB), as well as a library of over 500 literature references ( Figure S5).

Conclusions
To summarize, HydDB is a definitive resource for hydrogenase classification and analysis. The classifier described here provides a reliable, efficient, and convenient tool for hydrogenase classification and functional prediction. HydDB also provides browsing tools for the rapid analysis and retrieval of hydrogenase sequences. Finally, the manually-curated repository of class descriptions, hydrogenase structures, and literature references provides a deep but accessible resource for understanding hydrogenases.

Methods
Sequence datasets. The database was constructed using the amino acid sequences of all curated non-redundant 3248 hydrogenase catalytic subunits represented in the NCBI RefSeq database in August 2014 2 (Dataset S1). In order to test the classification tool, additional sequences from newly-sequenced archaeal and bacterial phyla were retrieved from the Joint Genome Institute's Integrated Microbial Genomes database 43 . The mean precision ± two standard deviations of the folds is shown in the figure (note the y-axis). k = 1 provides the most accurate classifier. However, k = 4 provides almost the same precision and is more robust to errors in the training set (reflected by the lower standard deviation). In general, the standard deviation is very small, indicating that the predictions are robust to changes in the training data.
Scientific RepoRts | 6:34212 | DOI: 10.1038/srep34212 Sequence similarity networks. Sequence similarity networks (SSNs) 18 constructed using Cytoscape 4.1 44 were used to visualize the distribution and diversity of the retrieved hydrogenase sequences. In this analysis, each node represents one of the 3248 hydrogenase sequences in the reference database (Dataset S1). Each edge represents the sequence similarity between them as determined by E-values from all-vs-all BLAST analysis, with all self and duplicate edges removed. Three networks were constructed, namely for the [NiFe]-hydrogenase large subunit sequences (Dataset S2), [FeFe]-hydrogenase catalytic domain sequences (Dataset S3), and [Fe]-hydrogenase sequences (Dataset S4). To control the degree of separation between nodes, logE cutoffs that were incrementally decreased from − 5 to − 200 until no major changes in clustering was observed. The logE cutoffs used for the final classifications are shown in Fig. 1 and Figure S1. Classification method. The k-NN method is a well-known machine learning method for classification 45 .
Given a set of data points x 1 , x 2 , … x N (e.g. sequences) with known labels y 1 , y 2 , … , y N (e.g. type annotations), the label of a point, x, is predicted by computing the distance from x to x 1 , x 2 , … x N and extracting the k labeled points closest to x, i.e. the neighbors. The predicted label is then determined by majority vote of the labels of the neighbors. The distance measure applied here is that of a BLAST search. Thus, the classifier corresponds to a homology search where the types of the top k results are considered. However, formulating the classification method as a machine learning problem allows the use of common evaluation methods to estimate the precision of the method and perform model selection. The classifier was evaluated using k-fold cross-validation. The dataset is first split in to k parts of equal size. − k 1 parts (the training set) are then used for training the classifier and the labels of the data points in the remaining part (the test set) are then predicted. This process, called a fold, is repeated k times. The predicted labels of each fold are then compared to the known labels and a precision can be computed.  Table 2. Predictive capacity of the HydDB. HydDB accurately determined hydrogenase content and predicted the physiological roles of H 2 metabolism in 12 newly-sequenced archaeal and bacterial species.