Abstract
H2 metabolism is the most ancient and diverse mechanism of energy-generation. 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]-hydrogenase subgroups, 8 [FeFe]-hydrogenase subtypes, [Fe]-hydrogenases). 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 also provides the capacity to browse 3248 annotated sequences and 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/
Introduction
Microorganisms conserve energy by metabolizing H2. 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. H2 metabolism contributes to the growth and survival of microorganisms across the three domains of life: chemotrophs and phototrophs, lithotrophs and heterotrophs, aerobes and anaerobes, mesophiles and extremophiles alike 1,2. On the ecosystem scale, H2 supports microbial communities in most terrestrial, aquatic, and host-associated ecosystems 1,3. It is also generally accepted that H2 was the primordial electron donor 4. In biological systems, metalloenzymes known as hydrogenases are responsible for oxidizing and evolving H2 1,5. Our recent survey showed there is a far greater number and diversity of hydrogenases than previously thought 2. It is predicted over 55 microbial phyla and up to half of all microorganisms harbor hydrogenases 2,6. Better understanding H2 metabolism and the enzymes that mediate it also has wider implications, particularly in relation to human health and disease 3,7, biogeochemical cycling 8, and renewable energy 9,10.
There are three classes 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 11, the [NiFe] and [FeFe] classes are widely distributed and functionally diverse. They comprise numerous 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) 1,5. For example, while Group 2a and 2b [NiFe]-hydrogenases share > 35% sequence identity, they have distinct roles as respiratory uptake hydrogenases and H2 sensors respectively 12,13. Building on previous work 14,15, we recently created a comprehensive hydrogenase classification scheme predictive of biological function 2. This scheme was primarily based on amino acid sequence phylogeny, but 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 with unique physiological roles 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. This tool is a more reliable, efficient, and user-friendly method for hydrogenase classification than standard approaches involved phylogenetic tree construction, with a precision of more than 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 16, in which nodes represent individual proteins and the distances between them reflect BLAST E-values. As reflected by our analysis of other protein superfamilies 17,18, 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 2,14,15, 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 (Figure 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 2,15, cannot be classified by sequence alone as they have principally diversified through changes in domain architecture and quaternary structure. It remains strictly 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 groups and subgroups that clustered together in the phylogenetic tree analysis 2 separate into several subclades of probable distinct function (Figure 1). On this basis, we refined and expanded the hydrogenase classification scheme to reflect the sequence diversification observed (Table 1). Three lineages originally classified as Group 1a [NiFe]-hydrogenases were reclassified as new subgroups, the Coriobacteria (Group 1 i), Archaeoglobi (Group 1j), and Methanosarcinales (Group 1i). The previously-defined 4b and 4d subgroups 2 were dissolved, as the SSN analysis confirmed they were highly polyphyletic. These sequences are reclassified here into five new subgroups: the formate-and carbon monoxide-respiring Mrp-linked complexes (Group 4b) 19, the ferredoxin-coupled Mrp-linked complexes (Group 4d) 20, the well-described methanogenic Eha (Group 4h) and Ehb (Group 4i) supercomplexes 21, and a more loosely clustered class of unknown function (Group 4g). Three crenarchaeotal hydrogenases were also classified as their own family (Group 2e); these enzymes enable certain crenarchaeotes to grow aerobically on O2 22,23 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 (Figure 1); these enzymes likely have a sensory role 2,15 and are each co-transcribed with different regulatory elements (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 two-step process following input of the catalytic subunit sequence. In the first, the Conserved Domain Database (CDD) 24 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). The sequence is subsequently 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 accuracy for each k. The results are shown in Figure 2. The classifier predicted the classes of the 3248 hydrogenase sequences with 99.8% accuracy 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 S1. The classifier has also been trained to detect and exclude protein families that are homologous to hydrogenases but do not metabolize H2 (Nuo, Ehr, NARF, HmdII 1,2) using reference sequences of these proteins.
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 accuracy, the classifier is superior to other approaches due to its usability (Figure S2). 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. When analysis has completed, results are presented in a table that can be downloaded as a CSV file. This provides an efficient and user-friendly way to classify hydrogenases, in contrast to the previous standard which requires visualization of multiple sequence alignments in phylogenetic trees 25.
HydDB infers the physiological roles of H2 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 H2 metabolism were predicted (Table 2). For five of the organisms, these predictions are confirmed or supported by previously published data 23,26–⇓⇓29. Other predictions are in line with metabolic models derived from metagenome surveying 30–⇓32. In some cases, the capacity for organisms to metabolize H2 was not tested or inferred in previous studies despite the presence of hydrogenases in the sequenced genomes 27,33–⇓35.
While HydDB serves as a reliable initial predictor of hydrogenase class and function, further analysis is recommended to verify predictions. Hydrogenase sequences only provide organism with the genetic capacity to metabolise H2; their function is ultimately modulated by their expression and integration within the cell 1,36. 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 H2-scavenging functions as the core group 8. Likewise, it seems probable that the Group 3a [NiFe]-hydrogenases of Thermococci and Aquificae use a distinct electron donor to the main class 37. 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 sequence, hydrogenase class, taxonomic affiliation, and predicted behavior (Figure S2). 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 1000 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 provide a deep but accessible resource for understanding hydrogenases.
Materials and 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 bacteria phyla were retrieved from the Joint Genome Institute's Integrated Microbial Genomes database 38.
Sequence similarity networks
Sequence similarity networks (SSNs) 16 were used to visualize the distribution and diversity of the 3248 retrieved hydrogenase sequences. In this analysis, nodes represent individual proteins and edges represent the all-versus-all BLAST E-values. Three networks were constructed using Cytoscape, namely for the [NiFe]-hydrogenase large subunit sequences, [FeFe]-hydrogenase catalytic domain sequences, and [Fe]-hydrogenase sequences. The relationships between them were viewed at different logE cutoffs using different subsets of sequences.
Classification method
The k-NN method is a well-known machine learning method for classification 39. Given a set of data points x1,x2, …xN (e.g. sequences) with known labels y1,y2, …,yN (e.g. type annotations), the label of a point, x, is predicted by computing the distance from x to x1,x2, …xN 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 accuracy of the method and perform model selection. The classifier was evaluated using fc-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 an accuracy can be computed.
Author Contributions
CG and DS designed experiments. DS and CG performed experiments. CG, DS, and CNSP analysed data. CNSP supervised students. CG and DS wrote the paper.
The authors declare no conflict of interest.
Acknowledgements
We thank A/Prof Colin J. Jackson, Dr Hafna Ahmed, Dr Andrew Warden, and Dr Stephen Pearce for their helpful advice and comments regarding this manuscript. This work was supported by a PUMPkin Centre of Excellence PhD Scholarship awarded to DS, an Australian National University PhD Scholarship awarded to FHA, and a CSIRO Office of the Chief Executive Postdoctoral Fellowship awarded to CG.