BioHackathon series in 2013 and 2014: improvements of semantic interoperability in life science data and services

Ontology for locations on biological sequences

During the BioHackathon 2012⁴, it was recognized that a common schema ontology was desirable for the Semantic Web integration of sequence annotation across multiple databases. In depth group discussions including bioinformatics software developers and major database representatives identified common core needs in defining locations on biological sequences (both nucleic acids and proteins). This produced a draft specification for the Feature Annotation Location Description Ontology (FALDO), and proof of principle data conversion tools. This work continued at the BioHackathon 2013, with a specific focus on ensuring that all the existing annotations in the International Nucleotide Sequence Database Collaboration (INSDC)¹² feature tables could be converted into RDF triples using FALDO, as well as standardizing the coordinate system, and making sure that the starts of features are biologically sensible i.e. the start value is numerically higher than the end for genes located on the reverse strand. Subsequently, in May 2014, DBCLS organized a closed meeting, the RDF Summit, where a small group of developers from DBCLS, DNA Data Bank of Japan (DDBJ), Swiss Institute of Bioinformatics (SIB), European Bioinformatics Institute (EBI) and Stanford gathered to standardize the RDF representation of genomic annotations. The group agreed to use the FALDO ontology (see the section below) for annotating the coordinates of genomic annotations and to represent gene/transcript/exons in RDF. As a result, the RDF model of DDBJ, Ensembl¹³ and TogoGenome⁹ are now aligned such that common SPARQL queries can retrieve sequence annotations from these distinct data sources interoperably.

Human genome and variation

After defining a common RDF model to represent the INSDC feature tables, one of the major remaining needs was to standardize the RDF representation of genome variations, which was discussed during the BioHackathon 2014.

A group from EBI, DBCLS and Tohoku University surveyed existing databases that represent clinical annotation of variants. National Center for Biotechnology Information (NCBI) ClinVar¹⁴ provides information on the relationships between human genetic variation and phenotypes along with supporting evidence; Online Mendelian Inheritance in Man (OMIM)¹⁵ provides relationships between genes and disease; Leiden Open Variant Database (LOVD)¹⁶ provides gene variants related to colon cancer; Human Gene Mutation Database (HGMD)¹⁷ is commercial but widely used; Thomson Reuters Gene Variant Database (GVDB) is also a commercial database. Tohoku Medical Megabank had a license to jointly develop the RDF version of the GVDB with Thomson Reuters and they completed the initial version to test queries like "find shared variations among diseases" and "find related variations from a specific disease". In parallel, the EBI group started to convert Ensembl variation data into RDF in which an "allele" is related to "gene_variant", "sequence_alteration" and "regulatory_region_variant" instances in the sequence ontology (SO), and its location is represented by means of a FALDO region [Figure 1].

Figure 1. Proposed schema for the Ensembl variation.

The H-invitational database (H-InvDB)^18,19 group developed RDF data and an ontology for their database covering ncRNA annotations. During the Biohackathon 2013, the RDF version of the H-InvDB was expanded and its ontology was published including recent advancement in understanding of non-coding RNA (ncRNA) function. To improve descriptions of the functional relationships between coding transcripts and ncRNA, links between transcripts in H-InvDB and two major RNA databases, Rfam²⁰ and miRBase²¹, were added. For miRBase, interactions between miRNA and transcripts were predicted using TargetScan²². For both of these databases, new classes were defined in the ontology to describe interaction events, such as binding between a transcript and a miRNA. At the BioHackathon 2014, the group tried to incorporate variant information into the RDF data.

Identifiers for sequences and annotations

There was discussion of how to represent gene names and chromosome Uniform Resource Identifiers (URIs). For gene names, it is recommended to use rdfs:label and dc:identifier for primary gene IDs and use skos:altLabel for gene synonyms. However, it is not mandatory because gene IDs are not always available, depending on the source of information. As for chromosome URIs, it would be useful if the bioinformatics community could agree to a common URI for each chromosome and version (e.g. human chromosome 19 in the GRCh38 assembly). However, we could not reach an agreement at the BioHackathon as it seemed to be impractical to cover every sequence assembly of all species, individuals, cells and samples in an unified manner as drafted in the RDF summit. In this section, we describe the current situation and proposals relating to this issue.

Universal Biological Sequence ID (UBSID). An essential step in the merging of datasets is relating primary identifiers i.e. any data can be joined if they contain the same identifiers. Therefore, all databases can be joined as fully connected Linked Data if appropriate universal identifiers are consistently used. To date, molecular biology has mainly developed around the Central Dogma concept in which higher levels of annotation (transcripts, proteins) are related to the underlying genomic sequence. Genes, as well as protein binding motifs and other features such as SNPs, can be related to DNA sequences, as can the transcriptome and proteome. Therefore, much of modern molecular biology data can in principle be related if the underlying nucleotide sequences are used as the basis for identifiers. However, the use of sequences per se as identifiers has several problems: for example, a sequence can be extremely long (e.g. human chromosome I), or very short (e.g. the location of a SNP), there can be multiple sequences that are highly similar or identical as in multi-copy paralogs, and a sequence feature can be on the sense or antisense strand. In order to overcome these problems, a universal sequence-based identifier scheme should incorporate position information, reference sequence information, the actual sequence (when there are differences, such as mutations, from the reference), strand information, and in addition, it would be ideal if all of such information is expressed as a short, human-comprehensible identifier. By using reference-based compression of DNA sequences based on offset and run-length encoding, the sequence can be expressed just by the mismatching positions and this can form the basis for an identifier system. Therefore, the G-language group proposed a Universal Biological Sequence ID (UBSID) to enable this encoding. For example, the human APOE mRNA sequence is encoded as <http://rest.g-language.org/ubsid/ubsid2seq/hg19-chr19:045409882+A42:=43-1092=193-580=718:> as a URI in the G-language REST service.

Identifiers used in the DDBJ and TogoGenome RDF. After the BioHackathon 2012, a group from DBCLS and DDBJ developed an ontology which can capture semantically the data model of the INSDC, such as the records from GenBank, DDBJ and ENA, with restrictions on terms used in the feature table and qualifier key-values. A converter for INSDC and RefSeq²³ entries to RDF was developed based on the ontology, and the RDFized data is used in the TogoGenome application. TogoGenome integrates information on genes, proteins, organisms, phenotypes and environments. Because the genes in TogoGenome are currently extracted from INSDC and RefSeq records, the URI for each annotation is constructed as a fragment using Identifiers.org URIs in the form: <http://identifiers.org/[insd c or refseq]/[entry_id]#[fragment]>. For example, the human APOE gene on chromosome 19 in the RefSeq record NC_000019.10 is internally represented in TogoGenome as <http://identifiers.org/refseq/NC_000019.10#feature:44905782-44909393:1:gene.1424> and the information can be accessed at <http://togogenome.org/gene/9606:APOE> where 9606 is the taxon ID corresponding to human in the NCBI Taxonomy database and APOE is the gene name used in the record. This approach is slightly different from the proposed UBSID model which encodes sequence alignment with comments but can distinguish the source of information and feature types annotated in the INSDC/RefSeq record. The location of each gene and exons in TogoGenome RDF are described by the FALDO ontology.

Identifiers used in the Ensembl RDF. Ensembl generates their own IDs for genes, transcripts and exons in their database. For example, the human APOE gene is given an ID of ENSG00000130203, which encodes five transcripts (one of them is ENST00000252486) and one of the exons of this transcript is ENSE00003577086. It is natural to use these IDs when constructing URIs for the RDF dataset. In the 2014 development version of the Ensembl RDF, the human APOE gene is indicated as <http://rdf.ebi.ac.uk/resource/ensembl/ENSG00000130203> within a graph identified as <http://rdf.ebi.ac.uk/dataset/ensembl/77/9606> for the human genome dataset in the Ensembl release 77. The location of this gene on human chromosome 19 is designated by <http://rdf.ebi.ac.uk/resource/ensembl/77/chromosome:GRCh38:19:44905754-44909393:1>. The strategy to generate unique URIs for each annotation in Ensembl is different from that employed by DDBJ/INSDC and TogoGenome, which all share both the same RDF model and use the FALDO ontology to describe the actual coordinates of annotations (e.g. genes and exons) on a chromosome. Thus at present further work is needed before all these providers are completely consistent and interchangeable.

Data integration beyond organisms

To facilitate more accurate and deeper integration of data, it is important to standardize metadata accompanying DNA sequences, orthologous gene relationships among organisms, phenotypic properties of organisms, inter-species and organism-environment interactions including host-pathogen relationships. We describe some of these efforts now.

Metadata on samples. DDBJ, EBI and NCBI are jointly hosting the BioSample database as an international collaboration. In this resource, metadata are accumulated on the samples from which DNA sequence in the INSDC database was collected and/or on which other research projects were conducted. The metadata includes species, type of samples (cell types etc.) and phenotypic or environmental information, and therefore it is valuable for data integration if the metadata is available as RDF. A group from DDBJ generated an RDF version of BioSample metadata during the BioHackathon 2014, using as a starting point 14,362 entries stored in the DDBJ BioSample database in XML format.

In addition, existing terminologies and ontologies for geological, archeological and morphological data were explored during the 2014 BioHackathon. For example, there are several resources for geolocations such as W3C Geospatial Ontologies, GeoRSS, GeoNames and Global Biodiversity Information Facility (GBIF). The National Aeronautics and Space Administration (NASA) has developed the Global Change Master Directory (GCMD) and the Semantic Web for Earth and Environmental Terminology (SWEET) which can be used to describe archaeological time scales. For morphology, the Foundational Model of Anatomy (FMA)²⁴, Anatomy Reference Ontology (AEO)²⁵, Vertebrate Skeletal Anatomy Ontology (VSAO)²⁶ and other domain specific ontologies²⁷ were surveyed. These ontologies are essential for encoding RDF data in environmental biology, such as biodiversity and biomolecular archeology. As a case study, a group developed a semantic resource with information about corals by integrating taxonomic, genomic, environmental, disease and coral bleaching information.

Ontologies for integration of microbial data. Within the field of microbiology, genomic and metagenomic data are expanding rapidly due to advances in next generation sequencing technologies. To effectively analyze these huge amounts of data, it is necessary to integrate various microbial data resources available on the Internet. Orthology can play an important role in summarizing such data by grouping corresponding genes across different organisms, and by annotating genes by transferring knowledge from highly curated model organism to newly sequenced genomes. Therefore RDF models were developed for representing the orthology data stored in the Microbial Genome Database for Comparative Analysis (MBGD)⁸, and these were used to construct an RDF version of MBGD. This also required the development of the OrthO ontology²⁸ for representing orthology and aligning concepts with the existing OGO ontology²⁹, with additional definitions mapped from OrthoXML³⁰. Orthology RDF data can now be linked with other databases published as RDF such as UniProt³¹, allowing the integrated dataset to be queried using SPARQL. When searching these data, ontologies can be utilized to specify complex search conditions. To assist making such precise queries, the Microbial Phenotype Ontology (MPO) was developed for describing microbial phenotypes such as microbial morphology, growth conditions, biochemical or physiological properties. During the hackathon, the ontology was updated to comply with a better classification of the hierarchical (is-a) and partonomical (part-of) structure. In addition, the Pathogenic Disease Ontology (PDO) was developed to describe pathogenic microbes that cause diseases in their hosts. An RDF dataset that describes pathogenic information relating to each microbial genome sequence was created using the PDO. Since the genes within these genomes are connected to the ortholog information in the MBGD ortholog database, it is possible to calculate the set of orthologous gene groups that is enriched in the disease related microbes.

Knowledge extraction of factors related to diseases. Information and knowledge of the relationships between genes/ mutations/ lifestyle/ environment and diseases is required in order to predict the risk of a disease and for prognosis after the onset of a disease. In practice, it will also be necessary to collect individual lifestyle and environmental profiles as well as personal genetic data such as genome sequences to allow such predictions for individual people. The necessary underlying relationships are often described in the literature, but are not yet systematically collected in a database. To extract these relationships from the literature, there are two key steps that need to be addressed. First, entities must be annotated automatically using text mining software and, second, these annotations must be represented in a curation interface to allow confirmation that the information has been extracted accurately. Genes, genetic variants, diseases, environmental factors and lifestyle factors are the entity types that need to be annotated on the corpus. Existing software for extracting genes (e.g. GNAT³², GenNorm³³ etc.), mutations (e.g. tmVar³⁴, MutationFinder³⁵ etc.) and diseases (e.g. BANNER³⁶ with disease model) are openly available, along with existing datasets such as BioContext³⁷ and EVEX DB³⁸. Before environmental factors and lifestyle factors can be extracted systematically it is necessary to decide on a controlled vocabulary (whether existing or not) to represent them. Pregnancy Induced Hypertension (PIH) was chosen as a case study and 86 relevant open access PubMed Central articles identified. It was possible to extract genes from 32 of these articles using the BioContext dataset, while the other 54 articles were published more recently than BioContext. Attempts were made to extract mutations from 86 articles. For lifestyle and environmental factors, controlled vocabularies were collected in preparation for entity recognition. After obtaining all the entities in the 86 articles, they were curated using interfaces such as PubAnnotation³⁹, and the curated relationships represented as an RDF graph.

Tools for semantic genome data

Genome annotations have historically been represented and distributed in non-standard domain-specific data formats (e.g., INSDC, GFF3, GTF). The data formats themselves often include implicit semantics, making automatic interpretation and integration of the data with other resources challenging. Therefore, tools to convert those data into RDF and ontologies to support semantic representation of data need to be developed. BioInterchange is a tool to convert those file formats into RDF and was originally developed in the BioHackathon 2012⁴, with its functionalities and ontologies being enhanced over successive hackathons. Other tools for high-throughput data processing of Sequence Alignment/Map format (SAM), Binary SAM format (BAM)⁴⁰, Variant Call Format (VCF)⁴¹, Genome Variation Format (GVF)⁴², Header-Dictionary-Triples (HDT)⁴³ files have also been developed and a middleware to enable SPARQL queries directly against these huge files on-the-fly for scalability was explored and results were incorporated into integrated semantic genome databases such as TogoGenome and MicrobeDB.jp.

Utilization of domain specific data formats in semantic web. In BioHackathon 2013, VCF2RDF was developed and subsequently published as a Ruby program to convert VCF files into RDF, which represent positions in FALDO and alleles in its own ad hoc ontology terms. The resulting RDF data was loaded into Fuseki and queries were tested in the Jena framework, taking three minutes for the cow genome on a laptop to plot quality scores of variant calls for a million base pairs. During BioHackathon 2014, a group developed middleware to interpret SPARQL queries against SAM/BAM/VCF files on the fly. The first implementation was prototyped in JRuby so that the Java library for samtools can be used in a Ruby program. The resulting application, VCFotf, is packaged as a Docker image that serves a query interface on the Web page. Also, another implementation (sparql-vcf) was developed with Jena for improving query execution time, in which Jena property functions are used to introduce a "special predicate" which accelerates search performance; however, this ‘boutique’ query violates the SPARQL standard.

Use of compressed RDF for large scale genomic data. BioInterchange was used in a Genomic HDT project as a feasibility study to convert a variety of genomic data files (e.g. GVF) containing coordinate-annotated genomic features into an ontology-annotated RDF representation. The RDF data file is then processed into an RDF/HDT file, which is a compressed, indexed, and queryable data archive. Using Ensembl's human somatic variation data (81MB, 9MB gzipped), it was found that the RDF/HDT archive is only 20MB (1.5M triples; 15MB data + 5MB index), which is a significant reduction from the 313MB RDF N-triples representation. A JSON RESTful API was made available using Sinatra to provide access to the RDF/HDT file, and this allowed a demonstration of genome-based browsing of the RDF/HDT data file using the JBrowse genome browser.

Integrated semantic genome databases. TogoGenome⁹ was developed to integrate heterogeneous biomedical data using Semantic Web technologies. This utilizes the representation of genomic data in the standard RDF format, enabling interoperation with any other Linked Open Data (LOD) around the world. To support these efforts we have collaborated with DDBJ, UniProt, and the EBI RDF group to develop ontologies for representing locations and annotations of genome sequences and used these developments for all prokaryotic genomes and, later, eukaryotic genomes. To complement the above work we developed ontologies for taxonomies, phenotypes, environments and diseases related to organisms, so enabling faceted browsing of the entire datasets. Every TogoGenome report page is made up of modular components called TogoStanza, which is a generic framework to generate Web components querying SPARQL endpoints and rendering them as HTML elements. Stanzas are re-usable modules which can be shared and embedded easily into other databases, and which have been developed in collaborations with MicrobeDB.jp, MBGD⁸ and CyanoBase⁴⁴, resulting in over 100 TogoStanzas being available so far.

Visualization of semantic annotations in JBrowse. JBrowse⁴⁵ was used by several projects within the BioHackathon as a demonstration platform. JBrowse running on top of the SPARQL endpoints, e.g. TogoGenome or a prototype InterMine⁴⁶ endpoint, or from indexed files produced by GenomicHDT, were comparable in performance with typical RDB back-ended settings. In addition, an unusual use of JBrowse was to view text instead of DNA sequence, with the annotation viewed being the output of natural language processing.

TogoGenome: JBrowse was extended to support the TogoGenome's SPARQL endpoint as a data source to retrieve and visualize genes on a chromosomal track (Figure 2). This enhancement is already merged into the official JBrowse release since version 1.10 in 2013. SPARQL queries are customizable in the JBrowse configuration file as long as they return start, end, strand, type (label), uniqueID and parentUniqueID of the annotation objects in a given range within a sequence. When scrolling to neighboring regions, the performance is good enough for browsing.

Figure 2. SPARQL back-ended JBrowse is integrated into the TogoGenome database.

InterMine: Representatives from the InterMine project⁴⁶ produced proof-of-concept demonstrations of semantic extensions to InterMine data-warehouses. These included on the one hand a draft of how to model InterMine data as linked data, producing both an ontology of relationships and triples that conform to that ontology, and on the other hand a draft of a very limited SPARQL engine capable of operating on an InterMine data source directly. Together these investigations indicate that given some development effort, it is likely that significant progress can be made to integrating InterMine into the semantic web. An area that needs work, and is receiving attention, is the production of stable URIs for InterMine entities. In addition to this, work was done to implement a simple adaptor allowing, as described above, JBrowse to request data directly from InterMine RESTful web services.

Text-mining: In the community of BioHackathon, text mining resources were developed around PubAnnotation, a public repository of literature annotation data sets. Usually text mining requires its own set of tools, e.g. viewers or editors. However, an interesting experiment was carried out during BioHackathon 2013 and 2014 to use JBrowse as a viewer of text annotation data. The idea behind the experiment was that both genomic data and text data are represented as character sequences, and that annotations of both type of data are attached to specific regions on the sequences. A simple script was developed to convert annotations in PubAnnotation to JBrowse format, and it was observed that text annotations can be nicely viewed in JBrowse. The result raises the possibility of further interoperability between tools for genomics and text mining.

Protein structures, interactions and expressions

The European Bioinformatics Institute’s (EBI) SIFTS "Structure Integration with Function, Taxonomy and Sequences" resource provides regularly updated residue-level mappings between UniProt and PDB entries⁴⁷. SIFTS has been distributed in Comma Separated Values (CSV) and Extensible Markup Language (XML) formats. Like many other proteome-related databases, SIFTS uses the classical protein chain ID specified by the author. However, in 2016, the Worldwide Protein Data Bank (wwPDB) will abolish the conventional PDB format and instead will distribute RDF/XML based on the PDB exchange dictionary / macromolecular Crystallographic Information Format (PDBx/mmCIF) [PDBx/mmCIF]. At the same time wwPDB will start assigning protein chain identifiers, which will also be encoded as URIs in the wwPDB/RDF.

During the BioHackathon, an RDF version of SIFTS (RDF-SIFTS) was designed and implemented to provide residue-to-residue correspondence between PDB and UniProt entries in RDF⁴⁸. RDF-SIFTS links both the protein chain ID assigned by authors and the one assigned by wwPDB to SIFTS. RDF-SIFTS uses existing ontologies of PDB, UniProt, EMBRACE Data and Methods (EDAM)⁴⁹ as well as FALDO, and resources are linked to Identifiers.org⁵⁰ URIs.

The University of Tokyo Proteins (UTProt)⁵¹ is a project that is collecting and building RDF to support interactome linked data. During the BioHackathon, the UTProt group extended RDF-SIFTS to cover intermolecular interactions, and this resulted in six billion triples including, for each pair of residues in the interacting surfaces, their separation distance. This resource will be useful for analysis of structure and sequence in proteomics and interactomics. Serialization Ruby code, RDF-SIFTS maker, is available through GitHub as open source software which can be used to convert new release of SIFTS data from EBI.

"Omics" technologies are primarily aimed at the universal detection of genes (genomics), mRNAs (transcriptomics), proteins (proteomics) and metabolites (metabolomics) in a specific biological sample. Proteomics and metabolomics in particular have gained a lot of attention in recent years due the possibility of studying reactions, post-translational modifications, and pathways⁵². The proteomics community has been working for more than ten years in the standardization of file formats and proteomics data⁵³. Different XML-based file formats and open-source libraries have been released to handle proteomics data from spectra to quantitation results^54–56.

In contrast, metabolomics is a relatively new "omics" field where the standardization of exchange formats is difficult, due to the variety of measurement methodologies ranging from nuclear magnetic resonance (NMR) spectroscopy to a variety of mass spectrometers (MS). Moreover, currently no single system can provide enough resolution to measure the entire set of small molecules within a biological sample; instead, data from multiple systems are combined to gain more comprehensive coverage, for instance combining Liquid Chromatography (LC), Gas Chromatography (GC), and Capillary Electrophoresis (CE) separation prior to analysis in a mass spectrometer. Recently, the mzTab data exchange format was introduced by the Human Proteome Organization (HUPO) Proteomics Standards Initiative, as a standardized format to report both qualitative and quantitative metabolomics and proteomics experiments in a simple tabular format⁵⁷. In BioHackathon 2014, a Perl library was developed to standardize the metabolomics data obtained from MasterHands software⁵⁸. MasterHands is a proprietary software for the analysis of CE-MS-based metabolomics used in the Institute for Advanced Biosciences, Keio University, and at Human Metabolome Technologies Inc. The library allows the annotation of KEGG compound information using the KEGG REST API, and also allows the annotation of Reactome and MetaCyc information.

In the age of systems biology and data integration, proteomics data represent a crucial component to understand the “whole picture” of life. In this context, well-established databases for proteomics data include the Global Proteome Machine Database (GPMDB), PeptideAtlas, ProteomicsDB, and the Proteomics Identification (PRIDE) database among others⁵⁹. In addition, at BioHackathon 2014, the "omics" group worked on the standardization to RDF of different web services and APIs for proteomics and protein expression data. The GPMDB2RDF and PRIDE2RDF library allow the export of expression data from the GPMDB Database⁶⁰ and PRIDE Database⁶¹ respectively. The development of a standard interface for providing protein expression data will allow, in the future, exchange and proper reuse of public proteomics data. To this end, the "omics" group in the BioHackathon 2014 made the first steps towards the development of the ProteomeXchange Interface (PROXI) for protein expression data exchange⁵⁹.

Glycoinformatics

The glycoscience group participated in a satellite BioHackathon in Dalian, China, in parallel to the GLYCO 22 Meeting held June 23–28, 2013. Although a preliminary RDF format was developed at the previous BioHackathon in 2012⁶², there was a need to address not only glycan structures (sequences) but also supporting experimental data, the biological source of the sample analyzed, and publication information. Therefore, during BioHackathon 2013, a formal ontology to represent these features, as well as the glycan structures to which they relate, was discussed. The aim of the GlycoRDF group was to define a standard RDF representation, in the form of an ontology by integrating features from existing ontologies where possible and creating new classes and relationships where needed.

As it would be impossible, in a week, to create an ontology that could cover the full spectrum of glycomics information and experimental data, it was decided that the group would limit the first version to the data that currently exists in glycomics databases. On the other hand, the developers also attempted to define the ontology so that it could be easily extended with additional predicates and classes if needed, in case more data or more glyco-related databases utilize the proposed RDF format. As a result, by the end of BioHackathon 2013, the first version of the GlycoRDF ontology was agreed upon and is currently available at the GlycoRDF repository at 63. In 2014, work progressed to the point where all glyco-scientists who attended previous BioHackathons had now generated GlycoRDF-formatted versions of their databases. The updated list of these databases are listed and documented on the GlycoRDF repository.

Enzymatic reaction ontology

Entities can be classified based on a variety of features, such as their function(s), structures/sub-structures, or chemical properties. For example, genes and proteins are independently classified based on their functions, role, and cellular location, organized by the Gene Ontology (GO)⁶⁴. At the same time, gene and proteins are also classified based on their conserved partial substructures, such as protein domains in Pfam. ChEBI⁶⁵ classifies chemical substances by their overall functions (ChEBI role ontology) and by their partial structures (ChEBI molecular structure ontology). For enzymes, their overall functions are classified by the Enzyme List of International Union of Biochemistry and Molecular Biology, often referred to as the Enzyme Commission (EC) numbers⁶⁶. To date, however, there have been no standard ways to classify enzymes based on the partial structures of their enzymatic reactions. Therefore, during BioHackathon 2013 we discussed the development of an ontology that deals with the partial structures of enzymatic reactions, i.e. substrate-product pairs derived from reaction equations. This led to the Enzyme Reaction Ontology for annotating Partial Information of biochemical transformation (PIERO) being published in 2014⁶⁷. In BioHackathon 2014, we had further discussions to refine the PIERO data to establish the PIERO Ver0.3 Schema. This ontology was later used in de novo metabolic pathway reconstruction analysis⁶⁸ and for ortholog predictions⁶⁹.

Modeling text annotations on the Semantic Web

Text mining is becoming an increasingly common component of biological curation pipelines and biological data analysis, and as such there is increasing demand for both text that has been automatically annotated with natural language processing tools, and annotated document resources that can be used in development and evaluation of those tools. This demand in turn leads to a need for standard, interoperable representations for annotations over documents. Several proposals for general linguistic annotation representations have been made⁷⁰, including ones specifically for biomedical text annotation representations^71,72, as well as data models underpinning standard modular architectures such as Unstructured Information Management Architecture (UIMA)⁷³. However, these approaches have not been adapted to the Semantic Web. Recently, the Open Annotation Core Data Model has been proposed to enable interoperable annotations on the web⁷⁴. This project explored the application of the Open Annotation Model to the use case of capturing text mining output, by harmonizing the data models of the existing proposals.

The existing RDF-based representation of the PubAnnotation tool³⁹ was used as a starting point, and adapted for compatibility with the Open Annotation model. The Open Annotation model provides an annotation class that relates a web resource to information that is about that resource; this representational choice is different from other models yet critically allows separation of metadata (e.g. provenance information) about the annotation itself, from meta-data about the content of the annotation⁷⁵. Several core requirements for text-based annotations were identified: (1) representation of document spans as annotation targets; (2) representation of "simple" associations, e.g. between a span of text and a concept such as an ontology identifier; (3) representation of "complex" associations, e.g. between several spans of text and a relation or event. In addition, the overall structure of a document corpus, which can consist of several documents, must be modeled in such a way as to allow those documents to have internal structure such as chapters, sections, passages or sentences. PubAnnotation models text spans relative to these internal structural elements, while BioC and UIMA have adopted absolute character offsets across a complete document. The model developed here allows for both, by allowing the target of annotation to be either a full document, or a document element as appropriate. It is hoped that the proposals made for web-based document annotation representations will enable interoperability with other Open Annotation-based data and tools, while also addressing the need to move linguistic annotation into the web.

During BioHackathon 2014, the integration of literature annotation resources was pursued with actual data sets. Colorado Richly Annotated Full-Text (CRAFT)⁷⁶ is a recent important achievement of biomedical text mining, which included 67 full papers with rich annotation based on 7 biomedical ontologies.

The GRO corpus⁷⁷ is a richly annotated corpus based on the Gene Regulation Ontology⁷⁸. Allie is an acronym-annotated collection of all PubMed titles and abstracts⁷⁹. They were all converted into PubAnnotation-compatible format, and submitted to PubAnnotation. The whole-PubMed-scale dataset, Allie, triggered the issue of scalability. However, integration of the two corpora, CRAFT and GRO, into PubAnnotation, demonstrated significantly improved utility.

Natural language query

SPARQL is a standard language for querying triple stores. However, SPARQL queries can be difficult to write, even for experts. Usability studies have shown natural language interfaces to SPARQL to be the preferred method of SPARQL query formation assistance⁸⁰. For this reason, software developers are encouraged to create applications that allow users to ask biomedical questions against triple stores using natural (i.e. human) language.

Building on the work in BioHackathon 2012 on querying Systematized Nomenclature of Medicine-Clinical Terms (SNOMED-CT), during BioHackathon 2013 effort was focused on the Online Mendelian Inheritance in Man (OMIM) SPARQL endpoint, with a simultaneous focus on building an evaluation data set. Social networking was used to obtain use cases from biologists and informaticians, and it was quickly discovered that the system had an issue with differentiating between broad semantic types and specific instances. For example, “heart disease” was correctly mapped to a specific entity, but the word “genes” was incorrectly mapped to one specific gene. For this reason, dealing with the issue of recognizing broad semantic classes was the major focus of the development work, and testing semantic class recognition was the main focus of the testing effort. OMIM uses Type Unique Identifiers (TUIs), in the Unified Medical Language System (UMLS)⁸¹, to semantically type subjects and objects in its triple store, so we approached the problem of recognizing broad semantic classes as recognizing mentions of TUIs. Accordingly, a TUI concept recognizer was implemented into the open source LODQA system for automatic generation of SPARQL queries from natural language queries.

Efforts to develop a natural language interface were continued in BioHackathon 2014, during which the LODQA system was configured for two large scale RDF datasets, Bio2RDF and BioGateway. In this way, it was demonstrated that technology like LODQA can answer a question like, “Which genes are involved in calcium binding?”, based on RDF data sets like Bio2RDF. However, it also revealed remaining performance issues.

VoID for InterMine and UniProt

As VoID is a vocabulary for describing datasets that can be used to generate documentation and assist users in finding key knowledge on how to write analytical data queries, the InterMine group worked on automatically generating VoID files for InterMine-based Model Organism Databases, while the UniProt group worked on the same for showing classes and predicates used in the named graphs on the UniProt SPARQL endpoint.

InterMine⁸⁶ is an open source graph-based data warehouse system built on top of PostgreSQL. Through a collaboration (the InterMOD consortium⁸⁷), with most of the main animal Model Organism Databases (MODs) there are now InterMine databases available for budding yeast (SGD)⁸⁸, rat (RGD⁸⁹), zebrafish (ZFIN⁹⁰), mouse (MGI⁹¹), nematode (WormBase, unpublished), Fly (InterMine group)⁹², and Arabidopsis⁹³ with further MOD InterMine instances expected. Extensive data from the modENCODE project⁹⁴ are also available through modMine⁹⁵. As a step towards exposing these rich data as RDF, code was developed that uses existing InterMine RESTful web services to interrogate the FlyMine database and to generate a VoID description of the database. Further work is required to adjust the core InterMine data model to include additional database metadata items. This will then allow the automatic generation of VoID descriptions for any InterMine database. Further work is also required to ensure that appropriate standards are adhered to, especially for RDF predicates. In addition to the above developments, progress was made in creating a Sesame-based SPARQL endpoint for InterMine databases to complement the existing web application and web services. At the moment the endpoint only supports a small range of simple queries. It is hoped that in the future such endpoints will make available the rich data assembled and curated by the world wide Model Organism Database community. In the process this should provide opportunities for interoperation and also a mechanism for federation across the different resources.

UniProt³¹ is available as RDF and can be queried via SPARQL and REST services. UniProt is a large and complicated database, that is difficult to explore due to its size. During the hackathon we implemented a procedure to generate a VoID file to describe UniProt data. The VoID file, now available on FTP and via the uniprot.org SPARQL Service Description (application/rdf+xml), is updated every release in synchrony with our production, and show users what types of data (and how much) are available in the UniProt datasets. We also document how many links to other databases UniProt provides, demonstrating the hub effect of UniProt.org in the life science domain. For the UniProt SPARQL endpoint this VoID description is used as a key part of the user documentation describing the schema of the UniProt data.

Schema.org and RDFa for biological databases

Schema.org is a collection of extensible schemas that webmasters can use to mark up structured data on their web pages with the aim of improving search engine performance and enabling the creation of other applications. The initiative was founded by Google, Bing and Yahoo! as a collaboration to improve the web and their search results by using such structured data. More than 700 item types have been listed in schema.org, some of which have been supported by these search engines. If webmasters mark up their content in an acceptable markup format (e.g. Microdata, microformats or RDFa), then web crawler programs can detect these structured data and they can be rendered as rich snippets in the search results.

During the BioHackathon, the members of this working group proposed two item types for a schema extension: "BiologicalDatabase" and "BiologicalDatabaseEntry". We discussed what item properties would be suitable for our purposes and how to label them in markup. Finally, we decided to use the Microdata format to mark up web pages and proposed five original properties: "entryID", "isEntryOf", “taxon”, "seeAlso" and "reference". Work in this area is now being carried forward by the bioschemas.org project.

We also publicized our proposal and encouraged BioHackathon members to mark up their databases. A Microdata crawler was created to extract these structured data. We modified "Sagace"⁹⁶, a web-based search engine for biomedical data and resources in Japan, developed at the NIBIOHN in collaboration with NBDC. We confirmed that marked-up data showed up as rich snippets in search results. Ten databases have been marked up with our new proposal and so can help improve the readability of search results. This service is freely available at http://sagace.nibiohn.go.jp.

Nanopublication

A nanopublication is defined as the smallest unit of publishable information that represents a finely-grained, but complete idea. Nanopublications are composed of such fine-grained assertions coupled with provenance metadata about the assertion, such as the methods used to create it and personal and institutional attributions, and finally additional metadata about the nanopublication itself, such as who or what created it, and when. The aim is to make a formal, predictable, and transparent relationship between data and its provenance. Nanopublications will be discussed here with respect to their application to FANTOM5^97,98 data, to track DBCLS literature curation, and within the Semantic Automated Discovery and Integration (SADI) framework⁹⁹.

The FANTOM5 project monitored transcription initiation at single base-pair resolution in mammalian genomes by Cap Analysis Gene Expression (CAGE) coupled with single molecule sequencing^97,98. Promoters were defined as upstream of CAGE peaks (transcription start site clusters) and their activities were quantified based on their read counts. The FANTOM5 promoters and their activities were described in nanopublications¹⁰⁰ to facilitate their open and interoperable exchange. Three classes of nanopublications, having the following assertions¹⁰¹, were generated: 1) A CAGE peak is defined in a specific region of the genome, 2) The CAGE peak is a transcription start site (TSS) region, which is part of a gene, 3) The CAGE peak is active at a certain level in a specified sample. Class 1 nanopublications (CAGE peaks) provide minimum information based on a model on genomic coordinates. They can be exported to genome browsers. Class 2 nanopublications (gene associations) are served as supplemental data to allow biological searches. This class of nanopublications may be re-released when a new data processing workflow is available or when different parameters or gene definitions are used. Class 3 nanopublications (activity levels of transcription in individual samples) are used only if the details of expression are relevant in a given biological search. By dissecting the whole data set into three classes of nanopublications with different granularities, its reusability is increased. These nanopublications are available at http://rdf.biosemantics.org, and they have been reported also in an article related to FANTOM5¹⁰¹.

The DBCLS has developed a web-based gene annotation tool, TogoAnnotation, has provides an easy way of accessing and adding annotations. Likewise Gene Indexing was developed as a simple named-entity recognition (NER) task in order to make connections between genomic loci and the literature. Gene Indexing generates micro-annotations by manually extracting gene and protein symbols from the text, tables and figures of full papers and connecting them to both PubMed IDs and genome location. A total of 10 curators cooperated over a five year period to manually annotate over 5,000 full papers relating to microbes. In this way over 200,000 gene/protein micro-annotations were generated.

Based on the above data, during the BioHackathon 2014, a Nanopublication model was developed for these literature curation data, as well as a converter to make any annotation in the TogoAnnotation system representable as a Nanopublication RDF (Figure 4). It is intended that the curation data be integrated into the TogoGenome system and be expanded as a standard distributed annotation platform in the future.

Figure 4. Proposed nanopublication data model for TogoAnnotation data.

The SADI Semantic Web services project also has a need to represent rich provenance data regarding how its services create their output. Given the rapid growth and notable success of the OpenPHACTS¹⁰² and NanoPublications¹⁰³ projects, it seems desirable that analytical Services - those following the SADI Semantic Web Service design patterns in particular - should output semantic data that follows the same NanoPublication paradigm. This would allow SADI services to publish new biomedical knowledge directly into the vast integrated NanoPublications space, and take advantage of their integration tools.

Extensions to the existing Perl SADI::Simple codebase in Comprehensive Perl Archive Network (CPAN) were undertaken at the hackathon. A key consideration was to ensure that the code could support distinct metadata for each triple, since SADI services are specifically designed to support multiplexed inputs potentially spread over a large number of processors for analysis, before being reassembled into an output message. As such, it is potentially the case that each triple has slightly distinct provenance information. The implemented solution guarantees globally unique identification of each of these nanopublications, for each execution, even over multiple iterations of the same input data.

NanoPublications are created when, through HTTP content negotiation, the client requests n-quads. The service responds with an RDF structure that follows the structure of the (proposed) NanoPublication Collection.

Requesting quads from a ‘legacy’ SADI Service that does not support NanoPublications will result in a HTTP 406 (Not Acceptable) response, with an output body in application/rdf-xml, as is allowed by HTTP 1.1.

Bio2RDF2SADI

Discovering and reusing data requires substantial expertise about where data are located and how to transform them into a more useable form for further analysis. While the Bio2RDF project transforms dozens of key bioinformatic resources into RDF, and is made available through public SPARQL endpoints, a key challenge still remained: how to identify which datasets contain the entities and relations that are of interest to solve a particular problem. To this end, Bio2RDF now generates and publishes summaries of the dataset contents in each of its SPARQL endpoints, thereby simplifying lookup, and reducing server load for expensive and common queries.

During the hackathon, an architecture was developed for an automated approach that utilizes the metadata from Bio2RDF’s content summaries to automatically generate SADI Semantic Web Services that provide discoverable access to this Bio2RDF data¹⁰⁴. SADI Services use ontologies to formally describe their inputs and outputs, such that it is possible to find services of interest by querying their ontological descriptions via a global Service metadata registry. In the case of these Bio2RDF SADI Services, the input data-type is a simple Bio2RDF typed-URI (for example [http://bio2rdf.org/mesh:C025643 rdf:type ctd:Chemical]) and the output is, as per the SADI specifications, the input node annotated with a Bio2RDF relation (for example [http://bio2rdf.org/mesh:C025643 sio:is-participant-in http://bio2rdf.org/go:0008380). Such metadata descriptions can be automatically generated from the Bio2RDF indexes, and moreover, the corresponding SPARQL queries that make up the business logic of the service can similarly be automatically constructed based on the information in these indexes. As such, both the service description, as well as the service itself, can be dynamically created to provide access to any Bio2RDF data of interest.

The advantage of exposing Bio2RDF as a set of SADI services is that the data in Bio2RDF becomes discoverable - software does not need to know, a priori, what data/relations exist in which Bio2RDF endpoint. Moreover, when exposed as SADI Services, Bio2RDF data can more easily be integrated into workflows using popular workflow editors such as Taverna¹⁰⁵ or as demonstrated by our use of these services within Galaxy workflows¹⁰⁶.

Quality assessment

A large amount of biomedical information is available via SPARQL endpoints, often in a redundant way. Life Sciences databases often integrate information from different sources to enrich the data they provide, and some information resources are pure aggregators whose value is in the harmonization of the information that they collect. As these resources publish their information on the Semantic Web, the result is that the same information is present in multiple endpoints. As a consequence, to decide which endpoint to use to access some particular data of interest is not a trivial task. Two hackathon activities addressed this issue. The development of a dataset descriptor is useful to know which data are present in an endpoint, with information on version, representation and update policies. But even if such a descriptor is provided, there is still an issue of the reliability of endpoints. It is also difficult to know which endpoints are actively maintained and which are not.

YummyData is a project that monitors endpoints by periodically running queries and performing a few tests. By collecting data over extended periods, it can provide a proxy for the reliability of an endpoint and the dynamism of the information it provides. More specifically, YummyData periodically queries datahub.io for datasets tagged as being of biomedical interest. It combines the result with a list of curated endpoints and, periodically, it runs a series of tests and queries and stores their results. YummyData performs some tests to determine whether the endpoint provides a VoID descriptor (see section above), as well as to measure response time. It also runs a series of queries that can be generic or endpoint-specific. Generic queries inspect aggregate information such as the number of statements, distinct resources, or properties. Specific queries are currently only implemented as a proof of concept, but they are intended to reveal aspects of the quality of the data provided by endpoints. For instance, a typical query would ask for the number of entities annotated via a given evidence code. Results over time are then aggregated in two types of rating: a SPARQL score that is a numeric value that results from a count of positive response codes over time windows; a star rating that is intended to provide a more qualitative assessment of features (e.g. the availability of a valid VoID descriptor, or of a copyright notice, yields +1 star). At the time of writing, YummyData has collected data for about a year on a few tens of information resources. A subset of these data are accessible via the http://yummydata.org website.

Underlying data

No data are associated with this article.

Extended data

Records of the BioHackathon 2013 and 2014 meetings are aggregated at https://github.com/dbcls/bh13/wiki and https://github.com/dbcls/bh14/wiki respectively.

Zenodo: dbcls/bh13: Included repositories related to BH13. http://doi.org/10.5281/zenodo.3271508¹⁰⁷

This project contains the following extended data:

Zenodo: dbcls/bh14: Included repositories related to BH14. http://doi.org/10.5281/zenodo.3271509¹⁰⁸

This project contains the following extended data:

Data are available under the terms of the Creative Commons Attribution 4.0 International license (CC-BY 4.0).