Science Forum: Wikidata as a knowledge graph for the life sciences

  1. Andra Waagmeester
  2. Gregory Stupp
  3. Sebastian Burgstaller-Muehlbacher
  4. Benjamin M Good
  5. Malachi Griffith
  6. Obi L Griffith
  7. Kristina Hanspers
  8. Henning Hermjakob
  9. Toby S Hudson
  10. Kevin Hybiske
  11. Sarah M Keating
  12. Magnus Manske
  13. Michael Mayers
  14. Daniel Mietchen
  15. Elvira Mitraka
  16. Alexander R Pico
  17. Timothy Putman
  18. Anders Riutta
  19. Nuria Queralt-Rosinach
  20. Lynn M Schriml
  21. Thomas Shafee
  22. Denise Slenter
  23. Ralf Stephan
  24. Katherine Thornton
  25. Ginger Tsueng
  26. Roger Tu
  27. Sabah Ul-Hasan
  28. Egon Willighagen
  29. Chunlei Wu
  30. Andrew I Su  Is a corresponding author
  1. Micelio, Belgium
  2. Department of Integrative Structural and Computational Biology, The Scripps Research Institute, United States
  3. Center for Integrative Bioinformatics Vienna, Max Perutz Laboratories, University of Vienna and Medical University of Vienna, Austria
  4. McDonnell Genome Institute, Washington University School of Medicine, United States
  5. Institute of Data Science and Biotechnology, Gladstone Institutes, United States
  6. European Bioinformatics Institute (EMBL-EBI), United Kingdom
  7. School of Chemistry, The University of Sydney, Australia
  8. Division of Allergy and Infectious Diseases, Department of Medicine, University of Washington, United States
  9. Wellcome Trust Sanger Institute, United Kingdom
  10. School of Data Science, University of Virginia, United States
  11. University of Maryland School of Medicine, United States
  12. Department of Animal Plant and Soil Sciences, La Trobe University, Australia
  13. Department of Bioinformatics-BiGCaT, NUTRIM, Maastricht University, Netherlands
  14. Retired researcher, Germany
  15. Yale University Library, Yale University, United States

Abstract

Wikidata is a community-maintained knowledge base that has been assembled from repositories in the fields of genomics, proteomics, genetic variants, pathways, chemical compounds, and diseases, and that adheres to the FAIR principles of findability, accessibility, interoperability and reusability. Here we describe the breadth and depth of the biomedical knowledge contained within Wikidata, and discuss the open-source tools we have built to add information to Wikidata and to synchronize it with source databases. We also demonstrate several use cases for Wikidata, including the crowdsourced curation of biomedical ontologies, phenotype-based diagnosis of disease, and drug repurposing.

Introduction

Integrating data and knowledge is a formidable challenge in biomedical research. Although new scientific findings are being discovered at a rapid pace, a large proportion of that knowledge is either locked in data silos (where integration is hindered by differing nomenclature, data models, and licensing terms; Wilkinson et al., 2016) or locked away in free-text. The lack of an integrated and structured version of biomedical knowledge hinders efficient querying or mining of that information, thus preventing the full utilization of our accumulated scientific knowledge.

Recently, there has been a growing emphasis within the scientific community to ensure all scientific data are FAIR – Findable, Accessible, Interoperable, and Reusable – and there is a growing consensus around a concrete set of principles to ensure FAIRness (Wilkinson et al., 2019; Wilkinson et al., 2016). Widespread implementation of these principles would greatly advance efforts by the open-data community to build a rich and heterogeneous network of scientific knowledge. That knowledge network could, in turn, be the foundation for many computational tools, applications and analyses.

Most data- and knowledge-integration initiatives fall on either end of a spectrum. At one end, centralized efforts seek to bring multiple knowledge sources into a single database (see, for example, Mungall et al., 2017): this approach has the advantage of data alignment according to a common data model and of enabling high performance queries. However, centralized resources are difficult and expensive to maintain and expand (Chandras et al., 2009; Gabella et al., 2018), at least in part because of bottlenecks that are inherent in a centralized design.

At the other end of the spectrum, distributed approaches to data integration result in a broad landscape of individual resources, focusing on technical infrastructure to query and integrate across them for each query. These approaches lower the barriers to adding new data by enabling anyone to publish data by following community standards. However, performance is often an issue when each query must be sent to many individual databases, and the performance of the system as a whole is highly dependent on the stability and performance of each individual component. In addition, data integration requires harmonizing the differences in the data models and data formats between resources, a process that can often require significant skill and effort. Moreover, harmonizing differences in data licensing can sometimes be impossible.

Here we explore the use of Wikidata (www.wikidata.org; Vrandečić, 2012; Mora-Cantallops et al., 2019) as a platform for knowledge integration in the life sciences. Wikidata is an openly-accessible knowledge base that is editable by anyone. Like its sister project Wikipedia, the scope of Wikidata is nearly boundless, with items on topics as diverse as books, actors, historical events, and galaxies. Unlike Wikipedia, Wikidata focuses on representing knowledge in a structured format instead of primarily free text. As of September 2019, Wikidata's knowledge graph included over 750 million statements on 61 million items (tools.wmflabs.org/wikidata-todo/stats.php). Wikidata was also the first project run by the Wikimedia Foundation (which also runs Wikipedia) to have surpassed one billion edits, achieved by a community of 12,000 active users, including 100 active computational ‘bots’ (Figure 1—figure supplement 1).

As a knowledge integration platform, Wikidata combines several of the key strengths of the centralized and distributed approaches. A large portion of the Wikidata knowledge graph is based on the automated imports of large structured databases via Wikidata bots, thereby breaking down the walls of existing data silos. Since Wikidata is also based on a community-editing model, it harnesses the distributed efforts of a worldwide community of contributors, including both domain experts and bot developers. Anyone is empowered to add new statements, ranging from individual facts to large-scale data imports. Finally, all knowledge in Wikidata is queryable through a SPARQL query interface (query.wikidata.org/), which also enables distributed queries across other Linked Data resources.

In previous work, we seeded Wikidata with content from public and authoritative sources of structured knowledge on genes and proteins (Burgstaller-Muehlbacher et al., 2016) and chemical compounds (Willighagen et al., 2018). Here, we describe progress on expanding and enriching the biomedical knowledge graph within Wikidata, both by our team and by others in the community (Turki et al., 2019). We also describe several representative biomedical use cases on how Wikidata can enable new analyses and improve the efficiency of research. Finally, we discuss how researchers can contribute to this effort to build a continuously-updated and community-maintained knowledge graph that epitomizes the FAIR principles.

The Wikidata Biomedical Knowledge Graph

The original effort behind this work focused on creating and annotating Wikidata items for human and mouse genes and proteins (Burgstaller-Muehlbacher et al., 2016), and was subsequently expanded to include microbial reference genomes from NCBI RefSeq (Putman et al., 2017). Since then, the Wikidata community (including our team) has significantly expanded the depth and breadth of biological information within Wikidata, resulting in a rich, heterogeneous knowledge graph (Figure 1). Some of the key new data types and resources are described below.

Genes and proteins: Wikidata contains items for over 1.1 million genes and 940 thousand proteins from 201 unique taxa. Annotation data on genes and proteins come from several key databases including NCBI Gene (Agarwala et al., 2018), Ensembl (Zerbino et al., 2018), UniProt (UniProt Consortium, 2019), InterPro (Mitchell et al., 2019), and the Protein Data Bank (Burley et al., 2019). These annotations include information on protein families, gene functions, protein domains, genomic location, and orthologs, as well as links to related compounds, diseases, and variants.

Genetic variants: Annotations on genetic variants are primarily drawn from CIViC (http://www.civicdb.org), an open and community-curated database of cancer variants (Griffith et al., 2017). Variants are annotated with their relevance to disease predisposition, diagnosis, prognosis, and drug efficacy. Wikidata currently contains 1502 items corresponding to human genetic variants, focused on those with a clear clinical or therapeutic relevance.

Chemical compounds including drugs: Wikidata has items for over 150 thousand chemical compounds, including over 3500 items which are specifically designated as medications. Compound attributes are drawn from a diverse set of databases, including PubChem (Wang et al., 2009), RxNorm (Nelson et al., 2011), the IUPHAR Guide to Pharmacology (Harding et al., 2018; Pawson et al., 2014; Southan et al., 2016), NDF-RT (National Drug File – Reference Terminology), and LIPID MAPS (Sud et al., 2007). These items typically contain statements describing chemical structure and key physicochemical properties, and links to databases with experimental data, such as MassBank (Horai et al., 2010; Wohlgemuth et al., 2016) and PDB Ligand (Shin, 2004), and toxicological information, such as the EPA CompTox Dashboard (Williams et al., 2017). Additionally, these items contain links to compound classes, disease indications, pharmaceutical products, and protein targets.

Pathways: Wikidata has items for almost three thousand human biological pathways, primarily from two established public pathway repositories: Reactome (Fabregat et al., 2018) and WikiPathways (Slenter et al., 2018). The full details of the different pathways remain with the respective primary sources. Our bots enter data for Wikidata properties such as pathway name, identifier, organism, and the list of component genes, proteins, and chemical compounds. Properties for contributing authors (via ORCID properties; Sprague, 2017), descriptions and ontology annotations are also being added for Wikidata pathway entries.

Diseases: Wikidata has items for over 16 thousand diseases, the majority of which were created based on imports from the Human Disease Ontology (Schriml et al., 2019), with additional disease terms added from the Monarch Disease Ontology (Mungall et al., 2017). Disease attributes include medical classifications, symptoms, relevant drugs, as well as subclass relationships to higher-level disease categories. In instances where the Human Disease Ontology specifies a related anatomic region and/or a causative organism (for infectious diseases), corresponding statements are also added.

References: Whenever practical, the provenance of each statement added to Wikidata was also added in a structured format. References are part of the core data model for a Wikidata statement. References can either cite the primary resource from which the statement was retrieved (including details like version number of the resource), or they can link to a Wikidata item corresponding to a publication as provided by a primary resource (as an extension of the WikiCite project; Ayers et al., 2019), or both. Wikidata contains over 20 million items corresponding to publications across many domain areas, including a heavy emphasis on biomedical journal articles.

Figure 1 with 1 supplement see all
A simplified class-level diagram of the Wikidata knowledge graph for biomedical entities.

Each box represents one type of biomedical entity. The header displays the name of that entity type (e.g., pharmaceutical product) and the number of Wikidata items for that entity type. The lower portion of each box displays a partial listing of attributes about each entity type and the number of Wikidata items for each attribute. Edges between boxes represent the number of Wikidata statements corresponding to each combination of subject type, predicate, and object type. For example, there are 1505 statements with 'pharmaceutical product' as the subject type, 'therapeutic area' as the predicate, and 'disease' as the object type. For clarity, edges for reciprocal relationships (e.g., 'has part' and 'part of') are combined into a single edge, and scientific articles (which are widely cited in statement references) have been omitted. All counts of Wikidata items are current as of September 2019. The most common data sources cited as references are available in Figure 1—source data 1. Data are generated using the code in https://github.com/SuLab/genewikiworld (archived at Mayers et al., 2020). A more complete version of this graph diagram can be found at https://commons.wikimedia.org/wiki/File:Biomedical_Knowledge_Graph_in_Wikidata.svg.

Figure 1—source data 1

Most frequent data sources cited as references for the biomedical subset of the Wikidata knowledge graph shown in Figure 1.

https://cdn.elifesciences.org/articles/52614/elife-52614-fig1-data1-v1.csv

Bot automation

To programmatically upload biomedical knowledge to Wikidata, we developed a series of computer programs, or bots. Bot development began by reaching a consensus on data modeling with the Wikidata community, particularly the Molecular Biology WikiProject. We then coded each bot to retrieve, transform, normalize and upload data from a primary resource to Wikidata via the Wikidata application programming interface (API).

We generalized the common code modules into a Python library, called Wikidata Integrator (WDI), to simplify the process of creating Wikidata bots (https://github.com/SuLab/WikidataIntegrator; archived at Burgstaller-Muehlbacher et al., 2020). Relative to accessing the API directly, WDI has convenient features that improve the bot development experience. These features include the creation of items for scientific articles as references, basic detection of data model conflicts, automated detection of items needing update, detailed logging and error handling, and detection and preservation of conflicting human edits.

Just as important as the initial data upload is the synchronization of updates between the primary sources and Wikidata. We utilized Jenkins, an open-source automation server, to automate all our Wikidata bots. This system allows for flexible scheduling, job tracking, dependency management, and automated logging and notification. Bots are either run on a predefined schedule (for continuously updated resources) or when new versions of original databases are released.

Applications of Wikidata

Translating between identifiers from different databases is one of the most common operations in bioinformatics analyses. Unfortunately, these translations are most often done by bespoke scripts and based on entity-specific mapping tables. These translation scripts are repetitively and redundantly written across our community and are rarely kept up to date, nor integrated in a reusable fashion.

An identifier translation service is a simple and straightforward application of the biomedical content in Wikidata. Based on mapping tables that have been imported, Wikidata items can be mapped to databases that are both widely- and rarely-used in the life sciences community. Because all these mappings are stored in a centralized database and use a systematic data model, generic and reusable translation scripts can easily be written (Figure 2). These scripts can be used as a foundation for more complex Wikidata queries, or the results can be downloaded and used as part of larger scripts or analyses.

Generalizable SPARQL template for identifier translation.

SPARQL is the primary query language for accessing Wikidata content. These simple SPARQL examples show how identifiers of any biological type can easily be translated using SPARQL queries. The top query demonstrates the translation of a small list of gene symbols (wdt:P353) to Entrez Gene IDs (wdt:P351), while the bottom example shows conversion of RxNorm concept IDs (wdt:P3345) to NDF-RT IDs (wdt:P2115). These queries can be submitted to the Wikidata Query Service (WDQS; https://query.wikidata.org/) to get real-time results. Translation to and from a wide variety of identifier types can be performed using slight modifications on these templates, and relatively simple extensions of these queries can filter mappings based on the statement references and/or qualifiers. A full list of Wikidata properties can be found at https://www.wikidata.org/wiki/Special:ListProperties. Note that for translating a large number of identifiers, it is often more efficient to perform a SPARQL query to retrieve all mappings and then perform additional filtering locally.

There are a number of other tools that are also aimed at solving the identifier translation use case, including the BioThings APIs (Xin et al., 2018), BridgeDb (van Iersel et al., 2010), BioMart (Smedley et al., 2015), UMLS (Bodenreider, 2004), and NCI Thesaurus (de Coronado et al., 2009). Relative to these tools, Wikidata distinguishes itself with a unique combination of the following: an almost limitless scope including all entities in biology, chemistry, and medicine; a data model that can represent exact, broader, and narrow matches between items in different identifier namespaces (beyond semantically imprecise 'cross-references'); programmatic access through web services with a track record of high performance and high availability.

Moreover, Wikidata is also unique as it is the only tool that allows real-time community editing. So while Wikidata is certainly not complete with respect to identifier mappings, it can be continually improved independent of any centralized effort or curation authority. As a database of assertions and not of absolute truth, Wikidata is able to represent conflicting information (with provenance) when, for example, different curation authorities produce different mappings between entities. (However, as with any bioinformatics integration exercise, harmonization of cross-references between resources can include relationships other than ‘exact match’. These instances can lead to Wikidata statements that are not explicitly declared, but rather the result of transitive inference.)

Integrative Queries

Wikidata contains a much broader set of information than just identifier cross-references. Having biomedical data in one centralized data resource facilitates powerful integrative queries that span multiple domain areas and data sources. Performing these integrative queries through Wikidata obviates the need to perform many time-consuming and error-prone data integration steps.

As an example, consider a pulmonologist who is interested in identifying candidate chemical compounds for testing in disease models (schematically illustrated in Figure 3). They may start by identifying genes with a genetic association to any respiratory disease, with a particular interest in genes that encode membrane-bound proteins (for ease in cell sorting). They may then look for chemical compounds that either directly inhibit those proteins, or finding none, compounds that inhibit another protein in the same pathway. Because they have collaborators with relevant expertise, they may specifically filter for proteins containing a serine-threonine kinase domain.

A representative SPARQL query that integrates data from multiple data resources and annotation types.

This example integrative query incorporates data on genetic associations to disease, Gene Ontology annotations for cellular compartment, protein target information for compounds, pathway data, and protein domain information. Specifically, this query (depicted schematically at right) retrieves genes that are (i) associated with a respiratory system disease, (ii) that encode a membrane-bound protein, and (iii) that sit within the same biochemical pathway as (iv) a second gene encoding a protein with a serine-threonine kinase domain and (v) a known inhibitor, and reports a list of those inhibitors. Aspects related to Disease Ontology in blue; aspects related to biochemistry in red/orange; aspects related to chemistry in green. Properties are shown in italics. Real-time query results can be viewed at https://w.wiki/6pZ.

Almost any competent informatician can perform the query described above by integrating cell localization data from Gene Ontology annotations, genetic associations from GWAS Catalog, disease subclass relationships from the Human Disease Ontology, pathway data from WikiPathways and Reactome, compound targets from the IUPHAR Guide to Pharmacology, and protein domain information from InterPro. However, actually performing this data integration is a time-consuming and error-prone process. At the time of publication of this manuscript, this Wikidata query completed in less than 10 s and reported 31 unique compounds. Importantly, the results of that query will always be up-to-date with the latest information in Wikidata.

This query, and other example SPARQL queries that take advantage of the rich, heterogeneous knowledge network in Wikidata are available at https://www.wikidata.org/wiki/User:ProteinBoxBot/SPARQL_Examples. That page additionally demonstrates federated SPARQL queries that perform complex queries across other biomedical SPARQL endpoints. Federated queries are useful for accessing data that cannot be included in Wikidata directly due to limitations in size, scope, or licensing.

Crowdsourced curation

Ontologies are essential resources for structuring biomedical knowledge. However, even after the initial effort in creating an ontology is finalized, significant resources must be devoted to maintenance and further development. These tasks include cataloging cross references to other ontologies and vocabularies, and modifying the ontology as current knowledge evolves. Community curation has been explored in a variety of tasks in ontology curation and annotation (see, for example, Bunt et al., 2012; Gil et al., 2017; Putman et al., 2019; Putman et al., 2017; Wang et al., 2016). While community curation offers the potential of distributing these responsibilities over a wider set of scientists, it also has the potential to introduce errors and inconsistencies.

Here, we examined how a crowd-based curation model through Wikidata works in practice. Specifically, we designed a hybrid system that combines the aggregated community effort of many individuals with the reliability of expert curation. First, we created a system to monitor, filter, and prioritize changes made by Wikidata contributors to items in the Human Disease Ontology. We initially seeded Wikidata with disease items from the Disease Ontology (DO) starting in late 2015. Beginning in 2018, we compared the disease data in Wikidata to the most current DO release on a monthly basis.

In our first comparison between Wikidata and the official DO release, we found that Wikidata users added a total of 2030 new cross references to GARD (Lewis et al., 2017) and MeSH (https://www.nlm.nih.gov/mesh/meshhome.html). These cross references were primarily added by a small handful of users through a web interface focused on identifier mapping (Manske, 2020). Each cross reference was manually reviewed by DO expert curators, and 2007 of these mappings (98.9%) were deemed correct and therefore added to the ensuing DO release. 771 of the proposed mappings could not be easily validated using simple string matching, and 754 (97.8%) of these were ultimately accepted into DO. Each subsequent monthly report included a smaller number of added cross references to GARD and MeSH, as well as ORDO (Maiella et al., 2018), and OMIM (Amberger and Hamosh, 2017; McKusick, 2007), and these entries were incorporated after expert review at a high approval rate (>90%).

Addition of identifier mappings represents the most common community contribution, and likely the most accessible crowdsourcing task. However, Wikidata users also suggested numerous refinements to the ontology structure, including changes to the subclass relationships and the addition of new disease terms. These structural changes were more nuanced and therefore rarely incorporated into DO releases with no modifications. Nevertheless, they often prompted further review and refinement by DO curators in specific subsections of the ontology.

The Wikidata crowdsourcing curation model is generalizable to any other external resource that is automatically synced to Wikidata. The code to detect changes and assemble reports is tracked online at https://github.com/SuLab/scheduled-bots (archived at Stupp et al., 2020) and can easily be adapted to other domain areas. This approach offers a novel solution for integrating new knowledge into a biomedical ontology through distributed crowdsourcing while preserving control over the expert curation process. Incorporation into Wikidata also enhances exposure and visibility of the resource by engaging a broader community of users, curators, tools, and services.

Interactive pathway pages

In addition to its use as a repository for data, we explored the use of Wikidata as a primary access and visualization endpoint for pathway data. We used Scholia, a web app for displaying scholarly profiles for a variety of Wikidata entries, including individual researchers, research topics, chemicals, and proteins (Nielsen et al., 2017). Scholia provides a more user-friendly view of Wikidata content with context and interactivity that is tailored to the entity type.

We contributed a Scholia profile template specifically for biological pathways (Scholia, 2019). In addition to essential items such as title and description, these pathway pages include an interactive view of the pathway diagram collectively drawn by contributing authors. The WikiPathways identifier property in Wikidata informs the Scholia template to source a pathway-viewer widget from Toolforge (https://tools.wmflabs.org/admin/tool/pathway-viewer) that in turn retrieves the corresponding interactive pathway image. Embedded into the Scholia pathway page, the widget provides pan and zoom, plus links to gene, protein and chemical Scholia pages for every clickable molecule on the pathway diagram see, for example, Scholia (2019). Each pathway page also includes information about the pathway authors. The Scholia template also generates a participants table that shows the genes, proteins, metabolites, and chemical compounds that play a role in the pathway, as well as citation information in both tabular and chart formats.

With Scholia template views of Wikidata, we were able to generate interactive pathway pages with comparable content and functionality to that of dedicated pathway databases. Wikidata provides a powerful interface to access these biological pathway data in the context of other biomedical knowledge, and Scholia templates provide rich, dynamic views of Wikidata that are relatively simple to develop and maintain.

Phenotype based disease diagnosis

Phenomizer is a web application that suggests clinical diagnoses based on an array of patient phenotypes (Köhler et al., 2009). On the back end, the latest version of Phenomizer uses BOQA, an algorithm that uses ontological structure in a Bayesian network (Bauer et al., 2012). For phenotype-based disease diagnosis, BOQA takes as input a list of phenotypes (using the Human Phenotype Ontology [HPO; Köhler et al., 2017]) and an association file between phenotypes and diseases. BOQA then suggests disease diagnoses based on semantic similarity (Köhler et al., 2009). Here, we studied whether phenotype-disease associations from Wikidata could improve BOQA’s ability to make differential diagnoses for certain sets of phenotypes. We modified the BOQA codebase to accept arbitrary inputs and to be able to run from the command line (code available at https://github.com/SuLab/boqa; archived at Köhler and Stupp, 2020) and also wrote a script to extract and incorporate the phenotype-disease annotations in Wikidata (code available at https://github.com/SuLab/Wikidata-phenomizer; archived at Tu et al., 2020).

As of September 2019, there were 273 phenotype-disease associations in Wikidata that were not in the HPO's annotation file (which contained a total of 172,760 associations). Based on parallel biocuration work by our team, many of these new associations were related to the disease Congenital Disorder of Deglycosylation (CDDG; also known as NGLY-1 deficiency) based on two papers describing patient phenotypes (Enns et al., 2014; Lam et al., 2017). To see if the Wikidata-sourced annotations improved the ability of BOQA to diagnose CDDG, we ran our modified version using the phenotypes taken from a third publication describing two siblings with suspected cases of CDDG (Caglayan et al., 2015). Using these phenotypes and the annotation file supplemented with Wikidata-derived associations, BOQA returned a much stronger semantic similarity to CDDG relative to the HPO annotation file alone (Figure 4). Analyses with the combined annotation file reported CDDG as the top result for each of the past 14 releases of the HPO annotation file, whereas CDDG was never the top result when run without the Wikidata-derived annotations.

BOQA analysis of suspected cases of the disease Congenital Disorder of Deglycosylation (CDDG).

We used an algorithm called BOQA to rank potential diagnoses based on clinical phenotypes. Here, clinical phenotypes from two cases of suspected CDDG patients were extracted from a published case report (Caglayan et al., 2015). These phenotypes were run through BOQA using phenotype-disease annotations from the Human Phenotype Ontology (HPO) alone, or from a combination of HPO and Wikidata. This analysis was tested using several versions of disease-phenotype annotations (shown along the x-axis). The probability score for CDDG is reported on the y-axis. These results demonstrate that the inclusion of Wikidata-based disease-phenotype annotations would have significantly improved the diagnosis predictions from BOQA at earlier time points prior to their official inclusion in the HPO annotation file. Details of this analysis can be found at https://github.com/SuLab/Wikidata-phenomizer (archived at Tu et al., 2020).

This result demonstrated an example scenario in which Wikidata-derived annotations could be a useful complement to expert curation. This example was specifically chosen to illustrate a favorable case, and the benefit of Wikidata would likely not currently generalize to a random sampling of other diseases. Nevertheless, we believe that this proof-of-concept demonstrates the value of the crowd-based Wikidata model and may motivate further community contributions.

Drug repurposing

The mining of graphs for latent edges has been an area of interest in a variety of contexts from predicting friend relationships in social media platforms to suggesting movies based on past viewing history. A number of groups have explored the mining of knowledge graphs to reveal biomedical insights, with the open source Rephetio effort for drug repurposing as one example (Himmelstein et al., 2017). Rephetio uses logistic regression, with features based on graph metapaths, to predict drug repurposing candidates.

The knowledge graph that served as the foundation for Rephetio was manually assembled from many different resources into a heterogeneous knowledge network. Here, we explored whether the Rephetio algorithm could successfully predict drug indications on the Wikidata knowledge graph. Based on the class diagram in Figure 1, we extracted a biomedically-focused subgraph of Wikidata with 19 node types and 41 edge types. We performed five-fold cross validation on drug indications within Wikidata and found that Rephetio substantially enriched the true indications in the hold-out set. We then downloaded historical Wikidata versions from 2017 and 2018 and observed marked improvements in performance over time (Figure 5). We also performed this analysis using an external test set based on Drug Central, which showed a similar improvement in Rephetio results over time (Figure 5—figure supplement 1).

Figure 5 with 1 supplement see all
Drug repurposing using the Wikidata knowledge graph.

We analyzed three snapshots of Wikidata using Rephetio, a graph-based algorithm for predicting drug repurposing candidates (Himmelstein et al., 2017). We evaluated the performance of the Rephetio algorithm on three historical versions of the Wikidata knowledge graph, quantified based on the area under the receiver operator characteristic curve (AUC). This analysis demonstrated that the performance of Rephetio in drug repurposing improved over time based only on improvements to the underlying knowledge graph. Details of this analysis can be found at https://github.com/SuLab/WD-rephetio-analysis (archived at Mayers and Su, 2020).

This analysis demonstrates the value of a community-maintained, centralized knowledge base to which many researchers are contributing. It suggests that scientific analyses based on Wikidata may continually improve irrespective of any changes to the underlying algorithms, but simply based on progress in curating knowledge through the distributed, and largely uncoordinated efforts of the Wikidata community.

Outlook

We believe that the design of Wikidata is well-aligned with the FAIR data principles. 

Findable: Wikidata items are assigned globally unique identifiers with direct cross-links into the massive online ecosystem of Wikipedias. Wikidata also has broad visibility within the Linked Data community and is listed in the life science registries FAIRsharing (https://fairsharing.org/; Sansone et al., 2019) and Identifiers.org (Wimalaratne et al., 2018). Wikidata has already attracted a robust, global community of contributors and consumers. 

Accessible: Wikidata provides access to its underlying knowledge graph via both an online graphical user interface and an API, and access includes both read- and write-privileges. Wikidata provides database dumps at least weekly (https://www.wikidata.org/wiki/Wikidata:Database_download), ensuring the long-term accessibility of the Wikidata knowledge graph independent of the organization and web application. Finally, Wikidata is also natively multilingual. 

Interoperable: Wikidata items are extensively cross-linked to other biomedical resources using Universal Resource Identifiers (URIs), which unambiguously anchor these concepts in the Linked Open Data cloud (Jacobsen et al., 2018). Wikidata is also available in many standard formats in computer programming and knowledge management, including JSON, XML, and RDF. 

Reusable: Data provenance is directly tracked in the reference section of the Wikidata statement model. The Wikidata knowledge graph is released under the Creative Commons Zero (CC0) Public Domain Declaration, which explicitly declares that there are no restrictions on downstream reuse and redistribution.

The open data licensing of Wikidata is particularly notable. The use of data licenses in biomedical research has rapidly proliferated, presumably in an effort to protect intellectual property and/or justify long-term grant funding (see, for example, Reiser et al., 2016). However, even seemingly innocuous license terms (like requirements for attribution) still impose legal requirements and therefore expose consumers to legal liability. This liability is especially problematic for data integration efforts, in which the license terms of all resources (dozens or hundreds or more) must be independently tracked and satisfied (a phenomenon referred to as 'license stacking'). Because it is released under CC0, Wikidata can be freely and openly used in any other resource without any restriction. This freedom greatly simplifies and encourages downstream use, albeit at the cost of not being able to incorporate ontologies or datasets with more restrictive licensing.

In addition to simplifying data licensing, Wikidata offers significant advantages in centralizing the data harmonization process. Consider the use case of trying to get a comprehensive list of disease indications for the drug bupropion. The National Drug File – Reference Terminology (NDF-RT) reported that bupropion may treat nicotine dependence and attention deficit hyperactivity disorder, the Inxight database listed major depressive disorder, and the FDA Adverse Event Reporting System (FAERS) listed anxiety and bipolar disorder. While no single database listed all these indications, Wikidata provided an integrated view that enabled seamless query and access across resources. Integrating drug indication data from these individual data resources was not a trivial process. Both Inxight and NDF-RT mint their own identifiers for both drugs and diseases. FAERS uses Medical Dictionary for Regulatory Activities (MedDRA) names for diseases and free-text names for drugs (Stupp and Su, 2018). By harmonizing and integrating all resources in the context of Wikidata, we ensure that those data are immediately usable by others without having to repeat the normalization process. Moreover, by harmonizing data at the time of data loading, consumers of that data do not need to perform the repetitive and redundant work at the point of querying and analysis.

As the biomedical data within Wikidata continues to grow, we believe that its unencumbered use will spur the development of many new innovative tools and analyses. These innovations will undoubtedly include the machine learning-based mining of the knowledge graph to predict new relationships (also referred to as knowledge graph reasoning; Das et al., 2017; Lin et al., 2018; Xiong et al., 2017).

For those who subscribe to this vision for cultivating a FAIR and open graph of biomedical knowledge, there are two simple ways to contribute to Wikidata. First, owners of data resources can release their data using the CC0 declaration. Because Wikidata is released under CC0, it also means that all data imported in Wikidata must also use CC0-compatible terms (e.g., be in the public domain). For resources that currently use a restrictive data license primarily for the purposes of enforcing attribution or citation, we encourage the transition to CC0 (+BY), a model that "move[s] the attribution from the legal realm into the social or ethical realm by pairing a permissive license with a strong moral entreaty’ (Cohen, 2013). For resources that must retain data license restrictions, consider releasing a subset of data or older versions of data using CC0. Many biomedical resources were created under or transitioned to CC0 (in part or in full) in recent years , including the Disease Ontology (Schriml et al., 2019), Pfam (El-Gebali et al., 2019), Bgee (Bastian et al., 2008), WikiPathways (Slenter et al., 2018), Reactome (Fabregat et al., 2018), ECO (Chibucos et al., 2014), and CIViC (Griffith et al., 2017).

Second, informaticians can contribute to Wikidata by adding the results of data parsing and integration efforts to Wikidata as, for example, new Wikidata items, statements, or references. Currently, the useful lifespan of data integration code typically does not extend beyond the immediate project-specific use. As a result, that same data integration process is likely performed repetitively and redundantly by other informaticians elsewhere. If every informatician contributed the output of their effort to Wikidata, the resulting knowledge graph would be far more useful than the stand-alone contribution of any single individual, and it would continually improve in both breadth and depth over time. Indeed, the growth of biomedical data in Wikidata is driven not by any centralized or coordinated process, but rather the aggregated effort and priorities of Wikidata contributors themselves.

FAIR and open access to the sum total of biomedical knowledge will improve the efficiency of biomedical research. Capturing that information in a centralized knowledge graph is useful for experimental researchers, informatics tool developers and biomedical data scientists. As a continuously-updated and collaboratively-maintained community resource, we believe that Wikidata has made significant strides toward achieving this ambitious goal.

Data availability

Links to all data and code used in this manuscript have been provided.

References

  1. Report
    1. Ayers P
    2. Mietchen D
    3. Orlowitz J
    4. Proffitt M
    5. Rodlund S
    6. Seiver E
    7. Taraborelli D
    8. Vershbow B
    (2019)
    WikiCite 2018-2019: Citations for the Sum of All Human Knowledge
    Wikimedia Foundation.
  2. Book
    1. Bastian F
    2. Parmentier G
    3. Roux J
    4. Moretti S
    5. Laudet V
    6. Robinson-Rechavi M
    (2008) Bgee: Integrating and Comparing Heterogeneous Transcriptome Data Among Species
    In: Bairoch A, Cohen-Boulakia S, Froidevaux C, editors. Data Integration in the Life Sciences, Lecture Notes in Computer Science. Berlin Heidelberg: Springer. pp. 124–131.
    https://doi.org/10.1007/978-3-540-69828-9_12
    1. Cohen D
    (2013) CC0 (+BY)
    CC0 (+BY), https://dancohen.org/2013/11/26/cc0-by/.
  3. Book
    1. Gil Y
    2. Garijo D
    3. Ratnakar V
    4. Khider D
    5. Emile-Geay J
    6. McKay N
    (2017) A Controlled Crowdsourcing Approach for Practical Ontology Extensions and Metadata Annotations
    In: d’Amato C, Fernandez M, Tamma V, Lecue F, Cudré-Mauroux P, Sequeda J, Lange C, Heflin J, editors. The Semantic Web – ISWC 2017, Lecture Notes in Computer Science. Springer International Publishing. pp. 231–246.
    https://doi.org/10.1007/978-3-319-68204-4
    1. Jacobsen A
    2. Kaliyaperumal R
    3. Stupp GS
    4. Schriml LM
    5. Thompson M
    6. Su AI
    7. Roos M
    (2018)
    Proceedings of the 11th International Conference Semantic Web Applications and Tools for Life Sciences, {SWAT4LS} 2018, Antwerp, Belgium, December 3-6, 2018
    Wikidata as an intuitive resource towards semantic data modeling in data FAIRification, Proceedings of the 11th International Conference Semantic Web Applications and Tools for Life Sciences, {SWAT4LS} 2018, Antwerp, Belgium, December 3-6, 2018, 2275, CEUR-WS.org.
    1. Manske M
    (2020) Mix’n’match
    Mix’n’match, https://tools.wmflabs.org/mix-n-match/#/.
  4. Book
    1. Nielsen FA
    2. Mietchen D
    3. Willighagen E
    (2017) Scholia, Scientometrics and Wikidata
    In: Blomqvist E, Hose K, Paulheim H, Ławrynowicz A, Ciravegna F, Hartig O, editors. The Semantic Web: ESWC 2017 Satellite Events, Lecture Notes in Computer Science. Springer International Publishing. pp. 237–259.
    https://doi.org/10.1007/978-3-319-70407-4_36
    1. Scholia
    (2019) Scholia, ACE Inhibitor Pathway
    Scholia, ACE Inhibitor Pathway, https://tools.wmflabs.org/scholia/pathway/Q29892242.
    1. Sprague ER
    (2017) ORCID
    Journal of the Medical Library Association 105:207–208.
    https://doi.org/10.5195/jmla.2017.89
  5. Conference
    1. Vrandečić D
    (2012) Wikidata: a new platform for collaborative data collection
    Proceedings of the 21st International Conference on World Wide Web.
    https://doi.org/10.1145/2187980.2188242
    1. Wang M
    2. Carver JJ
    3. Phelan VV
    4. Sanchez LM
    5. Garg N
    6. Peng Y
    7. Nguyen DD
    8. Watrous J
    9. Kapono CA
    10. Luzzatto-Knaan T
    11. Porto C
    12. Bouslimani A
    13. Melnik AV
    14. Meehan MJ
    15. Liu WT
    16. Crüsemann M
    17. Boudreau PD
    18. Esquenazi E
    19. Sandoval-Calderón M
    20. Kersten RD
    21. Pace LA
    22. Quinn RA
    23. Duncan KR
    24. Hsu CC
    25. Floros DJ
    26. Gavilan RG
    27. Kleigrewe K
    28. Northen T
    29. Dutton RJ
    30. Parrot D
    31. Carlson EE
    32. Aigle B
    33. Michelsen CF
    34. Jelsbak L
    35. Sohlenkamp C
    36. Pevzner P
    37. Edlund A
    38. McLean J
    39. Piel J
    40. Murphy BT
    41. Gerwick L
    42. Liaw CC
    43. Yang YL
    44. Humpf HU
    45. Maansson M
    46. Keyzers RA
    47. Sims AC
    48. Johnson AR
    49. Sidebottom AM
    50. Sedio BE
    51. Klitgaard A
    52. Larson CB
    53. P CAB
    54. Torres-Mendoza D
    55. Gonzalez DJ
    56. Silva DB
    57. Marques LM
    58. Demarque DP
    59. Pociute E
    60. O'Neill EC
    61. Briand E
    62. Helfrich EJN
    63. Granatosky EA
    64. Glukhov E
    65. Ryffel F
    66. Houson H
    67. Mohimani H
    68. Kharbush JJ
    69. Zeng Y
    70. Vorholt JA
    71. Kurita KL
    72. Charusanti P
    73. McPhail KL
    74. Nielsen KF
    75. Vuong L
    76. Elfeki M
    77. Traxler MF
    78. Engene N
    79. Koyama N
    80. Vining OB
    81. Baric R
    82. Silva RR
    83. Mascuch SJ
    84. Tomasi S
    85. Jenkins S
    86. Macherla V
    87. Hoffman T
    88. Agarwal V
    89. Williams PG
    90. Dai J
    91. Neupane R
    92. Gurr J
    93. Rodríguez AMC
    94. Lamsa A
    95. Zhang C
    96. Dorrestein K
    97. Duggan BM
    98. Almaliti J
    99. Allard PM
    100. Phapale P
    101. Nothias LF
    102. Alexandrov T
    103. Litaudon M
    104. Wolfender JL
    105. Kyle JE
    106. Metz TO
    107. Peryea T
    108. Nguyen DT
    109. VanLeer D
    110. Shinn P
    111. Jadhav A
    112. Müller R
    113. Waters KM
    114. Shi W
    115. Liu X
    116. Zhang L
    117. Knight R
    118. Jensen PR
    119. Palsson BO
    120. Pogliano K
    121. Linington RG
    122. Gutiérrez M
    123. Lopes NP
    124. Gerwick WH
    125. Moore BS
    126. Dorrestein PC
    127. Bandeira N
    (2016) Sharing and community curation of mass spectrometry data with global natural products social molecular networking
    Nature Biotechnology 34:828–837.
    https://doi.org/10.1038/nbt.3597

Decision letter

  1. Peter Rodgers
    Senior and Reviewing Editor; eLife, United Kingdom
  2. Chris Mungall
    Reviewer

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Thank you for submitting your article "Wikidata as a FAIR knowledge graph for the life sciences" for consideration by eLife. Your article has been reviewed by two peer reviewers, and the evaluation has been overseen by the eLife Features Editor (Peter Rodgers). The following individuals involved in review of your submission have agreed to reveal their identity: Chris Mungall (Reviewer #3).

The reviewers have discussed the reviews with one another and the Features Editor has drafted this decision to help you prepare a revised submission. We hope you will be able to submit the revised version within two months.

SUMMARY

The manuscript describes the life sciences component of the Wikidata knowledge base, which combines multiple disaggregated knowledge/databases into a single source enabling integrated querying. The authors describe the process for keeping Wikidata up to date, and they describe compelling use cases showing the power. Although there are many efforts in the life sciences to create integrated knowledge bases / knowledge graphs, Wikidata is unique in the breadth, scope, and in the community/crowdsourcing aspects. It is an increasingly important resource in the biomedical landscape, and the manuscript provides a clear description of the authors' significant efforts in building this component of the resource. However, the manuscript would benefit from addressing a number of issues in greater detail - see below.

ESSENTIAL REVISIONS:

1. An argument is made that Wikidata combines centralized and distributed approaches. This is an interesting concept. It would be important to know the relative importance of those two approaches. In other words, what proportion of edits is centralized (bots) and what proportion is distributed (contributors). Authors repeatedly refer to the fact that anyone is empowered to add new content, but how much of this potential is realized, and how this compares to other Wiki projects, e.g., Wikipedia?

2. Performance metrics and measurable assessment are often missing from the description of the applications. For example, when the process of translation of identifiers from various databases is described, no information is provided if the translation is done deterministically or probabilistically, how it is done algorithmically, how the performance can be evaluated (AUC ROC?), how Wikidata compares to other systems (e.g., UMLS). It is admirable that the authors admit that "Wikidata is certainly not complete with respect to identifier mappings...", but readers would benefit from knowing more about the proportion of mappings available and those not, and how this differs across databases integrated within Wikidata.

3. In such a broad overview, including multiple use cases, some selectivity in the presentation of examples is probably inevitable. This is understandable. Yet one can make those choices more informed, and offer, in addition, examples less favorable to the system (Wikidata), and in this way provide a balanced assessment. For example, on pages 10-11, for the purpose of disease diagnosis, annotations from the Human Phenotype Ontology (HPO) were compared to combined HPO + Wikidata annotations. However, comparison to only one set of annotations with respect to only one disease could be uninformative. At least, performance should be evaluated against a set of different randomly selected diseases.

4. An argument is made that the "Wikidata is among the most FAIR biomedical resources available". What are the other resources, and how Wikidata compares to those resources with respect to the four FAIR criteria?

5. The FAIR principles do not address the issue of completeness, as they were designed for datasets which are inherently self-contained and presumably complete with respect to the experiment/study. For biomedical knowledge bases (KBs), especially warehouse-style, a key question is: how complete is the KB? How often will my queries have false negatives? (Of course, KBs are always incomplete due to both knowledge and curation gaps. However, it is still important to provide users a sense of the areas in which there are major gaps, for example due to inability to integrate a key resource due to licensing).

An example is on figure 1: there are only 636 disease to symptom/phenotype associations. This is a very small amount, when compared to reference resources like the Human Phenotype Ontology. This may have negative consequences for applications built around Wikidata, causing people to lose trust. Another example is in the subset of genes imported - from 200 species. What determines the set of species included? Are there performance implications in including all genes in all species?

Of course, this is a difficult challenge for any KB, and completeness is not expected. I would still recommend a section that acknowledges and addresses this, addressing questions such as:

- how is a user best able to determine whether a portion of Wikidata is complete enough for their use case

- what is the process for deciding what is included vs excluded?

6. Identifier merging details: The identifier mapping use case is well articulated, but some details are omitted. Some identifier mapping resources store pairwise mappings without privileging any one resource. Wikidata is more like UMLS, in which they mint a new concept ID for the unified concept, and map everything to this. The challenge with this scheme is deciding on the criteria for lumping and splitting. Many source mappings are not 1:1, which can lead to excessive merging when these are traversed transitively

- For bot-derived entities, what is the algorithm for doing this?

- Is one source (e.g. NCBI gene) taken as canonical?

Additionally, it would be useful to see a stricter comparison with sources like bridgedb, in terms of completeness and consistency.

7. Trustworthiness of curation: The strength of Wikidata is in the crowdsourcing of knowledge. While this can be scaled up more easily due less of a need for funded curators, the downside is that the information may be less accurate and reliable. The paper provides good evidence of reliability via the disease cross-referencing experiment, in which 99% of 2030 crowd-sourced mappings were reviewed and accepted.

I am a little skeptical of the generalizability of these results. Mappings are generally quite easy and can be done with reasonable reliability by an automated string-matching process (with some caveats). It would be useful to know more about how many different people contributed to the 2030 (was it one person running a script)?

Refinements to the ontology structure are harder, and it's not clear how often these were incorporated. The mapping results are still a nice example, but they need to be qualified more, and there should be more discussion on reliability.

8. Phenomizer analysis: It's not clear if there is circularity in the Phenomizer analysis:

- were there publications that incorporated the case reports that were then annotated? Was this controlled for?

- What happened on 2018-09 when the HPO-only semsim score jumps? Were the WD annotations incorporated into HPO? Or was this independent annotation? The subsequent drop is curious.

9. Some lines read as if written for a grant proposal, e.g., "Wikidata has a proven track record for leveraging...", "Wikimedia Foundation... has a long track record of developing and maintaining..." see Introduction. In many instances, Wikidata is presented as "unique". Please reword such sentences/passages.

CODE/DATA

The Phenomizer analysis can't be reproduced at the moment, as the curated case reports with HPO IDs are not made available. Additionally, the settings for Phenomizer should be made available, also Phenomizer provides p-values, it would be useful to see this in the analysis results.

https://doi.org/10.7554/eLife.52614.sa1

Author response

We thank you and the reviewers for the helpful and detailed evaluation of our manuscript. We have made many changes in response to these comments, and believe the manuscript is substantially strengthened as a result. A point-by-point description of these changes is included below. We repeat the reviewers’ points here in italic, followed by our response in plain text.

SUMMARY

The manuscript describes the biomedical life sciences component of the Wikidata knowledge base, which combines multiple disaggregated knowledge/databases into a single source enabling integrated querying. The authors describe the process for keeping Wikidata up to date, and they describe compelling use cases showing the power. Although there are many efforts in the life sciences to create integrated knowledge bases / knowledge graphs, Wikidata is unique in the breadth, scope, and in the community/crowdsourcing aspects. It is an increasingly important resource in the biomedical landscape, and the manuscript provides a clear description of the authors' significant efforts in building this component of the resource. However, the manuscript would benefit from addressing a number of issues in greater detail - see below.

ESSENTIAL REVISIONS:

1. An argument is made that Wikidata combines centralized and distributed approaches. This is an interesting concept. It would be important to know the relative importance of those two approaches. In other words, what proportion of edits is centralized (bots) and what proportion is distributed (contributors). Authors repeatedly refer to the fact that anyone is empowered to add new content, but how much of this potential is realized, and how this compares to other Wiki projects, e.g., Wikipedia?

The Wikimedia Statistics tracker shows that over the last two years, 46% of all edits on Wikidata were attributed to normal "user" accounts while ~54% are attributed to accounts that are registered as bots, with the trend over the last year showing an increasing trend toward non-bot user edits ((Wikimedia Foundation, 2020); Figure 1-figure supplement 1 in revised manuscript). For comparison, that same site reports that English Wikipedia had 69% user edits and 14% bot edits (and an additional 17% anonymous edits).

In addition to the statistics above that demonstrate a balance between human and bot edits, the absolute number of editors is noteworthy. The number of "active editors" (five or more edits per month) has steadily grown from 10k to 12k over the last two years, with the number of bot accounts staying relatively stable over that time.

Finally, we also note that while Wikidata bots tend to focus on importing large-scale 'centralized' resources, the bot community itself is made up of the decentralized and distributed efforts of many bot developers.

These points have been clarified in the manuscript (including the addition of a new Figure 1-figure supplement 1).

2. Performance metrics and measurable assessment are often missing from the description of the applications. For example, when the process of translation of identifiers from various databases is described, no information is provided if the translation is done deterministically or probabilistically, how it is done algorithmically, how the performance can be evaluated (AUC ROC?), how Wikidata compares to other systems (e.g., UMLS). It is admirable that the authors admit that "Wikidata is certainly not complete with respect to identifier mappings...", but readers would benefit from knowing more about the proportion of mappings available and those not, and how this differs across databases integrated within Wikidata.

It is important to note that Wikidata is designed to be a database of assertions (with provenance) and not a database that attempts to resolve disagreements in search of absolute truth. Therefore, all identifier mappings in Wikidata are added deterministically – each Wikidata statement is an assertion from a database or curation authority. If different mapping resources disagree, then Wikidata can and should reflect that disagreement. Each consumer of Wikidata content is then free to add filters to prioritize or ignore specific mappings according to their own rules and biases. However, we also recognize that identifiers are a special class of statements because they establish the entities that are the subjects/objects of other statements. We provide more discussion of these topics below in response to point #6.

3. In such a broad overview, including multiple use cases, some selectivity in the presentation of examples is probably inevitable. This is understandable. Yet one can make those choices more informed, and offer, in addition, examples less favorable to the system (Wikidata), and in this way provide a balanced assessment. For example, on pages 10-11, for the purpose of disease diagnosis, annotations from the Human Phenotype Ontology (HPO) were compared to combined HPO + Wikidata annotations. However, comparison to only one set of annotations with respect to only one disease could be uninformative. At least, performance should be evaluated against a set of different randomly selected diseases.

Our goal was not to conclusively state that HPO + Wikidata is uniformly better for Phenomizer analyses. We acknowledge that the NGLY1 deficiency example was specifically selected to demonstrate a favorable case, and that the benefit of Wikidata would likely not currently generalize to a random sampling of other diseases. In this manuscript, we have attempted to balance examples in which Wikidata already has a rigorously-proven benefit to researchers, and proof-of-concept examples that demonstrate the value of the crowd-based Wikidata model and that motivate community contributions. Phenotype-based disease classification falls into the latter category. We have clarified this point in the manuscript.

4. An argument is made that the "Wikidata is among the most FAIR biomedical resources available". What are the other resources, and how Wikidata compares to those resources with respect to the four FAIR criteria?

Our intent is primarily to describe how Wikidata aligns with the FAIR data principles, and not to perform a comparative analysis of FAIRness relative to other biomedical resources. Therefore, we have reworded the text accordingly.

5. The FAIR principles do not address the issue of completeness, as they were designed for datasets which are inherently self-contained and presumably complete with respect to the experiment/study. For biomedical knowledge bases (KBs), especially warehouse-style, a key question is: how complete is the KB? How often will my queries have false negatives? (Of course, KBs are always incomplete due to both knowledge and curation gaps. However, it is still important to provide users a sense of the areas in which there are major gaps, for example due to inability to integrate a key resource due to licensing).

An example is on figure 1: there are only 636 disease to symptom/phenotype associations. This is a very small amount, when compared to reference resources like the Human Phenotype Ontology. This may have negative consequences for applications built around Wikidata, causing people to lose trust. Another example is in the subset of genes imported - from 200 species. What determines the set of species included? Are there performance implications in including all genes in all species?

Of course, this is a difficult challenge for any KB, and completeness is not expected. I would still recommend a section that acknowledges and addresses this, addressing questions such as:

- how is a user best able to determine whether a portion of Wikidata is complete enough for their use case

- what is the process for deciding what is included vs excluded?

There are many important issues raised in this point. Our overarching strategy for conveying the completeness is the class diagram in Figure 1. We believe that that Figure contains key information on the number of each entity type and the numbers of relationships between entity types. This information is relevant for each user to assess whether the appropriate data exist within Wikidata for their particular use case.

What is not included in Figure 1 are denominators – estimates of the total number of nodes/edges of a given type. Those denominators are extremely difficult to provide, since there are many reasonable ways to define them. For example, they could be based on resources that have been identified and screened but not yet loaded, based on resources that have been identified but are not suitable for Wikidata (e.g., for licensing reasons), based on an estimate of all structured knowledge (and/or unstructured knowledge) that is currently known, or based on an estimate of all knowledge that will eventually be known and discovered. We believe that the wide differences in how those denominators would be interpreted (and significant challenges in computing them) would counter any benefit to potential users. Instead, we have added a new supplemental table that shows the data sources that are cited as references for the most common properties. While this is not exactly the information that was requested by the reviewers, we believe this information will be more clearly interpretable.

A related question is raised here about how we prioritize resources for inclusion in Wikidata. In principle, we could write bots to load nearly any suitable data resources we come across. However, the bot writing process is actually a small proportion of the overall effort necessary to load a new resource. Each additional data resource requires a time-consuming data modeling process (in coordination with the broader Wikidata community and often the maintainers of the resource to be imported), adds overhead to our bot maintenance requirements, and introduces complexity in our data synchronization routines. Therefore, we avoid indiscriminate "stamp collecting" efforts in favor of targeted data loading driven by our own data-mining priorities. Accordingly, readers should feel empowered to import new data of interest to them, as that has been the primary mechanism for Wikidata's growth.

These points have been clarified in the manuscript.

6. Identifier merging details: The identifier mapping use case is well articulated, but some details are omitted. Some identifier mapping resources store pairwise mappings without privileging any one resource. Wikidata is more like UMLS, in which they mint a new concept ID for the unified concept, and map everything to this. The challenge with this scheme is deciding on the criteria for lumping and splitting. Many source mappings are not 1:1, which can lead to excessive merging when these are traversed transitively.

- For bot-derived entities, what is the algorithm for doing this?

- Is one source (e.g. NCBI gene) taken as canonical?

Wikidata bots generally do take one identifier as canonical (e.g., NCBI gene for genes, UniProt for proteins), but there is no requirement that other bots nor human editors use that same canonical identifier. Wikidata is based on a community process for harmonizing disconnected data resources. That means that any differing views on identifier mappings, like all statements on Wikidata items, are resolved via community discussion and consensus. Community input is solicited at many levels, starting at the approval process for bots to import large resources, all the way to discussions over individual statements on individual Wikidata items.

So while the process for resolving identifier mappings is the same as for any other Wikidata statements, we recognize that choices in identifier mappings have greater downstream implications (related to, for example, transitivity). However, the community-focused design of Wikidata means that there is no systematic scheme for making consistent lumping and splitting decisions. We recognize (and the original Wikidata creators recognized) that this design choice is not without its flaws, but also that it does have significant advantages with respect to usability at query time.

Finally, it is important to recognize that the issue of lumping and splitting is not specific to Wikidata, but one that exists across the biomedical knowledge management community. This issue has been exacerbated by the widespread use of semantically-imprecise cross references ('hasDbXref') in biomedical ontologies (as explained in this detailed blog post from Chris Mungall (Mungall, 2019)). So while we agree that this issue is not completely solved within Wikidata, we believe that the core issue lies further upstream in the knowledge management ecosystem. (If, for example, source databases and ontologies differentiated exact cross-references from non-exact cross references, Wikidata would be able to more precisely model these relationships.)

We have clarified these points in the manuscript.

Additionally, it would be useful to see a stricter comparison with sources like bridgedb, in terms of completeness and consistency.

We believe that a quantitative comparison to other mapping resources like BridgeDb would not be meaningful for a variety of reasons. First, both Wikidata and BridgeDb aggregate mappings from other 'authoritative' community resources (e.g., Ensembl for genes and proteins, ChEBI for chemicals, etc.). So a comparison between Wikidata and BridgeDb would really boil down to a comparison of the resources that were imported into each system. Second, for chemical compounds, BridgeDb actually imports mappings from Wikidata because it has a clear and transparent data model, and it allows people to fix inconsistencies and add missing content. Third, the scope of these resources is drastically different. BridgeDb currently focuses on genes, proteins, metabolic reactions, and metabolites, while Wikidata also includes many additional closely-related entity types (e.g., variants, diseases, organisms) as well as many more distantly-related types (e.g. clinical trials, people, countries).

7. Trustworthiness of curation: The strength of Wikidata is in the crowdsourcing of knowledge. While this can be scaled up more easily due less of a need for funded curators, the downside is that the information may be less accurate and reliable. The paper provides good evidence of reliability via the disease cross-referencing experiment, in which 99% of 2030 crowd-sourced mappings were reviewed and accepted.

I am a little skeptical of the generalizability of these results. Mappings are generally quite easy and can be done with reasonable reliability by an automated string-matching process (with some caveats). It would be useful to know more about how many different people contributed to the 2030 (was it one person running a script)?

Refinements to the ontology structure are harder, and it's not clear how often these were incorporated. The mapping results are still a nice example, but they need to be qualified more, and there should be more discussion on reliability.

We agree that identifier mapping represents the most common community contribution, and likely the most accessible crowdsourcing task. And the majority of the disease identifier mappings were performed by a handful of editors through the automated Mix'n'Match ID mapping interface (currently seven users with > 20 mappings). So while disease mapping through Mix'n'Match is skewed to a relatively small number of editors, the broader identifier mapping activities across all of Mix'n'Match is distributed across a much larger set of users, each with expertise and/or interest in a different area.

Regarding automated mapping via string matching, we agree that in many instances, automated methods are effective. However, in the case of the 2030 proposed MeSH and GARD mappings we reported, 771 (38%) were based on something other than simple string matching. We believe this supports the idea that crowdsourcing human mappings from Wikidata is a useful bridge between automated methods and expert curation.

We chose to focus on identifier mapping because the evaluation of accuracy by expert curators was relatively straightforward. We agree the refinements to ontology structure and the addition of other statements are likely to be more difficult and error-prone, and as we allude to in the manuscript, much more difficult to quantitatively evaluate for accuracy.

We have clarified all of these points in the manuscript.

8. Phenomizer analysis: It's not clear if there is circularity in the Phenomizer analysis:

- were there publications that incorporated the case reports that were then annotated? Was this controlled for?

- What happened on 2018-09 when the HPO-only semsim score jumps? Were the WD annotations incorporated into HPO? Or was this independent annotation? The subsequent drop is curious.

As part of a separate manuscript focused on NGLY1 deficiency (in revision), we curated disease-related phenotypes from two papers by Enns et al. (Enns et al., 2014) and Lam et al. (Lam et al., 2017, p. 1). These phenotypes were added as statements in Wikidata. In our analysis for this paper, we extracted phenotypes from suspected cases of CDDG / NGLY1-deficiency described in a third paper by Caglayan et al. (Caglayan et al., 2015) and used those as input to the Phenomizer analysis. So while our team was responsible for the majority of curation of phenotypes associated with NGLY1 deficiency, there was no "circularity" in the sense that different phenotype sets were used as the Phenomizer input and Wikidata additions. We have revised the manuscript to clarify this point.

After curation of the Enns et al. and Lam et al. papers, we also submitted those annotations for review and curation by curators with the Human Phenotype Ontology (https://github.com/obophenotype/human-phenotype-ontology/issues/3287). Based on annotations contributed by us and others, new phenotypes associated with CDDG / NGLY1-deficiency were added over the period shown in Figure 4. The sharp increase in the probability score around 2018-09 was primarily due to the addition of corneal ulceration (HP:0012804), which is a phenotype that is relatively specific for CDDG / NGLY1-deficiency and one of the phenotypes from Caglayan et al. that served as an input to the analysis. While there is a slight decline in the HPO-only probability score after 2018-09, we do not believe that reflects substantial changes to the underlying phenotypes annotated to CDDG / NGLY1-deficiency.

9. Some lines read as if written for a grant proposal, e.g., "Wikidata has a proven track record for leveraging...", "Wikimedia Foundation... has a long track record of developing and maintaining..." see Introduction. In many instances, Wikidata is presented as "unique". Please reword such sentences/passages.

We believe that the substance of those sentences is both correct and important to emphasize to readers for several reasons. First, many initiatives claim to be based on "crowdsourcing", but the vast majority fail to build and/or sustain a critical mass of users. It is noteworthy that the Wikimedia Foundation (WMF) has been successful at this on multiple occasions, and that Wikidata is one such example of a mature community-based resource.Second, the bioinformatics community continually generates innovative new tools, but the scalability of these tools are rarely tested with large numbers of users and large datasets. The WMF runs several online information resources (including Wikipedia and Wikidata) that run at a high level of performance and availability, which are important features for information management infrastructure. Third, Wikidata is sustained by funding streams that are different from the vast majority of biomedical resources (which are mostly funded by the NIH). Insulation from the 4-5 year funding cycles that are typical of NIH-funded biomedical resources does make Wikidata quite unique.

We have reworded several sections to hopefully strike a better balance between factual observation and advocacy. However, the intent of these sentences was not (and is not) to gratuitously praise Wikidata, but to convey to the reader that Wikidata is different than other online biomedical resources in several fundamental ways. We believe these points are important to communicate to readers as they decide how much effort to devote to using and/or contributing to Wikidata.

CODE/DATA

The Phenomizer analysis can't be reproduced at the moment, as the curated case reports with HPO IDs are not made available. Additionally, the settings for Phenomizer should be made available, also Phenomizer provides p-values, it would be useful to see this in the analysis results.

On more careful examination, we realized that we were imprecise with how we described our analysis. The published version of Phenomizer (Köhler et al., 2009) and the original Phenomizer web tool (http://compbio.charite.de/phenomizer/) are based on an "ontological similarity search", which includes the computation of a p-value. The most recent version of Phenomizer ("Orphamizer"; http://compbio.charite.de/phenomizer_orphanet/) is based on a newer algorithm for Bayesian network inference (Bauer et al., 2012) and a software package called BOQA (Köhler, 2020), which outputs probability scores. Our results are based on a BOQA analysis (with and without phenotype annotations from Wikidata). We have updated the manuscript text accordingly.

With that clarification, a Jupyter notebook that demonstrates the entire BOQA analysis from raw files to published figure has been added to our previously-cited Github repository: https://github.com/SuLab/Wikidata-phenomizer/ (Tu et al., 2019) (now also archived at (Tu et al., 2020)).

References

Bauer S, Köhler S, Schulz MH, Robinson PN. 2012. Bayesian ontology querying for accurate and noise-tolerant semantic searches. Bioinforma Oxf Engl 28:2502–2508. doi:10.1093/bioinformatics/bts471

Caglayan AO, Comu S, Baranoski JF, Parman Y, Kaymakçalan H, Akgumus GT, Caglar C, Dolen D, Erson-Omay EZ, Harmanci AS, Mishra-Gorur K, Freeze HH, Yasuno K, Bilguvar K, Gunel M. 2015. NGLY1 mutation causes neuromotor impairment, intellectual disability, and neuropathy. Eur J Med Genet 58:39–43. doi:10.1016/j.ejmg.2014.08.008

Enns GM, Shashi V, Bainbridge M, Gambello MJ, Zahir FR, Bast T, Crimian R, Schoch K, Platt J, Cox R, Bernstein JA, Scavina M, Walter RS, Bibb A, Jones M, Hegde M, Graham BH, Need AC, Oviedo A, Schaaf CP, Boyle S, Butte AJ, Chen Rui, Chen Rong, Clark MJ, Haraksingh R, FORGE Canada Consortium, Cowan TM, He P, Langlois S, Zoghbi HY, Snyder M, Gibbs RA, Freeze HH, Goldstein DB. 2014. Mutations in NGLY1 cause an inherited disorder of the endoplasmic reticulum-associated degradation pathway. Genet Med Off J Am Coll Med Genet 16:751–758. doi:10.1038/gim.2014.22

Köhler S. 2020. BOQA | Phenomics and Machine Learning @ Berlin. https://phenomics.github.io/software-boqa.html

Köhler S, Schulz MH, Krawitz P, Bauer S, Dölken S, Ott CE, Mundlos C, Horn D, Mundlos S, Robinson PN. 2009. Clinical diagnostics in human genetics with semantic similarity searches in ontologies. Am J Hum Genet 85:457–464. doi:10.1016/j.ajhg.2009.09.003

Lam C, Ferreira C, Krasnewich D, Toro C, Latham L, Zein WM, Lehky T, Brewer C, Baker EH, Thurm A, Farmer CA, Rosenzweig SD, Lyons JJ, Schreiber JM, Gropman A, Lingala S, Ghany MG, Solomon B, Macnamara E, Davids M, Stratakis CA, Kimonis V, Gahl WA, Wolfe L. 2017. Prospective phenotyping of NGLY1-CDDG, the first congenital disorder of deglycosylation. Genet Med Off J Am Coll Med Genet 19:160–168. doi:10.1038/gim.2016.75

Mungall CJ. 2019. Never mind the logix: taming the semantic anarchy of mappings in ontologies. Monkeying OWL. https://douroucouli.wordpress.com/2019/05/27/never-mind-the-logix-taming-the-semantic-anarchy-of-mappings-in-ontologie/

Tu R, Stupp GS, Su AI. 2020. SuLab/Wikidata-phenomizer: Release v1.0 on 2020-01-15. Zenodo. doi:10.5281/zenodo.3609142

Tu R, Stupp GS, Su AI. 2019. SuLab/Wikidata-phenomizer 7b25781. https://github.com/SuLab/Wikidata-phenomizer

Wikimedia Foundation. 2020. Wikimedia Statistics - Wikidata - Edits. https://stats.wikimedia.org/v2/#/wikidata.org/contributing/edits/normal|line|2-year|editor_type~anonymous*group-bot*name-bot*user|monthly

https://doi.org/10.7554/eLife.52614.sa2

Article and author information

Author details

  1. Andra Waagmeester

    Andra Waagmeester is at Micelio, Antwerp, Belgium

    Contribution
    Conceptualization, Data curation, Software, Formal analysis, Validation, Visualization, Writing - review and editing
    Contributed equally with
    Gregory Stupp
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9773-4008
  2. Gregory Stupp

    Gregory Stupp is in the Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA, United States

    Contribution
    Conceptualization, Data curation, Software, Formal analysis, Validation, Visualization, Writing - original draft
    Contributed equally with
    Andra Waagmeester
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0644-7212
  3. Sebastian Burgstaller-Muehlbacher

    Sebastian Burgstaller-Muehlbacher is in the Center for Integrative Bioinformatics Vienna, Max Perutz Laboratories, University of Vienna and Medical University of Vienna, Vienna, Austria

    Contribution
    Conceptualization, Data curation, Software
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-4640-3510
  4. Benjamin M Good

    Benjamin M Good is in the Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA, United States

    Contribution
    Conceptualization, Data curation, Software, Supervision, Project administration
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7334-7852
  5. Malachi Griffith

    Malachi Griffith is in the McDonnell Genome Institute, Washington University School of Medicine, St. Louis, MO, United States

    Contribution
    Data curation, Funding acquisition
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-6388-446X
  6. Obi L Griffith

    Obi L Griffith is in the McDonnell Genome Institute, Washington University School of Medicine, St. Louis, MO, United States

    Contribution
    Data curation, Funding acquisition, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0843-4271
  7. Kristina Hanspers

    Kristina Hanspers is in the Institute of Data Science and Biotechnology, Gladstone Institutes, San Francisco, CA, United States

    Contribution
    Data curation, Software
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-5410-599X
  8. Henning Hermjakob

    Henning Hermjakob is at the European Bioinformatics Institute (EMBL-EBI), Hinxton, United Kingdom

    Contribution
    Data curation, Supervision, Funding acquisition
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-8479-0262
  9. Toby S Hudson

    Toby S Hudson is in the School of Chemistry, University of Sydney, Sydney, Australia

    Contribution
    Data curation, Validation
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-3348-3622
  10. Kevin Hybiske

    Kevin Hybiske is in the Division of Allergy and Infectious Diseases, Department of Medicine, University of Washington, Seattle, WA, United States

    Contribution
    Data curation, Funding acquisition
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-2967-3079
  11. Sarah M Keating

    Sarah M Keating is at the European Bioinformatics Institute (EMBL-EBI), Hinxton, United Kingdom

    Contribution
    Data curation, Software
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-3356-3542
  12. Magnus Manske

    Magnus Manske is at the Wellcome Trust Sanger Institute, Hinxton, United Kingdom

    Contribution
    Data curation, Software
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-5916-0947
  13. Michael Mayers

    Michael Mayers is in the Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA, United States

    Contribution
    Data curation, Software, Formal analysis, Visualization
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7792-0150
  14. Daniel Mietchen

    Daniel Mietchen is in the School of Data Science, University of Virginia, Charlottesville, VA, United States

    Contribution
    Data curation, Validation, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9488-1870
  15. Elvira Mitraka

    Elvira Mitraka is in the University of Maryland School of Medicine, Baltimore, MD, United States

    Contribution
    Data curation, Validation
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-0719-3485
  16. Alexander R Pico

    Alexander R Pico is in the Institute of Data Science and Biotechnology, Gladstone Institutes, San Francisco, CA, United States

    Contribution
    Data curation, Supervision, Writing - original draft, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-5706-2163
  17. Timothy Putman

    Timothy Putman is in the Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA, United States

    Contribution
    Data curation, Software
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4291-0737
  18. Anders Riutta

    Anders Riutta is in the Institute of Data Science and Biotechnology, Gladstone Institutes, San Francisco, CA, United States

    Contribution
    Data curation, Software
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4693-0591
  19. Nuria Queralt-Rosinach

    Núria Queralt-Rosinach is in the Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA, United States

    Contribution
    Data curation, Software
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-0169-8159
  20. Lynn M Schriml

    Lynn M Schriml is in the University of Maryland School of Medicine, Baltimore, MD, United States

    Contribution
    Data curation, Supervision, Funding acquisition
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-8910-9851
  21. Thomas Shafee

    Thomas Shafee is in the Department of Animal Plant and Soil Sciences, La Trobe University, Melbourne, Australia

    Contribution
    Data curation, Visualization, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-2298-7593
  22. Denise Slenter

    Denise Slenter is in the Department of Bioinformatics-BiGCaT, NUTRIM, Maastricht University, Maastricht, Netherlands

    Contribution
    Data curation, Software
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-8449-1318
  23. Ralf Stephan

    Ralf Stephan is a retired researcher based in Berlin, Germany

    Contribution
    Data curation, Validation
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4650-631X
  24. Katherine Thornton

    Katherine Thornton is at Yale University Library, Yale University, New Haven, CT, United States

    Contribution
    Software, Validation
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4499-0451
  25. Ginger Tsueng

    Ginger Tsueng is in the Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA, United States

    Contribution
    Data curation, Software
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9536-9115
  26. Roger Tu

    Roger Tu is in the Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA, United States

    Contribution
    Software, Formal analysis, Visualization
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7899-1604
  27. Sabah Ul-Hasan

    Sabah Ul-Hasan is in the Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA, United States

    Contribution
    Data curation, Software
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-6334-452X
  28. Egon Willighagen

    Egon Willighagen is in the Department of Bioinformatics-BiGCaT, NUTRIM, Maastricht University, Maastricht, Netherlands

    Contribution
    Data curation, Software
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-7542-0286
  29. Chunlei Wu

    Chunlei Wu is in the Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA, United States

    Contribution
    Data curation, Software, Supervision, Funding acquisition
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-2629-6124
  30. Andrew I Su

    Andrew I Su is in the Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA, United States

    Contribution
    Conceptualization, Formal analysis, Supervision, Funding acquisition, Validation, Writing - original draft, Project administration, Writing - review and editing
    For correspondence
    asu@scripps.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-9859-4104

Funding

National Institute of General Medical Sciences (R01 GM089820)

  • Andrew I Su

National Institute of General Medical Sciences (U54 GM114833)

  • Henning Hermjakob
  • Andrew I Su

National Institute of General Medical Sciences (R01 GM100039)

  • Alexander R Pico

National Human Genome Research Institute (R00HG007940)

  • Malachi Griffith

National Cancer Institute (U24CA237719)

  • Malachi Griffith

V Foundation for Cancer Research (V2018-007)

  • Malachi Griffith

National Institute of Allergy and Infectious Diseases (R01 AI126785)

  • Kevin Hybiske

National Center for Advancing Translational Sciences (UL1 TR002550)

  • Andrew I Su

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

The authors thank the thousands of Wikidata contributors for curating knowledge, both directly related and unrelated to this work, much of which has been organized under the WikiProjects for Molecular Biology, Chemistry and Medicine. The authors also thank the Wikimedia Foundation for financially supporting Wikidata, and many developers and administrators for maintaining Wikidata as a community resource.

Publication history

  1. Received: October 9, 2019
  2. Accepted: February 28, 2020
  3. Version of Record published: March 17, 2020 (version 1)

Copyright

© 2020, Waagmeester et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 9,072
    Page views
  • 515
    Downloads
  • 67
    Citations

Article citation count generated by polling the highest count across the following sources: Scopus, Crossref, PubMed Central.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Andra Waagmeester
  2. Gregory Stupp
  3. Sebastian Burgstaller-Muehlbacher
  4. Benjamin M Good
  5. Malachi Griffith
  6. Obi L Griffith
  7. Kristina Hanspers
  8. Henning Hermjakob
  9. Toby S Hudson
  10. Kevin Hybiske
  11. Sarah M Keating
  12. Magnus Manske
  13. Michael Mayers
  14. Daniel Mietchen
  15. Elvira Mitraka
  16. Alexander R Pico
  17. Timothy Putman
  18. Anders Riutta
  19. Nuria Queralt-Rosinach
  20. Lynn M Schriml
  21. Thomas Shafee
  22. Denise Slenter
  23. Ralf Stephan
  24. Katherine Thornton
  25. Ginger Tsueng
  26. Roger Tu
  27. Sabah Ul-Hasan
  28. Egon Willighagen
  29. Chunlei Wu
  30. Andrew I Su
(2020)
Science Forum: Wikidata as a knowledge graph for the life sciences
eLife 9:e52614.
https://doi.org/10.7554/eLife.52614
  1. Further reading

Further reading

    1. Computational and Systems Biology
    James D Brunner, Nicholas Chia
    Research Article Updated

    The microbial community composition in the human gut has a profound effect on human health. This observation has lead to extensive use of microbiome therapies, including over-the-counter ‘probiotic’ treatments intended to alter the composition of the microbiome. Despite so much promise and commercial interest, the factors that contribute to the success or failure of microbiome-targeted treatments remain unclear. We investigate the biotic interactions that lead to successful engraftment of a novel bacterial strain introduced to the microbiome as in probiotic treatments. We use pairwise genome-scale metabolic modeling with a generalized resource allocation constraint to build a network of interactions between taxa that appear in an experimental engraftment study. We create induced sub-graphs using the taxa present in individual samples and assess the likelihood of invader engraftment based on network structure. To do so, we use a generalized Lotka-Volterra model, which we show has strong ability to predict if a particular invader or probiotic will successfully engraft into an individual’s microbiome. Furthermore, we show that the mechanistic nature of the model is useful for revealing which microbe-microbe interactions potentially drive engraftment.

    1. Computational and Systems Biology
    Sean R Johnson, Meghana Peshwa, Zhiyi Sun
    Short Report

    Accurately detecting distant evolutionary relationships between proteins remains an ongoing challenge in bioinformatics. Search methods based on primary sequence struggle to accurately detect homology between sequences with less than 20% amino acid identity. Profile- and structure-based strategies extend sensitive search capabilities into this twilight zone of sequence similarity but require slow pre-processing steps. Recently, whole-protein and positional embeddings from deep neural networks have shown promise for providing sensitive sequence comparison and annotation at long evolutionary distances. Embeddings are generally faster to compute than profiles and predicted structures but still suffer several drawbacks related to the ability of whole-protein embeddings to discriminate domain-level homology, and the database size and search speed of methods using positional embeddings. In this work, we show that low-dimensionality positional embeddings can be used directly in speed-optimized local search algorithms. As a proof of concept, we use the ESM2 3B model to convert primary sequences directly into the 3D interaction (3Di) alphabet or amino acid profiles and use these embeddings as input to the highly optimized Foldseek, HMMER3, and HH-suite search algorithms. Our results suggest that positional embeddings as small as a single byte can provide sufficient information for dramatically improved sensitivity over amino acid sequence searches without sacrificing search speed.