PaperlayerView exact revision on arXiv

Source-linked AI summary

A Review of Biomedical Datasets Relating to Drug Discovery: A Knowledge Graph Perspective

Stephen Bonner, Ian P Barrett, Cheng Ye, Rowan Swiers, Ola Engkvist, Andreas Bender, Charles Tapley Hoyt, William L Hamilton

arXiv:2102.10062v4cs.AI

TL;DR

Drug discovery is complex and costly, and suitable primary data is needed to construct useful KGs. This review surveys public data sources, existing KGs, and case studies; it concludes that KGs and machine learning show promising early applications, while data use and graph composition still need improvement.

  • Problem

    Drug discovery is complex and costly, and constructing a suitable KG requires capturing the correct primary data.

    Source: §1 · pp. 1–2

  • Method

    The review surveys public drug-discovery data sources for KG use, compares existing public KGs, and evaluates selected case studies.

    Source: §1 · p. 2

  • Results

    The case studies show considerable potential and successes for KGs in diverse drug-discovery tasks, while data use and graph composition remain areas for improvement.

    Source: §7.4 · p. 24, §9 · p. 26

  • Takeaways & Limitations

    KGs combined with machine learning have promising early applications in drug repositioning, drug-drug interactions, and gene prioritisation.

    Source: §9 · p. 26

    Using relations from multiple sources can duplicate edges and risk overlap between training and evaluation data if graph composition and evaluation are not handled carefully.

    Source: §6.2.3 · p. 20

Abstract

from arXiv · show

Drug discovery and development is a complex and costly process. Machine learning approaches are being investigated to help improve the effectiveness and speed of multiple stages of the drug discovery pipeline. Of these, those that use Knowledge Graphs (KG) have promise in many tasks, including drug repurposing, drug toxicity prediction and target gene-disease prioritisation. In a drug discovery KG, crucial elements including genes, diseases and drugs are represented as entities, whilst relationships between them indicate an interaction. However, to construct high-quality KGs, suitable data is required. In this review, we detail publicly available sources suitable for use in constructing drug discovery focused KGs. We aim to help guide machine learning and KG practitioners who are interested in applying new techniques to the drug discovery field, but who may be unfamiliar with the relevant data sources. The datasets are selected via strict criteria, categorised according to the primary type of information contained within and are considered based upon what information could be extracted to build a KG. We then present a comparative analysis of existing public drug discovery KGs and a evaluation of selected motivating case studies from the literature. Additionally, we raise numerous and unique challenges and issues associated with the domain and its datasets, whilst also highlighting key future research directions. We hope this review will motivate KGs use in solving key and emerging questions in the drug discovery domain.

1 Introduction

Drug discovery KGs can represent nuanced relationships among drugs, diseases, and genes, but constructing useful graphs depends on capturing appropriate primary data. This review surveys public data sources and existing KGs, outlines domain challenges, and identifies future research directions.

1 Introduction

  • KGs represent genes, diseases, and drugs as entities and capture different kinds of relationships, while drug-discovery work must contend with data reliability, integration, uncertainty, and contextual or quantitative information.The review notes that the domain has many curated public data sources, and that KG use could support machine-learning approaches.Source: §1 · p. 2
  • The review surveys public drug-discovery data sources for KG use, compares existing public KGs, and presents case studies, challenges, and future research directions.Its contributions also include an introduction to the domain for KG and machine-learning practitioners.Source: §1 · p. 2

1.1 Dataset Selection Criteria

  • The review selects datasets that are publicly accessible, regarded as high quality based primarily on their popularity in drug-discovery literature, and actively maintained.The review focuses on publicly accessible resources rather than commercial datasets.Source: §1.1 · p. 3

1.2 Review Organisation

  • The paper proceeds from background and prior work through ontologies, datasets, existing KGs, and case studies before discussing future challenges and conclusions.The review's stated aim is to signpost resources and challenges for researchers considering KG methods in drug discovery.Source: §1 · p. 2, §1.2 · p. 3

2 Background

The background introduces drug-discovery routes and tasks, defines KGs as heterogeneous graphs, and outlines opportunities and challenges in applying them to biomedical data. It emphasizes graph composition, uncertainty, evolving sources, bias, and careful evaluation.

2.1 An Introduction To Drug Discovery and Development

  • Drug discovery can proceed through target selection, phenotypic screening, or repositioning; candidate compounds are then tested in preclinical models and clinical trials for efficacy and toxicity.The review describes the field as multidisciplinary and driven by the need to address diseases or conditions lacking suitable treatments.Source: §2.1 · p. 3
  • Drug active molecules include small chemicals and antibodies as well as peptides, nucleotides, macromolecules, and polymers; newer types are collectively termed drug modalities.These types have different advantages and disadvantages and broaden potential drug targets beyond historical approaches alone.Source: §2.1 · p. 3

2.1.1 Subtasks Within Drug Discovery

  • Common subtasks include drug repositioning, disease-target identification, drug-target interaction prediction, drug-combination effects, and toxicity prediction.Disease-target identification is also called target identification or gene-disease prioritisation; drug-target interaction prediction is also called target binding or target activity.Source: §2.1.1 · p. 4

2.2 Knowledge Graphs

  • The review defines KGs as heterogeneous graphs with distinct vertex and edge types, representing edges as relation-bearing triples and vertices as typed entities.In drug discovery, entity types can distinguish genes, proteins, and drugs, while relation types can encode distinctions such as up- or down-regulation.Source: §2.2 · pp. 4–5, Figure 1 · p. 5

2.3 Knowledge Graph Use in Drug Discovery: Research Challenges

  • Biomedical relationships offer opportunities for new knowledge, and the review identifies potential at the intersection of graph analysis, machine learning, and quantitative systems pharmacology.The review presents these opportunities alongside challenges in constructing KGs for drug-discovery tasks.Source: §2.3 · pp. 5–6
  • Biomedical KGs pose graph-composition choices, combine heterogeneous data with differing confidence and context, and depend on sources that evolve over time.The STRING figure tracks increases in organisms, interactions, and proteins across major versions, underscoring the source-versioning issue.Source: §2.3 · pp. 5–6, Figure 2 · p. 6
  • Biomedical KG work must account for under-represented negative data and ascertainment bias, possible train-test leakage, relation-specific evaluation, and metrics that are not comparable across differently sized KGs.The review recommends meaningful data splits, reporting metrics by relation type, and caution when comparing evaluation implementations.Source: §2.3 · p. 6

3 Prior Work

Prior reviews cover drug-repurposing, drug-target, drug-drug interaction, and biomedical KG topics, but often give limited attention to the overall drug-discovery landscape and datasets as KG components.

3 Prior Work

  • Drug-repurposing reviews span databases, machine-learning methods, and KG-specific approaches, including one review that partitions resources into four information-based categories.The cited KG-specific review selected six datasets to construct a graph for experimental evaluation.Source: §3 · pp. 6–7
  • Other prior work reviews drug-target and drug-drug interaction prediction, drug-related databases, biomedical KG embeddings, and graph-based machine learning in drug discovery.The review characterizes one graph-based machine-learning survey as a method-focused counterpart that does not discuss suitable public datasets.Source: §3 · p. 7
  • Existing reviews often focus on a specific subtask or experimental methods, discuss datasets less extensively, or treat resources as databases rather than potential KG components.Some reviews are also written primarily from a biological perspective, which may be less accessible to machine-learning practitioners new to the domain.Source: §3 · p. 7

4 Biomedical Ontologies

Biomedical ontologies provide controlled representations for consistent description of biomedical concepts, but proliferating ontologies make cross-resource matching and merging difficult. The section surveys ontology representations, matching challenges, and examples relevant to drug discovery.

4 Biomedical Ontologies

  • Ontologies define and categorize domain concepts, supporting consistency when diseases and gene functions admit multiple interpretations.They capture key entities and relationships in a concise machine-readable format.Source: §4 · pp. 7–8

4.1 Ontology Representations

  • Biomedical ontologies commonly use OBO, RDFS, or OWL; OBO is biologically oriented, while RDFS and OWL come from the semantic-web context.OBO can express the terms, relationships, and properties required for an ontology.Source: §4.1 · p. 8

4.2 Ontology Matching and Merging

  • Ontology proliferation complicates disease-term alignment, and OXO and DODO mappings do not guarantee a direct match between every pair of ontologies.Some ontology pairs may have no mapping at all.Source: §4.2 · p. 8
  • Mapping terms can introduce logical inconsistencies, so ontology merging often requires manual intervention and is time-consuming and error-prone.The OBO Foundry offers rules and advice intended to make ontologies easier to merge and match.Source: §4.2 · p. 8

4.3 Ontology Overviews

  • The overview selects major ontologies relevant to drug-discovery tasks, while noting that a full review is beyond this work’s scope.Readers seeking a full review are directed to.Source: §4.3 · p. 8, Table 2 · p. 8

4.3.1 Disease Ontologies

  • Disease ontologies differ by use: DO links datasets, MeSH indexes literature, HPO describes observable disease traits, and MonDO harmonizes disease definitions.The named resources are DO, MeSH, HPO, and MonDO.Source: §4.3.1 · p. 8

4.3.2 Gene Related Ontologies

  • GO links genes to molecular activity, locations, and biological processes, whereas DTO relates gene products to drug-discovery concepts such as druggability.The resources are Gene Ontology (GO) and Drug Target Ontology (DTO).Source: §4.3.2 · pp. 8–9

4.3.3 Integrator Ontologies

  • EFO describes experimental variables across domains, and Open Targets uses it to annotate disease entities with descriptions, phenotypes, cross-references, synonyms, and classifications.Its scope includes disease, anatomy, cell types, cell lines, chemical compounds, and assays.Source: §4.3.3 · p. 9

5 Primary Domain-Specific Datasets

The section surveys trusted public datasets for key drug-discovery entities and their relationships, considering what information can be extracted for knowledge graphs. It covers integrated resources and gene or gene-product datasets, then compares the latter’s potential relations and features.

5 Primary Domain-Specific Datasets

  • The survey signposts popular, trusted public resources for genes, gene products, diseases, drugs, and their interactions, pathways, and processes.It organizes the datasets by primary information type and suggests relations they could contribute to a knowledge graph.Source: §5 · p. 9

5.1 Integrated Drug Discovery Resources

  • Open Targets combines data on genes, drugs, and diseases, and its target–disease associations include evidence-class scores aggregated into an overall association score.Those scores could weight knowledge-graph relations and provide a notion of confidence; the review reports that Open Targets had not yet been integrated into existing drug-discovery KGs.Source: §5.1 · p. 10
  • Pharos can provide protein–disease links, weighted protein–protein interactions, and protein-level structural and expression features.Pharos is the access point for TCRD, part of an NIH program.Source: §5.1 · p. 10

5.2 Gene and Gene Products

  • Gene and gene-product resources vary in curation: UniProtKB’s Swiss-Prot is expert-curated, while TrEMBL stores automatically extracted information.As of UniProt version 2020 6, TrEMBL contained 195M entities versus 563K in Swiss-Prot.Source: §5.2 · p. 10
  • Ensembl provides experimentally based automated annotation of variants, transcripts, and genomic positions, while RNAcentral integrates non-coding RNA data and identifier mappings across vocabularies.RNAcentral integrates 28 expert databases covering 296 organisms.Source: §5.2 · pp. 10–11
  • Entrez Gene’s GeneID is a unique integer that can translate between resources and serves as the primary gene identifier in many existing KGs.Entrez Gene combines curated and automatically extracted information.Source: §5.2 · p. 11

5.2.1 Gene and Gene Products Resource Comparison

  • Gene and protein resources support sequence, structure, and expression features, gene or protein interactions, and links to GO functional annotations.The resource comparison summarizes potential relations and features in Table 5.Source: §5.2.1 · p. 11, Table 5 · p. 11

5.3 Interactions, Pathways and Biological Processes

The section reviews resources for biological interactions, pathways, and processes, then compares the relations and features they may provide. These sources offer interaction and pathway links as well as interaction weights, text descriptions, and graph-based representations.

5.3 Interactions, Pathways and Biological Processes

  • The section details resources that link entities through interactions, processes, and pathways; interaction resources are presented in Table 6, while processes-and-pathways resources are detailed in Table 8.Source: §5.3 · p. 11, Table 6 · p. 11

5.3.1 Interaction Resources

  • STRING collects protein interactions across over 5K organisms, while BioGRID, IntAct, MINT, and OmniPath provide complementary interaction or signaling data.STRING combines experimentally curated and literature-mined sources; IntAct uses UniProt IDs, and OmniPath integrates over 100 literature-curated resources.Source: §5.3.1 · pp. 11–12

5.3.2 Interactions Resource Comparison

  • Interaction resources can yield gene–gene or protein–protein links; IntAct and BioGRID can also link gene products with compounds, and sources provide interaction weights or types.The comparison of potential relations and features is summarized in Table 7.Source: §5.3.2 · p. 12, Table 7 · p. 12

5.3.3 Pathway Resources

  • Pathway resources are expert-curated interaction subsets relevant to biological processes or pathogenic mechanisms, and their definitions can carry implicit biases.Harmonizing and combining definitions has been shown to mitigate these biases when possible.Source: §5.3.3 · p. 12
  • Reactome covers reactions and pathways across species, WikiPathways uses community curation, and KEGG provides manually curated pathways but restricts redistribution and bulk access.Reactome data already appears in Hetionet; WikiPathways is designed to complement resources such as Reactome.Source: §5.3.3 · pp. 12–13

5.3.4 Pathway Resource Comparison

  • All reviewed pathway resources support gene–pathway links; Reactome and KEGG also link drugs to affected pathways, and all offer text and graph representations.These pathway features are summarized in Table 9.Source: §5.3.4 · p. 13, Table 9 · p. 13

5.4 Diseases

The section surveys datasets focused on diseases and compares their potential relations and features. The resources primarily support gene–disease links, with some also providing drug, pathway, or disease-similarity links.

5.4 Diseases

  • The disease-resource survey includes KEGG DISEASE, DISEASES, DisGeNET, OMIM, and the GWAS Catalog.The primary sources are listed in Table 10.Source: §5.4 · p. 13, Table 10 · p. 13
  • The GWAS Catalog provides searchable, freely available SNP–trait associations from published and unpublished studies, with expert-curated literature data.The catalog covers associations reported by genome-wide association studies.Source: §5.4 · p. 14

5.4.1 Disease Resource Comparison

  • Disease datasets primarily support gene–disease links; KEGG DISEASE adds disease–drug and disease–pathway links, while DisGeNET includes disease-similarity links.All provide some link evidence, and KEGG DISEASE and OMIM include text descriptions that could be mined for features.Source: §5.4.1 · p. 14, Table 11 · p. 14

5.5 Drugs and Compounds

The section surveys drug and compound resources, then compares the relations and entity features they could contribute to a knowledge graph.

5.5 Drugs and Compounds

  • The surveyed resources range from literature-curated ChEMBL and approved-drug-focused DrugCentral to PubChem, which aggregates over 700 resources.DrugBank spans bioinformatics and cheminformatics, while BindingDB focuses on interactions between potential drug-target proteins and drug-like molecules.Source: §5.5 · pp. 14–15
  • RepoDB focuses on drug-disease links for repositioning and includes failed or withdrawn drugs, which could supply negative drug-disease edges.It covers both approved drugs and drugs that failed in clinical trials or were withdrawn.Source: §5.5 · p. 15

5.5.1 Drug Resource Comparison

  • Drug resources broadly provide drug-gene links, while PubChem and DrugBank include drug-drug interactions and DrugCentral offers potential drug-disease links.Almost all also provide compound structures and numerical attributes; the passage notes text descriptions in KEGG DISEASE and OMIM.Source: §5.5.1 · p. 15, Table 13 · p. 15

5.6 Dataset Evaluation

The evaluation compares dataset provenance and annotation, potential KG relations, and opportunities for enrichment, while noting practical constraints in using some resources.

5.6.1 Data Trust

  • Many resources have human curation, but this does not guarantee accuracy; predicted data and integrator resources also require attention to bias and replicated knowledge.Integrators such as Omnipath and PubChem aggregate primary datasets and can bring diverse information into a KG from one resource.Source: §5.6.1 · p. 15, Table 14 · p. 16

5.6.2 Relation Mining

  • Gene-Gene and Gene-Drug relations have many potential sources, whereas Disease-Pathway links appear in only one source.Using multiple resources for the same relation type requires care to avoid duplicated edges; many reviewed resources are already available as edgelists or RDF.Source: §5.6.2 · p. 16

5.6.3 Graph Enrichment

  • Entity features beyond relational interactions, such as UniProtKB protein sequence and functional properties, could enrich KGs but may require manual feature engineering.The passage describes that process as complicated, domain-specific, and iterative [9] [78].Source: §5.6.3 · p. 16

5.6.4 Untapped Resources

  • OpenTargets and Pharos have drug-discovery-specific information, but their formats require manual conversion for KG incorporation.The review notes that neither had yet been incorporated into a public KG.Source: §5.6.4 · p. 16

6 Existing Biomedical Knowledge Graphs

The section surveys existing biomedical KGs and compares their composition, source datasets, documentation, and reproducibility. It also identifies resource limitations and construction alternatives relevant to drug-discovery applications.

6.1 Biomedical Knowledge Graphs Overviews

  • DRKG builds on Hetionet, adding recent COVID-19-related data from STRING, DrugBank, and GNBR, plus precomputed GNN-based molecule embeddings; BioKG draws on 13 sources, while PharmKG integrates seven sources before quality filtering.DRKG has no associated features for entities other than molecules; at the time of writing, PharmKG's filtered graph and entity feature vectors had not been released.Source: §6.1 · pp. 17–18
  • OpenBioLink combines data from 17 datasets and includes explicit true negatives for selected relation types, while CKG integrates 25 databases and 10 ontologies but distributes a smaller graph because of licensing restrictions.CKG adds -omics focus, including protein-state information such as post-translational modifications.Source: §6.1 · p. 18

6.2.1 Graph Composition: Entities

  • Across the KGs, shared entities include gene products, compounds, and diseases; Hetionet and DRKG include more ancillary entity types, potentially suiting fine-grained tasks.BioKG and CKG model proteins rather than genes, and BioKG alone separates genetic disorders from diseases.Source: §6.2.1 · p. 19, Table 16 · p. 19

6.2.2 Graph Composition - Relations

  • KGs differ in relation granularity and volume: DRKG has many Gene-Gene and Drug-Drug edges, while Gene-Gene or Drug-Drug relations form a majority in many graphs.OpenBioLink has no drug-drug edges; fewer gene-disease examples may leave target-discovery models with less training data.Source: Figure 6 · p. 19, §6.2.2 · p. 20, Table 18 · p. 20

6.2.3 Underlying Dataset Use

  • The KGs reuse underlying sources such as DrugBank, while DRKG alone draws four relation types from GNBR and Hetionet combines multiple datasets for some relation types.DRKG's use of both IntAct and STRING for gene-gene edges may duplicate interactions, potentially allowing training edges to overlap with evaluation edges.Source: §6.2.3 · p. 20, Figure 7 · p. 21

6.2.4 Evaluation of Documentation Quality and Reproducibility

  • Hetionet has the highest overall documentation quality in the comparison, yet two KGs lack graph-construction code and none states planned updates.The evaluation considers schema, filtering, relation explanations, updates, data-relation mappings, construction code, and licensing information.Source: §6.2.4 · pp. 20–21, Table 19 · p. 21

6.3 Shortcomings of Existing KG Resources

  • The reviewed KGs lack update schedules, detailed documentation, entity or relation features, and dataset-version information, making some graphs harder to use for real-world problems.The passage specifically links missing documentation and source mappings to greater difficulty in real-world use.Source: §6.3 · pp. 21–22

6.4.1 Other Graph Datasets; 6.4.2 Graph Construction Resources

  • Other options include homogeneous BioSNAP graphs and construction frameworks such as PheKnowLator and Bio2BEL, which supports ad hoc KG assembly.Bio2BEL converts sources into BEL and semantically aligns and merges them; NDEx stores user-submitted biological graphs.Source: §6.4.1 · p. 22, §6.4.2 · p. 22

7 Case Studies

The review compares KG case studies across polypharmacy, drug–target interaction, and gene–disease prioritisation, highlighting both their predictive results and choices in graph construction. It concludes that graph composition and dataset selection remain unsettled.

7 Case Studies

  • The case studies span three tasks where KGs have been used: polypharmacy, drug–target interaction, and gene–disease prioritisation.The review selects one approach per area and discusses both successes and areas for improvement.Source: §7 · p. 22, Table 20 · p. 23

7.1 Polypharmacy Prediction

  • Decagon predicted existing and proposed novel adverse drug–drug interactions better than non-graph-specific and homogeneous models.Its bipartite graph uses drug and protein entities, with 964 drug–drug side-effect relation types and one type for drug–protein and protein–protein interactions.Source: §7.1 · pp. 22–23, Table 20 · p. 23

7.2 Drug-Target Interaction Prediction

  • TriModel treats drug–target interaction prediction as link prediction and outperformed traditional and non-relational graph methods.The KG combines 11 entity types and 26 relation types; the review notes that the paper does not precisely justify these choices and uses a single relation for gene–drug interactions.Source: §7.2 · p. 23

7.3 Gene-Disease Prioritisation

  • Rosalind’s ComplEx model outperformed competing methods, including OpenTarget, by up to 20% extra recall across 198 diseases; evaluation used therapeutic edges and also tested future-edge prediction.The time-sliced evaluation trains on historical data and predicts future edges.Source: §7.3 · p. 24

7.4 Evaluative Summary

  • Across the case studies, KG composition varies substantially, with no consensus on an optimal design for machine-learning pipelines.The studies also vary in the sources used for common relations and usually do not justify source selection; none uses an existing KG, instead building task-specific graphs.Source: §7.4 · p. 24

8 Future Challenges & Key Issues

The review identifies challenges in graph construction, data use, and evaluation that affect drug-discovery KGs. It calls for improved reproducibility, metadata, feature use, bias handling, and uncertainty representation.

Graph Composition

  • Building useful drug-discovery KGs requires many choices, especially when non-relational sources are transformed into graphs; the review calls for domain-expert collaboration and validated public KGs.Unified, reproducible graph-construction toolkits could simplify bespoke KG creation.Source: §8 · p. 24

Data Value

  • The value of adding data modalities remains insufficiently assessed, while collection, storage, and processing can be costly when larger datasets do not improve the task.The review also leaves open whether a single super graph or task-specific projections yield better predictions.Source: §8 · pp. 24–25

Better Metadata

  • KGs often omit the versions of source resources used, limiting reproducibility and assessment of how updated knowledge affects predictive performance.Metadata could also record whether relationships came from expert curation or automated pipelines and include alternative entity identifiers.Source: §8 · p. 25

Incorporation of Features

  • Many KGs provide little beyond edge lists, and the benefits of adding entity or relation features still need fair assessment.Structural information for proteins and chemicals is one potential feature source discussed in the review.Source: §8 · p. 25

Addressing Bias

  • Under-represented relationships, curation, and negative-sampling choices can bias model performance; the review recommends mitigation such as reweighting and domain-aware sampling.The review notes that many graphs lack true negative samples, while one benchmark graph has recently included them.Source: §8 · p. 25

Fair Evaluation

  • Duplicate edges across sources and trivial inverse relations can skew evaluation, so the review calls for meaningful splits, such as by disease or protein family.Expert-created public splits for key tasks could support fairer evaluation.Source: §8 · p. 25

Uncertainty

  • Biological KG data can be uncertain because of experimental sources or literature mining, yet that uncertainty is rarely represented in the graph.The review argues that encoding uncertainty could let methods learn from it and potentially improve prediction robustness.Source: §8 · p. 25

Reproducibility

  • Reproducibility remains a major KG-field issue: papers often report results without releasing the exact graph used.The review considers improvement essential for continued field development.Source: §8 · p. 25

9 Conclusion

The review surveys data sources and public drug-discovery KGs, while identifying challenges in using drug-discovery data. It concludes that KGs combined with machine learning have promising early applications across several tasks.

9 Conclusion

  • KGs combined with machine learning have promising early applications in drug repositioning, drug–drug interactions, and gene prioritisation.The review presents these as potential contributions to key drug-discovery challenges.Source: §9 · p. 26
  • The review surveys datasets as building blocks for a hypothetical drug-discovery KG and evaluates existing public KGs.It also highlights pitfalls and challenges to consider when choosing data sources.Source: §9 · p. 26
  • The authors hope the review will guide KG-mining and machine-learning researchers applying techniques to drug discovery.They connect this guidance to suitable data sources and recent work on graph-specific machine-learning models.Source: §9 · p. 26
Loading 2102.10062v4…