Building and mining a heterogeneous biomedical knowledge graph
Download as PPTX, PDF0 likes147 views
This document summarizes Andrew Su's presentation on building and mining a heterogeneous biomedical knowledge graph. The presentation discusses integrating data from 29 public resources to create a knowledge graph called Hetionet containing over 47,000 nodes and 2.25 million relationships. It also describes Semantic MEDLINE, a database containing over 90 million biomedical relationships extracted from PubMed abstracts. Finally, it discusses challenges with using knowledge graphs for computational drug repurposing due to loss of predictive signals over time as relationships age and new knowledge is discovered.
![Querying a distributed KG
29
Query
template
Imatinib
[Disease]
[Gene]
[Pathway]
Key question
• Is this architecture
amenable to KB
completion?](https://image.slidesharecdn.com/20200623akbctrimmed-200623191243/85/Building-and-mining-a-heterogeneous-biomedical-knowledge-graph-29-320.jpg)
Building and mining a heterogeneous biomedical knowledge graph
- 1. Building and mining a heterogenous biomedical knowledge graph June 23, 2020 Slides: slideshare.net/andrewsu Andrew Su, Ph.D. @andrewsu http://sulab.org Q19857262
- 3. CML Pathway KEGG:hsa05220 Bcr-Abl UniProtKB:A9UF02 c-Kit UniProtKB:P10721 Imatinib CHEBI:45783 Mast Cell Degranulation GO:0043303 Chronic myelogenous leukemia DOID:8552 Asthma DOID:2841 inhibits inhibits part of part of causescauses treats treats? 3 3
- 4. 4
- 5. 5 Citation: https://doi.org/10.7554/eLife.26726 Method based on https://doi.org/10.1109/ASONAM.2011.112
- 6. “ 6 Daniel Himmelstein, https://think-lab.github.io/d/102/ We obtained and integrated data from 29 publicly available resources to create Hetionet v1.0. The hetnet contains 47,031 nodes of 11 types and 2,250,197 relationships of 24 types.
- 7. Algorithm: KB: Rephetio MINERVA, CBR, etc. Text-mining Semantic MEDLINE DB Crowd-curated Distributed KG 7
- 8. Semantic MEDLINE (SemMedDB) • Processed all PubMed abstracts (30 million+) every 6-12 months • 90 million+ triples covering key biomedical entity types (genes, drugs, diseases, anatomy, physiological processes, etc.) • All triples anchored to a publication (and publication date) 8 https://skr3.nlm.nih.gov/SemMedDB/
- 9. Time-resolved prediction analysis 9 1950 1970 1990 2010 1960 1980 2000 Drug-disease 1 Drug-disease 2 Drug-disease 3 Drug-disease 4 Drug-disease 5 Drug-disease 6 … Drug-disease n Approval dates KB year Train Test
- 10. Time-resolved prediction analysis 10 1950 1970 1990 2010 1960 1980 2000 Drug-disease 1 Drug-disease 2 Drug-disease 3 Drug-disease 4 Drug-disease 5 Drug-disease 6 … Drug-disease n Approval dates KB year Train Test
- 11. Time-resolved prediction analysis 11 1950 1970 1990 2010 1960 1980 2000 Drug-disease 1 Drug-disease 2 Drug-disease 3 Drug-disease 4 Drug-disease 5 Drug-disease 6 … Drug-disease n Approval dates KB year
- 12. Time-resolved prediction analysis 12 1950 1970 1990 2010 1960 1980 2000 Drug-disease 1 Drug-disease 2 Drug-disease 3 Drug-disease 4 Drug-disease 5 Drug-disease 6 … Drug-disease n Approval dates KB year
- 13. Time-resolved prediction analysis 13 1950 1970 1990 2010 1960 1980 2000 Drug-disease 1 Drug-disease 2 Drug-disease 3 Drug-disease 4 Drug-disease 5 Drug-disease 6 … Drug-disease n Approval dates KB year
- 14. Time-resolved prediction analysis 14 Area Under ROC https://doi.org/10.1186/s12859-019-3297-0
- 15. Time-resolved prediction analysis 15 Area Under ROC Area Under PRC https://doi.org/10.1186/s12859-019-3297-0 Key findings • Computational drug repurposing using KB reasoning is hard • We lose almost all predictive signal at < 5 years in the future
- 16. Time-resolved prediction analysis 16 Area Under ROC https://doi.org/10.1186/s12859-019-3297-0 Key findings • Computational drug repurposing using KB reasoning is hard • We lose almost all predictive signal at < 5 years in the future • Identified four edge types that are sensitive to information gain/loss compound – TREATS – disease compound – RELATEDTO – compound anatomy – LOCATIONOF – diseasedisease – ASSOCIATEDWITH – disease
- 17. Algorithm: KB: Rephetio MINERVA, CBR, etc. Crowd-curatedText-mining Semantic MEDLINE DB Distributed KG 17
- 18. is to data is to text biomedical Provide a database of the world’s knowledge that anyone can edit - Denny Vrandečić 18
- 20. Wikidata as a drug repositioning KB 20
- 21. Wikidata as a drug repositioning KB 21
- 22. Wikidata as a drug repositioning KB 22
- 23. Wikidata as a drug repositioning KB Key findings • Wikidata is comparable in size, scope, structure to be suitable for computational drug repositioning • Results improve over time as the KB content increases in size and quality 23 2017-01-16 2018-02-05 2019-09-13 https://doi.org/10.7554/eLife.52614
- 24. Algorithm: KB: Rephetio MINERVA, CBR, etc. Crowd-curated Distributed KGText-mining Semantic MEDLINE DB 24
- 25. Data federation is common, and sometimes necessary 25 20,000 genes 20,000genes Gene correlation matrices Health data Licensing restrictions The (Re)usable Data Project http://reusabledata.org/ Columbia Open Health Data http://cohd.io/
- 28. Notebooks ● Github ● colab Autonomous query planning and API call execution 28
- 29. Querying a distributed KG 29 Query template Imatinib [Disease] [Gene] [Pathway] Key question • Is this architecture amenable to KB completion?
- 30. Algorithm: KB: Evidence: Rephetio MINERVA, CBR, etc. Text-mining Crowd-curated Distributed KG Semantic MEDLINE DB DrugMechDB 30
- 31. CML Pathway KEGG:hsa05220 Bcr-Abl UniProtKB:A9UF02 c-Kit UniProtKB:P10721 Imatinib CHEBI:45783 Mast Cell Degranulation GO:0043303 Chronic myelogenous leukemia DOID:8552 Asthma DOID:2841 inhibits inhibits part of part of causescauses treats treats? 31
- 32. fomepizole Compound Alcohol dehydrogenase 1A Gene Ethanol oxidation Pathway formic acid Compound Methanol poisoning Disease MESH:D000077604 UniProt:P07327 REACT:R-HSA-71384 MESH:C030544 MESH:D000138 Inhibits Part of Produces Causes Indication: Fomepizole – TREATS – Methanol poisoning DrugMechDB – a database of drug mechanisms https://github.com/SuLab/DrugMechDB 32
- 33. Acknowledgements Mike Mayers Núria Queralt-Rosinach Toby Li Mike Mayers Andra Waagmeester (Micelio) Sabah Ul-Hasan Ginger Tsueng Roger Tu Greg Stupp Ben Good Tim Putman Sebastian Burgstaller-Muehlbacher Lynn Schriml (Univ. Maryland) Kevin Hybiske (Univ. Washington) Kevin Xin Chunlei Wu Marco Alvarado Jerry Zhou Semantic MEDLINE DB DrugMechDB: Mike Mayers 33 Funding