AstraZeneca - The promise of graphs & graph-based learning in drug discovery
0 likes384 views
The document discusses the promise of knowledge graphs and graph-based learning in drug discovery. It describes how graphs can intuitively represent the complex connections in biological data. The author details their process of building a knowledge graph from various data sources, using the graph for both predictive and exploratory applications like identifying new drug targets. Key learnings are the critical importance of data modeling, need for cross-disciplinary teams, and exploring multiple analytical methods rather than relying on a single approach.
![Analogy - if it was all about people:
• genes would be the general identifiable information: name, address, maybe social
security number etc
[Ufuk Kirik, The data scientist, Lives in Gothenburg]
20](https://image.slidesharecdn.com/neo4jsummit220906uk-220914130047-375d8c66/85/AstraZeneca-The-promise-of-graphs-graph-based-learning-in-drug-discovery-19-320.jpg)
![Analogy - if it was all about people:
• genes would be the general identifiable information: name, address, maybe social
security number etc
[Ufuk Kirik, The data scientist, Lives in Gothenburg]
• proteins would be the actual physical presence; what we do and how we look like
while doing that.
[The person standing in front of you, here I am talking, dressed smart casual]
21](https://image.slidesharecdn.com/neo4jsummit220906uk-220914130047-375d8c66/85/AstraZeneca-The-promise-of-graphs-graph-based-learning-in-drug-discovery-20-320.jpg)
![Analogy - if it was all about people:
• genes would be the general identifiable information: name, address, maybe social
security number etc
[Ufuk Kirik, The data scientist, Lives in Gothenburg]
• proteins would be the actual physical presence of me; what we do and how we look
like while doing that.
[The person standing in front of you, here I am talking, dressed smart casual]
• What a gene/protein is and what it does varies on the context
• Not only the activity changes but also with whom I do these activities change based
on time and location.
22](https://image.slidesharecdn.com/neo4jsummit220906uk-220914130047-375d8c66/85/AstraZeneca-The-promise-of-graphs-graph-based-learning-in-drug-discovery-21-320.jpg)
AstraZeneca - The promise of graphs & graph-based learning in drug discovery
- 1. The promise of graphs & graph-based learning in drug discovery Ufuk Kirik Data Science & Quantitative Biology, Discovery Sciences, R&D AstraZeneca Gothenburg, Sweden 2022-09-05
- 2. Outline • Background • Drug Discovery Landscape • Graphs in Life Science (Research) • The promise • How we use them • What have we learned so far • The “so what?” 2
- 3. Background 3 Who am I? Curious, nerdy kid Engineer/Applied mathematician Bioinformatician Data scientist Graph enthusiast
- 4. Background 4 Who am I? What do I do? • Lead a team working with KG Insights • Develop processes to enrich our graphs with a wide variety of in-house generated data • Develop tools to effectively explore and exploit our graphs, catering to a broad range of users. Curious, nerdy kid Engineer/Applied mathematician Bioinformatician Data scientist Graph enthusiast
- 5. Background What do I want to achieve with this talk? • Our journey into the graph world • Our challenges and approach to graph-based learning • Our learnings, generalized for wider audience 5 Who am I?
- 6. Background What do I want to achieve with this talk? • Our journey into the graph world • Our challenges and approach to graph-based learning • Our learnings, generalized for wider audience • Will not include any graphics or specifics from our graphs, due to the sensitive nature of our project. 6 Who am I?
- 7. Pre-clinical Clinical Follow-up Drug discovery landscape • What does it take to bring a new medicine to a patient? 7
- 8. Pre-clinical Clinical Follow-up Drug discovery landscape • What does it take to bring a new medicine to a patient? 8 Ph. I Ph. II Ph. III (x10)
- 9. Drug discovery landscape • What does it take to bring a new medicine to a patient? 9 Cook, D., Brown, D., Alexander, R. et al. Lessons learned from the fate of AstraZeneca's drug pipeline: a five-dimensional framework. Nat Rev Drug Discov 13, 419–431 (2014). https://doi.org/10.1038/nrd4309 Bottleneck: Phase II Reason(s): Efficacy & Safety Problem: Time & Money
- 10. Enter graphs…
- 11. What is so cool about graphs anyways? • Connected information everywhere • Graphs are a natural way to represent connected information • Graph theory is old and rich, most importantly it is incredibly useful 12
- 12. What is so cool about graphs anyways? • Connected information everywhere • Graphs are a natural way to represent connected information • Graph theory is old and rich, most importantly it is incredibly useful 13
- 13. What is so cool about graphs anyways? • Connected information everywhere • Graphs are a natural way to represent connected information • Graph theory is old and rich, most importantly it is incredibly useful 14
- 14. What is so cool about graphs anyways? • Connected information everywhere • Graphs are a natural way to represent connected information • Graph theory is old and rich, most importantly it is incredibly useful • ML on graphs is extremely useful: • Movie/Book/Product recommendations (e.g. Amazon, Netflix, Spotify) • Personal/Professional network (e.g. social media) • Like all things ML, sometimes it can go terribly wrong too! 15
- 15. Graphs in life sciences • Nothing works in isolation in the biological domain! 16
- 16. Graphs in life sciences • Nothing works in isolation in the biological domain! • Crash course in cell biology: • Central dogma DNA -> RNA -> Proteins -> Function 17
- 17. Graphs in life sciences • Nothing works in isolation in the biological domain! • Crash course in cell biology: • Central dogma DNA -> RNA -> Proteins -> Function • Pathways & Functional annotations • Cells & Tissues ... • Compounds & Diseases ... 18
- 18. Graphs in life sciences • Nothing works in isolation in the biological domain! 19
- 19. Analogy - if it was all about people: • genes would be the general identifiable information: name, address, maybe social security number etc [Ufuk Kirik, The data scientist, Lives in Gothenburg] 20
- 20. Analogy - if it was all about people: • genes would be the general identifiable information: name, address, maybe social security number etc [Ufuk Kirik, The data scientist, Lives in Gothenburg] • proteins would be the actual physical presence; what we do and how we look like while doing that. [The person standing in front of you, here I am talking, dressed smart casual] 21
- 21. Analogy - if it was all about people: • genes would be the general identifiable information: name, address, maybe social security number etc [Ufuk Kirik, The data scientist, Lives in Gothenburg] • proteins would be the actual physical presence of me; what we do and how we look like while doing that. [The person standing in front of you, here I am talking, dressed smart casual] • What a gene/protein is and what it does varies on the context • Not only the activity changes but also with whom I do these activities change based on time and location. 22
- 22. Analogy - if it was all about people: • If I do things I am not supposed to, or allowed to, then we end up having problems, which is essentially what happens in the case of most diseases. • Now that is where the analogy falls short, since although there is only 1 of me, proteins can, and typically do, have many copies co-existing at the same. • Nevertheless all of these interactions can be captured with a graph structure. I would actually go as far to speculate that a graph is the only intuitive way to represent information like this. 23
- 23. Biomedical knowledge graphs 24 Himmelstein, D. S., & Baranzini, S. E. (2015). Heterogeneous Network Edge Prediction: A Data Integration Approach to Prioritize Disease- Associated Genes. PLoS computational biology, 11(7), e1004259. https://doi.org/10.1371/journal.pcbi.1004259
- 24. How do we deploy graphs? • We refer to graphs in the context of knowledge graphs, where we aggregate data from various different sources; public and internal, conceptual and experimental alike. 25
- 25. Graph composition 26 Graph Composition • These data sources can then be transformed into a knowledge graph through a composition stage. • Unfortunately, many biases can be present during this phase.
- 26. How do we deploy graphs? • We refer to graphs in the context of knowledge graphs, where we aggregate data from various different sources; public and internal, conceptual and experimental alike. • We use these graphs for both predictive and explorative purposes 27
- 27. How do we deploy graphs? • We refer to graphs in the context of knowledge graphs, where we aggregate data from various different sources; public and internal, conceptual and experimental alike. • We use these graphs for both predictive and explorative purposes • Predictive examples: • which gene could be a good drug target for a given disease? • which compound could be a good/useful treatment for a given disease? 28
- 28. How do we deploy graphs? • We refer to graphs in the context of knowledge graphs, where we aggregate data from various different sources; public and internal, conceptual and experimental alike. • We use these graphs for both predictive and explorative purposes • Predictive examples: • which gene could be a good drug target for a given disease? • which compound could be a good/useful treatment for a given disease? • Explorative examples: • what do we know about gene X (in the context of disease D)? • what level of evidence do we have that supports/contradicts a putative use of compound C for a new indication (that is a new disease)? 29
- 29. Example • (Find all/some genes that associates with Parkinson's | Parkinson's -> associates -> ?). 30
- 30. Example • Find all/some genes that associates with Parkinson's | Parkinson's -> associates -> ?. • What is the likelihood of an associates edge between HP and Parkinson’s Disease? 31
- 31. Biomedical knowledge graphs 32 Entity Embedding matrix Relation Embedding matrix KGEM Embeddings are learned for each entity (gene, disease, compound, etc) and relation in the graph. The embedding dimension typically falls with the range of 32-1024. Himmelstein, D. S., & Baranzini, S. E. (2015). Heterogeneous Network Edge Prediction: A Data Integration Approach to Prioritize Disease- Associated Genes. PLoS computational biology, 11(7), e1004259. https://doi.org/10.1371/journal.pcbi.1004259
- 32. 33 Summary: KG use for target prediction The graph: constituent data + composition ML model Ranking targets Human review Experimental validation
- 33. 34 Summary: KG use for target prediction The graph: constituent data + composition ML model Ranking targets Human review Experimental validation
- 34. 35 The graph: constituent data + composition ML model Ranking targets Human review Experimental validation Summary: KG use for target prediction
- 35. What are our learnings? Lesson 1: First and foremost, data modelling is absolutely critical for your success in anything you might want to do with the graph 36
- 36. What are our learnings? Lesson 1: First and foremost, data modelling is absolutely critical for your success in anything you might want to do with the graph • We have noticed that the graph topology, that is the "structure”, has direct and large effects on predictive performance* • Predictive performance depends heavily on the choice of ML approach, and no method appear to be indisputably the best. Models have their shortcomings, and that ties in together with the graph topology and the underlying data model. 37 * Stephen Bonner, Ufuk Kirik, Ola Engkvist, Jian Tang, Ian P Barrett, Implications of topological imbalance for representation learning on biomedical knowledge graphs, Briefings in Bioinformatics, 2022;, bbac279, https://doi.org/10.1093/bib/bbac279
- 37. What are our learnings? Lesson 1: First and foremost, data modelling is absolutely critical for your success in anything you might want to do with the graph • Building a good data model is not easy: there are difficult decisions to be made, and you have to live with the consequences of these decisions. Ø Disease vs Phenotype vs Symptom Ø Gene vs Transcript vs Protein 38
- 38. What are our learnings? Lesson 2: Cross-disciplinary science can only be done in cross-disciplinary teams! 39 http://drewconway.com/zia/2013/3/26/the-data-science-venn-diagram
- 39. What are our learnings? Lesson 2: Cross-disciplinary science can only be done in cross-disciplinary teams! • Get your SMEs involved in the data modelling, in feature engineering, and in prediction (if you can). We have cross-functional teams with many different expertise typically working together on a daily basis. 40 http://drewconway.com/zia/2013/3/26/the-data-science-venn-diagram
- 40. What are our learnings? Lesson 2: Cross-disciplinary science can only be done in cross-disciplinary teams! • Get your SMEs involved in the data modelling, in feature engineering, and in prediction (if you can). We have cross-functional teams with many different expertise typically working together on a daily basis. • Corollary: Effective communication is key! Learn to speak each others language 41 http://drewconway.com/zia/2013/3/26/the-data-science-venn-diagram
- 41. What are our learnings? Lesson 3: Don't rely on a single analytical/predictive method, experiment with different models (that utilize different portions of the graph, or at least weight them differentially). 42
- 42. What are our learnings? Lesson 3: Don't rely on a single analytical/predictive method, experiment with different models (that utilize different portions of the graph, or at least weight them differentially). 43 • Continually evaluate and sanity-check your results. After all, a computational model is a fancy calculator, no matter how advanced.
- 43. What are our learnings? Lesson 3: Don't rely on a single analytical/predictive method, experiment with different models (that utilize different portions of the graph, or at least weight them differentially). 44 • Continually evaluate and sanity-check your results. After all, a computational model is a fancy calculator, no matter how advanced. • Also worth considering what the cost of a bad prediction is for your business: • for Facebook or LinkedIn it might not be the end of the world to suggest someone you don't know as a contact • that is not the world we live in, and may not be the world you live in either.
- 44. So what? I. Intuitive representation of massive amounts of connected data: Better exploration! II. Use years of wisdom coming from graph-theory as well as the recent advances in graph-based learning for better predictions (remember: the cost of failure/mistake) III. There are great tools provided by the community; the collective expertise is incredibly valuable, use it well! 45
- 45. Thank you! 46
- 46. Confidentiality Notice This file is private and may contain confidential and proprietary information. If you have received this file in error, please notify us and remove it from your system and note that you must not copy, distribute or take any action in reliance on it. Any unauthorized use or disclosure of the contents of this file is not permitted and may be unlawful. AstraZeneca PLC, 1 Francis Crick Avenue, Cambridge Biomedical Campus, Cambridge, CB2 0AA, UK, T: +44(0)203 749 5000, www.astrazeneca.com 47