Can drug repurposing be saved with AI 202405.pdf

Presented at DigiTechPharma, London May 2024.
What is drug repurposing. Why is it needed? What systematic approaches are there? Is AI a solution? Why not?

1. Can drug repurposing
be saved with artificial
intelligence?
Paul Agapow (GSK)
London, May 2024
[email protected]

2. Introduction & disclaimer
●Have been a biochemist, immunologist, molecular
evolutionist, evolutionary biologist, bioinformatician,
database manager, analyst, data scientist,
epi-informatician, “computer guy” ...
●Now I crunch numbers in drug development for big
pharma
●Nothing in this presentation implies the existence of any
project or policy at any company
●A side project
Portrait via ChatGPT

3. Is drug repurposing
necessary?

4. The usual litany of woes
●Drug development is too slow,
too long, too expensive & too
uncertain
●Erooms Law (it’s only getting
worse)
●It’s difficult to develop drugs
for rare diseases
Jones & Wilsdon (2018) The Biomedical Bubble

5. What is drug repurposing?
●A variety of definitions
●A.k.a. Repositioning, reprofiling
●Using an existing drug or drug candidate for a new
treatment indication
●Has been in clinical development for one indication, now
in clinical development for a distinct indication
●Excludes off-label use

6. Shortcuts to a successful drug
Mohi-ud-din et al. (2023)
Euro. J. med. Res.

7. Why should it work?
●A working drug has many qualities:
○Can be synthesized, is stable
○Can be tolerated by a patient
○Does something
○May have data associated with it
○May have passed some regulatory hurdles
●Islands in “drug space”
●60% of drugs are studied in 2+ diseases
●Diseases can share targets
●Making off-target effects on target
●Change formulation & delivery for new target
●Backlog of abandoned drugs

8. Successful (?) examples
●1/3 of approved drugs
●Obscured by COVID, drug classes, Traditional Chinese
Medicine
Drug Original indication Repurposed for
Viagra (Sildenafil) Angina Erectile dysfunction
Minoxidil Hypertension Hair loss
Thalidomide Morning sickness Myeloma
Dimethyl fumarate Psoriasis Multiple sclerosis
Rapamycin Immunosuppressant Anti-cancer

9. Can we systematically find
drugs for repurposing?

10. How to find drugs for
repurposing?
●Serendipity
●RWD
●Systematic searching (by starting point)
○Drugs vs Diseases vs Targets
○On target or off-target
●Systematic searching (by search method)
○Network analysis or AI

11. From which direction?
The Drug
●Find other use for a
drug
○ Approved
○ Investigational
or abandoned
○ Withdrawn
○ End-of-life
●Least common
●Look for off-label use
The Disease
●Identify the mechanism
underlying a disease
●Look for analogous or shared
mechanisms with other diseases
●Hence, link drug across diseases
The Target
●Identify the mechanism
underlying a disease
●Look for analogous or shared
mechanisms with other diseases
●Hence, link drug across diseases


https://www.elsevier.com/en-gb/industry/drug-repurposing

12. Every Cure
Testing 3000 drugs against
2000 diseases

13. What works?
●We don’t know
●Thanks to COVID, we now know even less
●Few measures of success
●Most don’t go to trial
○And we don’t know why

14. Can AI help us
systematically find drugs
for repurposing?

15. What works?
●No one knows

16. Networks and drugs
●Complete the triangle
●Broadly two computational approaches:
○Molecular similarity / structural analysis
■E.g. bromovinyldeoxyuridine (herpes -> HSP27 -> anti-tumor)
■“uni-modal”
○Networks, knowledge graphs, pathway analysis ...
■Network analysis
■“AI”
D1 D2
T

17. Why use AI?
●Deploy at scale
●Bring a more systematic approach
●Integrate multimodal and complex data

●Interested here more in identification than screening for
validation
●What can we learn if we survey AI-mediated repurposing

18. What’s a (knowledge) graph?
●A network or graph of entities
○Drug, diseases, molecules ...
●Bound by relationships
○suppresses, expressed_by, targets,
treats, has_phenotype ...
●Integrate multimodal and
complex data
●info about metabolic or
signaling pathways, protein
interaction networks, drug
interactions ...

19. Example 1: CoV-KGE
●Zeng et al. (2020) J. Proteome Res.
●Construct KG from 24M PubMed articles &
DrugBank
●Used deep learning over graph to identify 41
candidate drugs against COVID
●“Validated” by matching against SARS-CoV
transcriptomics & proteomics

20. Example 2: Opioid use disorder
●Feng et al. (2023) Computers in Biology & Medicine
●Using opioid inhibition data and DrugBank to
match “binding fingerprints” via ML
●Also predicted pharmacological properties of
drugs
●None subsequently taken to trial

21. Data sources
●Uses a variety of data sources & types
○Make comparison difficult
●But usually lean on a few common sources
○Mostly Drugbank
○Good but still limited size
○Most of the usual data type suspects
●Desk drawer problem?
●TA problem - infectious disease outweighs all

22. Biological knowledge is a
bottleneck
●If you’re looking for a drug that treats disease X,
what if X doesn’t exist as a single coherent
disease?
○E.g. COVID, various complex & contested diseases
●Biological knowledge is unevenly distributed
○E.g. our neurological or cancer knowledge is incomplete
●You can’t use “all the information”
●You can’t test “all the candidates”

23. Measures of success
●Few
●Poor and varied use of validation
○Some validate in biology
○Some validate in literature
○Some only validate internally
○Some don’t validate at all
●Most don’t go to trial
○And those that do appear to succeed no better than de novo drugs

24. Inevitably, there’s a GenAI angle
●Yan et al. (2024) NJP Digital Medicine
○Asked ChapGPT for repurposing candidates for AD
○Tested candidates in RWE
○Valid but not de novo candidates, more of a literature
search
○And similar to other non-AI studies

25. Practical barriers
●Data access to compound banks
●IP barriers
●Drug development still lengthy
●How do you make money from this?

26. Summary
●We should be good at it
○Repurposing is happening at scale already
○Considerable savings
●Yet there are no established best way(s)
○Unvalidated / untested candidates
○Lack of benchmarks
●Low-tech (screening) use of AI has been successful
●AI over graphs seems promising but still lack evidence of
systemic effectiveness
●Still failing to live up to promise