Chandigarh Call Girls Service ❤️🍑 9809698092 👄🫦Independent Escort Service Cha...
Filling the gaps in translational research
1. Filling the gaps in translational research
Paul Agapow, Health Informatics Director
D4, Basel
Public
October 2019
2. Disclosure
• No conflicts of interest
• Based on experience in current &
previous positions
• Health Informatics @AZ, RWD
• Data Science Institute @ICL,
precision medicine
• Does not reflect official AZ thought
or projects
2
3. Stated thesis
• Development of new therapies is largely a
matter of translating basic research into
actionable healthcare
• Too often such research focuses on the wrong
problems and approaches
• What needs to be done to close the gap?
3
5. A revolution in drug development?
• Every day we hear of new
advances & developments
• Acceleration in basic
biomedical research
• Constant development of
new molecular technologies
• An age of cheap
computation & powerful
machine learning
5
6. Drug development is increasingly unsustainable
• Eroom’s Law: cost of
developing new drug roughly
doubles every nine years
• Accelerated biomedical
research not reflected in
drug development
• (Not discussing regulation)
6
Pharmacelera (2014)
8. There is a conflict in priorities
“Academic”
• Interesting problems
• Ideal, clean data
• Isolated, simple biology
• “Proof of concept”
• Focus on early dev
• Often single instances &
statistics
8
“Industrial”
• Needful problems
• Real, messy data
• Real, systemic biology
• Operational
• Need help in late dev
• Usually a numbers game
Thesis: we tend to employ & incentivize approaches that
inhibit long-term therapy development
9. • Tendency to solve:
• Easy problems (low hanging fruit)
• Interesting problems
• Publishable problems
• Problems we have data for
• But:
• That’s not where the problems
are
• That’s not where the savings are
9
Translational research tends to solve the wrong problems
10. • Landmark AZ papers:
• Cook et al. 2014
• Morgan et al. 2018
• 5 Rs:
• right target
• right patient
• right tissue
• right safety
• right commercial potential
• Translational research focuses on
early Rs at expense of later Rs
10
We neglect the 5 Rs of drug development
11. • It costs ~ $1B and 10 years to
develop & launch a drug
• Each patient in a clinical trial costs
$1-10K
• The “valley of death”: most
candidate drugs will fail
• The later it fails, the more
expensive
11
We neglect the tough maths of drug development
13. • There is a tendency to treat drug
development as just a data
problem
• Machine learning
• Shift from whole-organism to
high throughput methods
• Simplified view of biology
• As we move later in the
development cycle, biology
grows more complex & more
important
13
Problem: Biology isn’t “just the domain”, it’s the problem
14. • Maybe all the low-hanging fruit has
been picked
• E.g. single gene / single system
diseases
• Most diseases are complex &
systemic
• Many patients are complex
• Lifestyle, exposure, co-
morbidities, co-medications
• A cohort is rarely just a simple table
14
Problem: Simple biology only helps with simple patients
15. • Work in real complex biology early
• More work in phase 2 saves time
and money in phase 3
• Work hand-in-hand with chemists,
epidemiologists, toxicologists
• Validate functionally early & often
15
Action: Try to fail early & fail often
Ferrero (2017) ODSC Europe
16. • Incorporate as much biological
information as possible as early as
possible in the discovery process
• Integrative analysis
• Constrain search with biology
• Accept complexity
• Polypills
• Polypharmacology
• Look for hostile data
• Adverse effects
16
Action: Recognise & work with complex biology early
Krassowski (unpub.)
17. ML is hungry for data but:
• Not enough labelled data
• Not enough of the right sort of data:
• e.g. adverse events
• Badly imbalanced data (+ve or -ve):
• e.g. “what is the effect of drug X
on cancer Y”
• Not enough data without weird
biases
• e.g. hospital data
17
Problem: We need more data
18. • If you use sparsely sampled data to
explore or describe a data space,
the apparent shape of that data
space has more to do with the data
sampling than the actual space.
• And just about everything is sparse
• We need better datasets
Problem: Train or learn from sparse data at your peril
18
19. • Solutions are inevitably shaped by
the composition of the test cohort,
which are usually:
• WEIRD
• Young male Caucasians
• Worried well
• But:
• Can models based on these
populations generalise?
• Do drugs behave differently in
different populations?
19
Problem: It is too easy to study the wrong populations
20. • Need data with relevant / real
populations
• Not just more validity, there is more
information in a diverse dataset
• Lowers barrier to exploration
• How do we do this?
• Harvest EHRs & other RWE
• Collaborate with national centres
• Build small, locally dense datasets
• Will require long-term funding &
broad collaboration to ensure
usefulness & sustainability (FAIR,
consortiums, public-private, IMI ..?)
20
Action: Build more diverse datasets
21. • On top of dataset bias, our data is too simple
• Therapies and algorithm always under-perform in the “real world” because:
• Disease is complex
• Patients are complex
• Co-morbidities
• RCT populations are unrealistic
• Desk drawer problem
21
Problem: if it’s not Real World Data, it’s not real
22. • Synthetic control arms
• Build tools over RWD to help clinical trials
recruitment
• Build new algorithms to exploit RWD
• Disease trajectories / Brunak
22
Action: use RWD, build algorithms for RWD
Hypertension Diabetes Retinal Dx
Acute
bronchitis Candidiasis
Menstruation
disorder
23. When:
• We have no good idea of the
“model” underlying the data
• Variables may interact in complex
ways
• There’s potentially a lot of variables
and we don’t know which ones are
important
• Lowers barrier to exploration
23
Machine learning is highly attractive for drug dev
24. Despite promise, we have little more than
proofs of concept:
• Chen et al. (2019) showed DUD-E dataset,
used by many “accurate” CNN models of
drug-target interactions, actually biased
• AI-radiomics shows incredible performance
in trials but mediocre performance in the
clinic
• Many ML studies are direly underpowered
• We keep solving the same “easy” problems
• Cultural issues
24
Problem: We are terrible at machine learning
25. • Use ML to prioritize research:
• “I can’t do anything with 1000
candidate molecules in Phase
III. Give me 5 good ones.”
• E.g. Prioritize & validate pro-
arrhythmic candidates
• Use ML later in the pipeline where
savings are greatest:
• E.g. adverse events
• E.g. screening trial candidates
• Don’t put data science in a box
25
Action: Use ML where it has impact
Costabal, et al. (2019) BioXRiv
26. • Take a drug approved for one
indication and use it for another
• Why:
• Cheap
• Starting with a drug that does
something is closer than
starting de novo
• Safer
• Drugs act on common pathways
26
Action: Drug repurposing & repositioning is smart
Naylor & Schonfeld (2014) DDW Winter
27. • Use data and ML where they have impact / saving
• This is often later in the drug development process
• Use data and ML to keep yourself honest
• Let’s build the datasets we don’t have
27
Summary