1. Drug Discovery is a Numbers Game
A Million to One:
Drug Discovery is a Numbers Game
Dr Graeme Robb
23-25th March 2015
JMP Discovery Summit Europe
Brussels

2. Contents What is drug discovery ? Beginning the journey
A bit of context to start
Beginning the journey
Hit Generation
Changing Tracks
Lead Identification
Reaching the destination
Lead Optimisation

3. What is Drug Discovery ? (an unabashedly chemistry-centric view)

4. An Introduction to Drug Discovery
Target protein function altered by drug
Iressa (drug) inhibits EGFR (protein)
Kills tumour cells
Graeme Robb IMED Biotech Unit

5. An Introduction to Drug Discovery
A million to one
Evidence and expertise used to select protein
Hit Generation
~106 molecules
e.g. high throughput screen. Selection of best hits
Is our screening set representative of wider chemistry?
Lead Identification
~104 molecules
Large compound libraries to explore the ‘chemical space’
How do we explore this chemical space efficiently ?
~103 molecules
Lead Optimisation
1 molecule
Narrow down to 1 molecule. Multi-objective.
How do you optimise when your inputs are non-linear
Target Selection
Questions:
Graeme Robb IMED Biotech Unit

6. An introduction to Drug Discovery
Massively Multi-objective Optimisation
Goal is to identify a molecule which...
Binds selectively to the target protein
Dissolves in the stomach/gut (oral drugs)
Permeable at gut-wall and cell-wall
Stable to metabolism
Safe and non-toxic
Synthesis is possible (at scale)
etc, etc...
But all we can alter is the chemical structure of the molecule
Determines: shape, flexibility, electrostatics, bond-strengths
All measureable properties flow from these
Challenging: industry averages show that it takes years to take a drug to the market , at a cost of $5 billion per successful drug.
A drug must find the optimal balance between many factors
(source: Forbes, 2014)
Graeme Robb IMED Biotech Unit

7. Statistical modelling of chemistry
Elephant in the Room
Statistical modelling of chemistry
Reality
Modelling
Various properties. Some measurable or readily predicted. Many unknown
Chemical structure
Measurable effect
Subset of properties, some measured, some predicted, some just wrong
Chemical structure
Measurable effect
Graeme Robb IMED Biotech Unit

8. Beginning the Journey Hit Generation – a BIG challenge

9. The Scale of the Problem
Hit Generation
The Scale of the Problem
Molecules are frameworks of atoms
Each vertice can be one of nine atoms (typically): C, N, O, F, P, S, Cl, & Br
Each edge can be a single, double or triple bond,
BUT, there are rules that define what can exist in nature
Enumerating all possibilities (JL Reymond, Acc.Chem.Res., 2015)
Using up to 13 atoms = 977 million real molecules ~ 10^9
Using up to 17 atoms = 166 billion real molecules ~ 10^11
Drugs typically have up to 40 atoms (many have more)
Estimated to be 10^60 possible drug-like molecules !!!!
How do we even hope to find hit molecules?
isopropanol
acetone
impossible!
Unlabelled vertices are carbon. Hydrogens are omitted.
Graeme Robb IMED Biotech Unit

10. Time for a demonstration
Hit Generation
How do we do it?
Traditionally use High Throughput Screening (HTS)
We can test ~ 10^6 compounds (individually)
Can we do better? Yes. DNA-encoded libraries
We can test ~ 10^9 compounds (in cocktails)
So, best-case is only a tiny fraction of what is possible?
Time for a demonstration
Million-to-one.jmp
Graeme Robb IMED Biotech Unit

11. Hit Generation How do we do it?
Traditionally use High Throughput Screening (HTS)
We can test ~ 10^6 compounds (individually)
Can we do better? Yes. DNA-encoded libraries
We can test ~ 10^9 compounds (in cocktails)
So, best-case is only a tiny fraction of what is possible?
Luckily, each compound is not an isolated point
“Similar chemical structures give similar biological activity”
But how do we define ‘similar’
By one common similarity measure, only 30% of ‘similar’ compounds to a known active compound are themselves active. Lower than you might expect, but significantly better than random (YC Martin, J.Med.Chem., 2002, 45, 4350)
Graeme Robb IMED Biotech Unit

12. How do we define similarity
Screening Set
How do we define similarity
By chemical similarity
Pairwise comparison of molecule features gives similarity score (0 to 1)
Essentially defines a ‘chemical space’ with thousands of dimensions
By pharmacophore
A pharmacophore defines the important interactions with a protein in 3D
By properties
Define compounds with properties, e.g. length, weight, polarity, etc.
By pharmacology
Similar pharmacology on one target may be similar against another
Nature has a habit of reusing successful designs
In reality we must consider all of the above
We do not know what is important in advance.
Graeme Robb IMED Biotech Unit

13. Covering chemical space
Screening set
Covering chemical space
We try to ensure even coverage across chemical space
Here, we would detect 9 actives
Reality is, we have patchy coverage of chemical space
Here, we would find no actives •
We frequently use our various similarity measures to buy/make compounds to plug ‘holes’ like this and increase our chances of finding hits in the future.
Graeme Robb IMED Biotech Unit

14. Time for a demonstration
Finding True Hits
Making sense of an HTS
potential hits
HTS results are typically only N=1 and very noisy
We use stats to define true active
e.g. Active if : Inhib > Mean + 3 x StdDev
Use chemical similarity clustering (outside JMP)
Examine clusters for interest (inside JMP)
Time for a demonstration
Mock-HTS-Workup.jmp
Graeme Robb IMED Biotech Unit

15. Changing Tracks Lead Identification – from initial hit to promising lead

16. Diversity Driven Exploration
Lead Identification
Diversity Driven Exploration
Goal is to find the best starting point for optimisation
Initially we know nothing, so we explore around each hit
We make/buy similar compounds to the hit compound(s)
Diversity is the key driver
We make ‘libraries’ of compounds
Typically based on a simple synthetic route
Try to ensure coverage of ‘chemical space’, i.e. features at 5 R-groups
How can we tell if this gives diversity?
Time for a demonstration
Systematic-design.jmp
Graeme Robb IMED Biotech Unit

17. Are compounds diverse (enough) ?
Lead Identification
Are compounds diverse (enough) ?
2D can be pretty removed from the 3D reality
Look at 3D pharmacophore
Time for a demonstration
Systematic-design-3D.jmp
Graeme Robb IMED Biotech Unit

18. Not only chemical similarity
Scaffold Hopping
Another method of exploring similarity is to use pharmacophore similarity
Ignore chemical similarity
Assuming the pharmacophore is relevant, will maintain biological activity
But it will be chemically dissimilar and so
will change many other properties.
How do we compare scaffolds?
Time for a demonstration
MatchedPairsAnalysis.jmp
Graeme Robb IMED Biotech Unit

19. Application of DOE to chemistry design
Diversity by Design
Application of DOE to chemistry design
Make a lot of compounds and assume/hope it is diverse
A long tradition of pharmaceutical deign - Surely we can do better
Quantify our chemistry with properties
Use DOE to define an optimal set of compounds with prescribed properties
Problem 1: we can only use a subset of properties
Problem 2: properties are continuous, chemistry is discrete
Example:
Available chemistry would allow us to make 100 different X’s
Available resource allows us to make only about 10
Time for a demonstration
DoE_for_chemical_space.jmp
Graeme Robb IMED Biotech Unit

20. Diversity by Design Failure and Success
See, I told you
Ten compounds were duly made and tested
No model existed with these inputs
Consider 3D once more
Different groups have a twisting effect on the molecule (dihedral angle)
Graeme Robb IMED Biotech Unit

21. (inc. negative controls)
Diversity by Design
Failure and Success
Ten compounds were duly made and tested
No model existed with these inputs
Consider 3D once more
Different groups have a twisting effect on the molecule (dihedral angle)
Make another 17
(inc. negative controls)
The model works!
Graeme Robb IMED Biotech Unit

22. Reaching the Destination Lead Optimisation - the best combination of properties

23. Modelling and Prediction
Using Chemical Structure
Can build Quantatative Structure-Activity Relationship (QSAR) models
Predictive models based on calculated (or easily measures) properties
These often fail - not a true representation of chemical structure
Free-Wilson Model (SM Free & JW Wilson, J.Med.Chem., 1964, 7, 395)
Use substructural features directly in the model
Molecule properties are the sum of the contributions of the parts
Represent each chemical substructure with SMILES language
Can model existing data
Predict best combinations of groups for multiple properties
CCO CC(C)O Cc1ccccc COCCBr
Time for a demonstration
FreeWilson-example.jmp
Graeme Robb IMED Biotech Unit

24. Modelling and Prediction
Dose-to-Patient
One of the most important parameters
A composite of many inputs
Defined (to an approximation) with known formula
Can look at how sensitive it is to the inputs
Profiler tool
Time for a demonstration
Graeme Robb IMED Biotech Unit

25. Summary
We can see that while statistically hit-finding should be impossible, invoking similarity arguments allows fuller exploration of chemical space than you might expect.
JMP, being statistically rigorous and very flexible is an ideal tool for exploring and examining HTS result sets.
Diversity in ‘chemical space’ is a subjective property and difficult to quantify. JMP is a useful tool to help visualise this.
Superior diversity/coverage is achieved by including this at the point of design. DOE can be applied to the problem, with care.
JMP’s modelling platforms are ideal for multi-objective optimisation problems we reach in the final stage of drug selection.
Graeme Robb IMED Biotech Unit

26. What Science Can Do
Graeme Robb IMED Biotech Unit

27. Graeme Robb IMED Biotech Unit