computational chemistry in drug discovery - embl … chemistry in drug discovery darren green ebi...

Computational Chemistry in Drug DiscoveryDarren GreenEBI 2016

The Discovery Process

gene proteintarget

screen and identify lead

Leadoptimisation

chemicaldiversity

(compoundlibrary)

test safety& efficacyin animals and

humans

Targets Hits Leads Candidates Drugs Products

Drug discovery and computational chemistry

TargetSelection

CandidateSelection

ProductDifferentiation

MarketEntry

Target ID LeadDiscovery

LeadOptimisation

PreclinicalDevelopment POC Full

Development Registration

Predictive ADMET

Structural Genomics

Sequence homology

Molecular Diversity

Library DesignMolecular Similarity

HTS analysis

(Q)SAR analysis

Attrition analysis/risk assessment

Structure Based DesignLigand based Design

Solid state/multi scale modelling

Industry success rates after candidate selection

26

CS-FTIH

CS-Phase II

62

26

8

6

CS-Phase II

CS-Phase III

CS-Launch

% of compounds achieving milestone

KMR Pharmaceutical Benchmarking Forum 2010

Computation in Drug Discovery: the good news!

–Only Two questions–What (protein/pathway) to target?–And what molecule to make?

–And One topic for today–Molecule Design

Types of Molecules studied by GSK comp chem

– Small molecules

– Peptides

– Protein therapeutics

– Bioconjugates

– Materials

– Enzymes

The Lead Optimisation cycle

Design

Make

Test

AnalyseSAR

“Lead ”

“Candidate”

“Screening Cascade”in vitro

in vivo

Binding Selectivity

Function

Safety Hazard

PK/PD Disease Decreasing throughput

“Candidate”

Surely Computation can help? “Rational” drug design

– Most design methodologies are aimed at reducing the number of cycles in lead optimisation- ideally to 1!

8

– All design methodologies, to date, have had limited success in this regard

Safety

XDrug

Potency

X

Lead

A multi-objective optimisation

9

Solubility

Absorption

Metabolicstability

XDrug

PC1

PC

2

Traditional Way : Sequential Process, Costly, Lengthy

Safety

XDrug

Potency

X

Lead

A multi-objective optimisation

10

Solubility

Absorption

Metabolicstability

XDrug

PC1

PC

2

Traditional Way : Sequential Process, Costly, LengthyDesired- faster navigation through multi-dimensional space, by reducing the cycles

SafetyPotency

Can computation help? Yes

Structure based drug discovery

11

Solubility

Absorption

Metabolicstability

PC1

PC

2

Solvation

Structure-Based Optimization of Naphthyridones into Potent ATAD2 Bromodomain Inhibitors

“Structure-Based Optimization of Naphthyridones into Potent ATAD2 Bromodomain Inhibitors”Bamborough, Chung, Furze et alJ. Med. Chem (2015) 58, p6151-6178

Computational analysis suggests ether linkage optim alfor accessing RVF shelf

Introduction of polarity to gain selectivity via di fferencesin shelf region

Adding back optimised C5 substituent

Design approach for ATP site fragment set

– Fragment set for screening against kinases

– Target mainly the purine hinge-binding site, which is small and flat

� Aromatic ring and 1-3 hydrogen bonds

� 3D pharmacophore

� 2D representations also suitable� 2D representations also suitable

Design approach for ATP site fragment set

– Fragment set for screening against kinases

– Target mainly the purine hinge-binding site, which is small and flat

3D pharmacophore

Available solids Reactivity filters

Kinase SMARTS

Cluster & visualise

Final set:936 fragments

with goodproperty profile

LCMS / NMR QC

Kinase fragment activity profiles

– Kinase fragments have selectivity profiles, just like larger molecules

– This may be unexpected, since they bind in the most conserved part of the domain

%I

30 kinases

N

N

NH

N

NH2 NH

N

N NNH

NHN N NH2

O

0 1 0 0 1 0 0 1 1 0 0 01 1 0 0 1 0 0 0 0 0 0 1

0

30

60

90

Kinase activity profile similarity e.g. Tanimoto coefficient

Kinase profile similarity tracks structural similar ity

– 2D descriptors

– 3D similarity (e.g. Openeye ROCS/ EON, Cresset FieldScreen…)

– 3D similarity metrics capture some of the features of fragments important for selectivity

As the 3D fields of fragments become more similar, so their kinase inhibition profiles* also

tend to become more similar

*Tanimoto activity profile similarity

� A good choice for fragment similarity or diversity“Selectivity of Kinase Inhibitor Fragments”Bamborough, Brown, Christopher, Chung, Mellor

J. Med. Chem (2011) 54, p5131-5143

Theoretical considerations when working with Protei n Structures

– What you see on a screen is not real– Proteins move– Hydrogens are often assigned not observed

– How is solvation being treated?– especially when looking to make h-bonds

– Am I taking entropy into account?– Am I taking entropy into account?– How are vdW energies being calculated?

– remember r12!

– Ligand conformational energies are as important as protein-ligand interaction energies

– Ligand based modelling methods are very effective when you also have a protein structure

The facts…..

GW813893XN O

N

N

O

S

O

O

P1

S ClX

S ClX

Ki(FXa) = 5 nMKi(Thr) = 184 nM

Ki(FXa) = 400 nMK (Thr) = 31 nM

M1N O

N

N

O

S

O

O

P1

S ClX

S ClX

Ki(Thrombin) = 367 nM

Ki(Thrombin) = 17 nM

Example: FXa/Thrombin inhibitors

GW833057XO

i

Ki(Thr) = 31 nM

M2O

Do they have a different binding mode in thrombin???

NoNo

0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

0 40 80 120 160 200 240 280 320 360

dihedral angle (C=C-S=O)

rela

tive

ener

gy in

kca

l/mol

N S

O

OCH3

H

N S

O

OS ClR

H

O=S C-C

M2

∆Econf(torsion1)3.1 0.0

N S

O

OS ClR

H

O=S C=C

M1

0.00.51.01.52.0

2.53.03.54.04.5

0 40 80 120 160 200 240 280 320 360

dihedral angle C-C-S=O

rela

tive

ener

gy in

kca

l/mol

N S

O

OCH3

H

∆Econf(torsion1)

in kcal/mol3.1 0.0

367 nM 17 nMKi(Thrombin)

Relaxed scan of the dihedral angle [C-C-S=O] (in 10º increments) The calculations were performed with Gaussian98 at the B3LYP/6-31G* level of theory. The orange circle is placed at the dihedral angle found when the ligands are bound to thrombin.

Virtual Screening

– The in silico equivalent of “wet” assays/HTS

– Requires a method of predicting active compounds

– protein structure (+docking)

– 3D pharmacophore

– QSAR equation

– Capable of screening many more molecules than can be made or tested in reality

– This “high throughput” use of comp. chem. requires us to make the most approximations

OO

NHN

P D/A Ar

A

D

Reduced Graphs

1 0 1 1 0 1 0 1… …

ReferenceShapes

Shape fingerprintMolecular fields

Chemical Descriptors

Cl

NHN

Cl

NH2

OO

NHN

N

O

N

O

NH2

Cl

Cl

N

OO

NN

Atom pairs and paths

Fragments

Shape3D Pharmacophores

1 0 1 1 0 1 0 0… …1 0 1 1 0 1 0 0… …

1 0 1 1 0 1 0 0… …

“3-point”Pharmacophores

“2D” fingerprints“3D” fingerprints

Chemical Descriptors

– Which ones to use?

– Which are “best”?

– GSK

– 2D fingerprints for diversity

– Reduced graphs, 3D pharmacophores. 3D fields and shape for knowledge based work

– The most important ingredient for success is the quality of the computational chemist, not the software (or descriptors) they have to hand (assuming a base level of scientific validation!)

Pharmacophores

� A reductionist approach to bioactivity

� “A pharmacophore is the ensemble of steric and electronic features that is necessary to ensure the optimal supramolecularinteractions with a specific biological target structure and to trigger (or to block) its biological response ” IUPAC

CO H

Hdon5.8

5.0

� Neglect of entropy, solvation and anything subtle!

CO2H

N+

Hdon

Hacc

5.0

2.97.6

5.4

6.5

9.3 angs

Example: 3D pharmacophore lead hopping 2. Endothelin antagonists

NHHN

OO

NH

O

O

NHO

HO2C

HNO

HO

OH

H

O

HO2C

+

5 BQ123 (IC50 22nM) 6 Shionogi 50-235 (IC50 78nM)

O

O

CO2HO

O

CO2H

CO2H

9.1 - 11.3 Angstrom

8 (IC50 9 M)7 (IC50 730 nM)

Molecular Field methodology- Cresset

Use of a special forcefield to better simulate elec trostatics

IsoStar- Cambridge Crystallographic Data Centre

Field searching example-Oxytocin antagonists

219 compounds

“2D Pharmacophores”

– Use pharmacophore concept, but not 3D distances or arrangements

– Encode pairs of pharmacophore points separated by a certain number of bonds (typically from1 to ~10)

– These are used to produce a vector, or fingerprint, which can then be used to compute molecular similarity (cf in 2D chemical then be used to compute molecular similarity (cf in 2D chemical database systems)

– Advantage– Don’t have to worry about active conformations

– Disadvantage– May be less selective that 3D methods (i.e. may need to screen

more molecules)

Reduced Graphs

– A powerful “2D” pharmacophore description– CB1 antagonist example

N

NH

O

NDonor / Acceptor

Aliphatic Ring

Acceptor

NN

Cl

Cl

Cl

Aromaticring

Aromaticring

Aromatic Ring

Acceptor

NN

Cl

NH

O

Cl

N

Cl

CB1 Antagonists

NN

Cl

NH

O

NC

MeO

N

N NH

O

N

Cl

Cl

N

NNH

O

N

MeO

Cl

Pfizer

OO

O

O

N

N

S

Cl

Cl

F

FO

O

N

N

Cl

NH

O

Me

Cl

Cl

NN

Cl

N

O

Cl

Cl

O

N

N NH

O

Cl

Cl

OH

NN

Cl

O

Cl

Cl

FF N N NH

N

Cl

SN

OH

OO

Eli Lilly Aventis Merck U. ConneticutBayer Va. Commonwealth Solvay

Solvay Astra

2D FP: 0.468

RG: 1

2D FP: 0.420

RG: 1

2D FP: 0.468

RG: 0.815

Reaching Beyond the Fog in HTS

– Data-driven analysis finds motifs, Reduced Graphs, frameworks etc that are enriched in the active population

Physical Properties

Solubility forecast index (SFI)Hill, Young, Drug Discovery Today 15, 648-655 (2010)

601 1109 2707 5847 8647 11436 13421 12999 9711 6024 3058 1229 620

60%

70%

80%

90%

100%

Proposed:“Solubility Forecast Index”

Simple metrics can be effective and highly competit ive with first principles computation

Measured CLND solubility < 30 µM; 30-200µM >200µM

Binned SI

x = 1 1 < x ... 2 < x ... 3 < x ... 4 < x ... 5 < x ... 6 < x ... 7 < x ... 8 < x ... 9 < x ... 10 < x... 11 < x... 12 < x0%

10%

20%

30%

40%

50%

SFI = clog D7.4 + #Ar

or

SFI = Chrom log D7.4 + #Ar

SafetyPotency

Can computation help? Sometimes

Toxicology predictionOff target effects

38

Solubility

Absorption

Metabolicstability

PC1

Prediction of permeationActive transport

Metabolite prediction

Physical Properties and flatness together

PFI and probability of success

Essential Reading

Poor solubility for Mpt 150° C(Yalkowsky)

Kinetic solubility >250 mg/ml

Low solubility/ionisation dependent

Human Clearnance >7.8 mL/min/kg

Human Clearnance >7.8 mL/min/kg

Bioprint receptor promiscuity Bioprint receptor promiscuity (AZ)

Human mice Clint >8.6

Poor permeability (PSA < 132)


permeability < 100 nm/s (MWt <400)permeability < 100 nm/s (MWt <400)

Renal clearance



Adapted from: (AZ)(AZ)

Bioprint receptor promiscuity (Roche)

Bioprint receptor promiscuity (Roche)

in vitro rat toxicology

Phospholipidosis(bases of pKa 9)

Phospholipidosis (all bases)

hERG >30% @ 10µM (neutral molecules)

hERG >30% @ 10µM (bases)

logD/P-3 -2 -1 0 1 2 3 4 5 6

Adapted from:“Lipophilicity in Drug Discovery”M. J. WaringExpert Opin. Drug Disc., 5, 235-248 (2010)

– Common problem in medicinal chemistry..

CF3

Key Issue : High intrinsic clearance

BioDig – Why ?

O

N

S OO

Hets

from Phenyl

‘Why don’t you change it for an oxadiazole – it worked for us in Prog X’.

Can we systematically mine existing data to find structural changes that lead to better clearance ?

“Standing on the shoulders of giants” Isaac Newton

– How ?

– How do we analyse our SAR ?– One way is to look for pairs of compounds which only differ by a

single change

Mining our data

single change

The change in activity can then be attributed to the structural change

– A database has been created where we have collected together the compound pairs (and associated structural changes) in our ADME data

BioDig ADME database

Property Number of compounds

Number of matched

molecular pairs

Number oftransforms

Clearance 54k 9M 7M

Solubility 267k 243M 217M

Represents a huge amount of knowledge to tap into

Solubility 267k 243M 217M

Log D 328k 316M 279M

hERG* 175k 98M 76M

PP Binding 248k 137M 117M

P450 3A4 262k 193M 167M

PGPefflux 6k 367k 306k

*old assay format

BioDig ADME

Substructure X: “What possible replacements have been tried and which

improve clearance ?”

The information in the database can be used to answer other questions

Binned ACTIVITY_CHANGE

1

2

3

5

4

17

21

7

2

-2.2 -1.8 -1.4 -1 -0.6 -0.2 0.2 0.6 1 1.4 1.8 2.20

5

10

15

20

PROFILE_CLUSTER - 22









- - - 0 + + + - - - 0 + + + - - - 0 + + +

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

Surely Computation can Help? What have we learnt?

ThreatsSmall, aging, communityPharma Industry uncertaintyHype (again)

StrengthsSmall, well-connected communityOur PeopleDataMoore’s law/GPUs

WeaknessesSlow progress in fundamental sciencePoor alignment of Academia and Industry

Opportunities“Principles of Computational Drug Design”Poor alignment of Academia and Industry

Performance Plateau of our techniques“Principles of Computational Drug Design”PPPs

Computer-aided molecular design under the SWOTlightDarren V. S. Green, Andrew R. Leach, Martha S. HeadJ Comput Aided Mol Des (2012) 26:51–56

Can we predict this? A molecule than binds to the protein mostly through the water network!

Sampl1, Docking Mode Prediction

Eric-Dock andMarti-Dock!

Surely Computation can Help? What have we learnt?

ThreatsSmall, aging, communityPharma Industry uncertaintyHype (again)

StrengthsSmall, well-connected communityOur PeopleDataMoore’s law/GPUs

WeaknessesSlow progress in fundamental sciencePoor alignment of Academia and Industry

Opportunities“Principles of Computational Drug Design”Poor alignment of Academia and Industry

Performance Plateau of our techniques“Principles of Computational Drug Design”PPPs


Compound Quality:Are you looking for a “probe”, a “tool”, a drug “pu blication”…. or a “publication”….

537 compounds• 115 pass GSK filters• only 71 with no issues!

Many promiscuous/interference chemotypes published time and time again :http://blogs.sciencemag.org/pipeline/archives/2012/06/01/return_of_the_rhodanome

None of these papers move the science of computatio nal chemistry forward

The world view of an industrialcomputational chemistry group

....despite the real limitations in our underlying methodology, computational chemistry still makes a significant impact on drug discovery. Much of the reason can be ascribed to the fact that very rarely are computational chemistry methods used without any intervention and guidance from a scientistintervention and guidance from a scientistoften with significant knowledge of their target and often many years of experience in the application of different computational chemistry methods


“Principles of Computational Drug Design”

SafetyPotency

Will computation improve? Yes

Structure based drug discoverySimulations, QM etc

54

Solubility

Absorption

Metabolicstability

PC1

PC

2

Crystal form predictionSolvation energies

SafetyPotency

Will computation help? Maybe

Toxicology predictionOff target effects

55

Solubility

Absorption

Metabolicstability

PC1

Prediction of permeationActive transport

Metabolite prediction

Summary

– Computational chemistry is a key component of modern drug discovery– With great reliance on experience & know how of individual computational

chemists

– Computation is beginning to find application for a wide variety of – Computation is beginning to find application for a wide variety of problems, not just small molecule discovery

– We have a long way to go before we can replicate what is done in other industries– But it is a journey we must begin

Acknowledgements

– Paul Bamborough

– Chris Luscombe

– Stephen Pickett

– Jameed Hussain

– Ceara Rea

– Rob Young– Rob Young

– Alan Hill

– Eric Manas

– Marti Head

Thank you for the invitation!

computational chemistry in drug discovery - embl … chemistry in drug discovery darren green ebi...

Documents