computational chemistry in drug discovery - embl … chemistry in drug discovery darren green ebi...
TRANSCRIPT
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)
Poor permeability (PSA < 132)
permeability < 100 nm/s (MWt <400)permeability < 100 nm/s (MWt <400)
Renal clearance
Poor permeability (PSA < 132)
Poor permeability (PSA < 132)
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
PROFILE_CLUSTER - 28
PROFILE_CLUSTER - 33
PROFILE_CLUSTER - 23
PROFILE_CLUSTER - 31
PROFILE_CLUSTER - 34
PROFILE_CLUSTER - 27
PROFILE_CLUSTER - 32
PROFILE_CLUSTER - 38
- - - 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
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
Computer-aided molecular design under the SWOTlightDarren V. S. Green, Andrew R. Leach, Martha S. HeadJ Comput Aided Mol Des (2012) 26:51–56
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
Computer-aided molecular design under the SWOTlightDarren V. S. Green, Andrew R. Leach, Martha S. HeadJ Comput Aided Mol Des (2012) 26:51–56
“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!