Structurebased Lead Optimization of Isosteric Analogs

1
Structure-based Lead Optimization of Isosteric
Analogs

• Roger Sayle, Bob Tolbert, Anthony Nicholls and
  Geoff Skillman
• OpenEye Scientific Software,
• Santa Fe, New Mexico

2
Overview

• Motivation
• The WABE/GIMBLE hypothesis
• Synergy and implementation
• Previous experimental systems
• Results of new validation studies
• The autopsy and future work

3
Motivation

• Specialized docking protocols for lead
  optimization rather then lead discovery.
• Methods for virtual library screening are
  inappropriate for chemistry optimization.
• Knowing the active site, binding modes and
  interactions helps significantly.
• Make maximum use of existing X-ray structures of
  protein-ligand complexes.

4
Docking is like a night club...

• its all about posing and scoring.
• What if we eliminate posing completely?
• Restrict ourselves to compounds with
  identical/similar binding modes.
• The frozen ligand approximation.
• Should cancel out entropic effects, VdW clashes,
  solvent disorder, protein flexibility,
  conformational strain...

5
Whats left? Electrostatics

• The remaining electrostatic terms hydrogen
  bonding, desolvation, hydrophobic effect,
  tautomerism and disassociation are better
  understood.
• Benefit of recent improvements in partial
  charging and continuous dielectric
  Poisson-Boltzmann methods.
• Dramatic computational simplification.

6
Electrostatics in water
7
Electrostatics in an active site
8
Gimble isostere evaluation

• Using Tidors formulation of the matrix its
  possible to calculate electrostatic binding
  energies of isosteres orders of magnitude faster
  than full Poisson-Boltzmann calculations.
• Full PB times (zap_bind)
• 1320 mols 54m44s 2.48 s/mol
• Gimble times
• 1320 mols 1m34s ltlt 0.07 s/mol

9
Gimble approximation accuracy
10
Wabe isostere generation

• Isosteres are compounds with the same shape.
• We define it as has the same heavy atom topology
  and non-terminal hybridizaton state (geometry).
• Isosteres typically adopt the same binding mode.
• Given a bound ligand-protein complex crystal
  structure, we can extrapolate isosteres with more
  confidence.
• The wabe algorithm refunctionalizes hybridization
  topologies (scaffolds) using fragments and atom
  types seen in a reference training database.

11
Example WABE solutions
12
Methotrexate/DHFR

• 33,250 analogues (2m41s) later

13
DHFR experimental support
Exptl IC50 3.1µM
Exptl IC50 0.125µM

• Hansch et al., J. Med. Chem., 20, 96-102, 1977.
• Piper et al., J. Med. Chem., 29, 1080-1086, 1986.
• Graffner-Nordberg et al., J. Med. Chem., 43,
  3852-3861, 2000.

14
p56 Lck kinase/AZ M532121
Astra Zeneca in-house structure at 1.5Å resolution
Cleave 2-hydroxyethyl and 1,720,000 analogues
later
15
Lck kinase experimental support
Ureas ? Thioureas lt Guanidines CH3/Cl lt CH3/CH3 ?
Cl/Cl lt F/F
16
Previous conclusions

• MMF94 better than Gasteiger charges.
• Explicit hydrogens important but there placement
  isnt critical.
• Residual problems with radii differences.
• Medicinal chemistry from basic physics.
• Different needs for smooth posing functions
  accurate scoring functions?

17
Round 2 The good, the bad the ugly

• Validation of the frozen ligand
• approximation against multiple systems.
• Ribosome inactivating protein
• L-Arabinose binding protein
• Neuraminidase (sialidase)
• HIV-1 Reverse transcriptase
• Cyclooxygenase 2 (COX-2)
• Glutamate receptor GluR2

18
Ribosome inactivating protein

• D.J. Miller, K. Ravikumar, H. Shen, J.-K. Suh,
  S.M. Kerwin and J.D. Robertus, Structure-based
  Design and Characterization of Novel Platforms
  for Ricin and Shiga Toxin Inhibition, J. Med.
  Chem. 45(1), pp. 90-98, 2002.

9-deazaguanine (1il4)
7-deazaguanine (1il9)
xanthine
19
RIP experimental support
Yellow 9DG pdb1il4 (2.6Å) Blue 7DG pdb1il3
(2.8Å) Ca RMS 0.324Å/267 points
9DG lt 7DG lt XAN
20
L-Arabinose binding protein

• J. Shen and F.A. Quiocho, Calculation of Binding
  Energy Differences for Receptor-Ligand Systems
  using the Poisson- Boltzmann Method, J. Comp.
  Chem. 16(4), pp. 445-448, 1995.

a-D-Galactose (5abp)
a-D-Fuctose (1abf)
a-L-Arabinose (1abe)
21
ABP experimental support
Yellow GLA pdb5abp (1.8Å) Blue FCA pdb1abf
(1.9Å) Ca RMS 0.170Å/305 points Red
ARA pdb1abe (1.7Å) Ca RMS 0.195Å/305 points
ARA lt FCA 1DA lt 6FA lt 2DA
R2 (except GLA) 0.6847
22
ABP post-mortem
Explicit hydrogen location is important.
Depending upon the orientation of the -OH rotor,
the interaction unique to a-D-galactose can
be favorable or not. We were lucky this isnt
an issue with heterocycles. This problem is
easily fixed.
23
Neuraminidase (sialidase)

• M. Fornabaio, P. Cozzini, A. Mozzarelli, D.J.
  Abraham and G.E. Kellogg, Simple, Intuitive
  Calculations of Free Energy of Binding for
  Protein-Ligand Complexes. 2. Computational
  Titration and pH effects in Molecular Models of
  Neuraminidase-Inhibitor Complexes, J. Med. Chem.
  46(21), pp. 4487-4500, 2003.

R1OH, R2OH 2,3-didehydro-2-deoxy- N-acetylneura
minic acid DANA (1f8b) R1OH, R2NH2
4-amino-DANA (1f8c) R1NH2, R2OH 9-amino-DANA
(1f8d) R1NH2, R2NH2 4,9-diamino-DANA (1f8e)
24
Neuraminidase experimental support
4A lt DANA ltlt 49A lt 9A
Yellow DAN pdb1f8b (1.8Å) Blue 4AM pdb1f8c
(1.7Å) Red 9AM pdb1f8d (1.4Å) Green 49A
pdb1f8e (1.4Å)
FDPB R2 0.9548 Gimble R2 0.8916
25
Neuraminidase post-mortem
Although this is a another success, there is a
caveat GLU276 had to be modeled as neutral (as
proposed by Kellogg et al.). This is supported
by the pH 5.5 conditions of the assay. In all
other validation case studies, the protein was
left at default protonation states for pH 7.0.
26
HIV-1 Reverse Transcriptase

• R.C. Rizzo, M. Udier-Blagovic, D.-P. Wang, E.K.
  Watkins, M.B. Kroeger Smith, R.H. Smith, J.
  Tirado-Rives and W.L. Jorgensen, Prediction of
  Activity for Nonnucleoside Inhibitors with HIV-1
  Reverse Transcriptase Based on Monte Carlo
  Simulations, J. Med. Chem., 45(14), pp.
  2970-2987, 2002.

R1 C, R2COCCO, XS HEPT (1rti) R1 C(C)C,
R2COCC, XC MKC-442 (1rt1)
27
HIV RT experimental support
Yellow HEF pdb1rti (3.0Å) Blue MKC pdb1rt1
(2.55Å) Ca RMS 1.054Å/538 points
Partial signal Effects of substituting the beta
and epsilon oxygens predicted, but not X or
cleavage.
28
HIV RT post-mortem
Only a single, but invariant, short-range
electrostatic (h-bonding) interaction. The
sidechain pruning extrapolation performs
less well in sites dominated by VdW interactions,
and surface area/cavity terms. Possible long
range electrostatic interactions and
crystallographic waters.
29
Cyclooxygenase 2 (COX-2)

• J.-C. Mozziconacci, E. Arnoult, P. Bernard, Q.T.
  Do, C. Marot and L. Morin Allory, Optimization
  and Validation of a Docking Scoring Protocol
  Application to Virtual Screening for COX-2
  Inhibitors, J. Med. Chem. 48(4), pp. 1055-1068,
  2005.

SC558 (1cx2)
30
COX2 experimental support
Identifies incorrect sulfonamide rotor, but very
little signal to noise.
31
COX2 post-mortem
Only one (very) low resolution (large) structure,
1cx2 is 3.0Å Sulfonamide rotor is
clearly incorrectly assigned and rings distorted
from planarity. The protein highly
restrained (almost homology model-like) Bromine
is a poor analogue/isostere. No significant
short range electrostatic (hbond) interactions.
32
Glutamate receptor GluR2

• T.B. Stensbol, P. Uhlmann, S. Morel, B.L.
  Eriksen, J. Felding, H. Kromann, M.B. Hermit,
  J.R. Greenwood, H. Brauner-Osborne, U. Madsen, F.
  Junager, P. Krogsgaard-Larsen, M. Begtrup and P.
  Vedso, Novel 1-Hydroxyazole Bioisosteres of
  Glutamic Acid Synthesis, Protolytic Properties
  and Pharmacology, J. Med. Chem. 45(1), pp.
  19-31, 2002.

AMPA (1ftm)
33
GluR2 post-mortem

• Subtle pKa (and tautomer?)
• effect on the assay.
• FDPB unable to differentiate
• neutral vs. ionized forms of
• acidic hydroxyl.
• MMFF94 partial charging
• model may be inappropriate
• (AM1BCC?)
• Potentially long range
• electrostatic interactions.

34
Conclusions

• Alas, the frozen ligand approximation is not a
  magic bullet, just a powerful tool.
• It works exceptionally well in many cases, but
  fares poorly on difficult systems.
• Gimble probably performs at least as well as
  traditional scoring functions (much faster).
• The results of wabe are still insightful even in
  cases where gimbles accuracy falters.

35
Future Work

• Address the known failings of explicit hydrogen
  placement (hydroxyl rotors).
• Improvements of protein pKa tools for active site
  preparation.
• Enhancements to isostere enumeration.
• Interactive graphical user interfaces.
• Automatic clipping and sprouting.

36
Acknowledgements

• Anthony, Bob and Geoff.
• Andrew Grant at AZ.
• OpenEyes VIDA2 team.
• Paul Hawkins at OE.
• Astra Zeneca.
• UCL/EBI Ligplot
• HP, IBM and SGI
• Lewis Carroll