pdf 3.1 M - OpenEye Scientific Software

Considerations of small molecule strain energy
Johannes Hermann
Ken Brameld
Deborah Reuter

Why calculate strain energies ?
• Probably - high strained conformations are not binding to target protein
• Strained conformations might bind but lose activity
• Conformational strain information could be used to guide/prioritize designs
Problems:
• Data sets to define thresholds are not clean
• Forcefields have holes in the parameter set
Q. Wang, Y.-P. Pang PLoS ONE 2007, 2(9) e820
J. Tirado-Rives, W.L. Jorgensen J. Med. Chem. 2006, 49, 5880-5884
E. Perola, P.S. Charifson J. Med. Chem. 2004, 47, 2499-2510
• Forcefields are not ideal for calculating non minimum energies
Scientifically accurate investigation/calculation
of strain energies is difficult,
but pragmatic conclusions may be feasible

Strain Energy Calculation Strategy
CSD
• Source of low energy
conformations
• Relatively clean
PDB
• Probably more strained
conformations
• less clean
Test methods
Use methods

Strain Energy Calculation Scheme CSD/PDB
xray
ΔE local
ΔE global
E – global minimum
Minimization with
0.2Å wall box
restraints
=
=
E – xray relaxed
E – xray relaxed
Free minimizations
Energy ranking
E – - xray relaxed
-
-
Conf-ensemble
Free minimization
E – local minimum
E – global minimum
E – local minimum
Conformer
generation

Strain Energy Calculation Scheme PDB
xray
ΔE local
ΔE global
E – global minimum
Minimization with
0.2Å wall box
restraints
=
=
E – xray relaxed
E – xray relaxed
Free minimizations
Energy ranking
E – - xray relaxed
-
-
Conf-ensemble
Free minimization
E – local minimum
E – global minimum
E – local minimum
Conformer
generation

Databases / methods
filtered to contain drug-likeish molecules
PDB
inhouse + public
CSD
Package
Macromodel
MOE
Jaguar
200 < MW < 700
No P,B,As…..
Rotatable bonds < 10
Ring size < 10
Internal other filters
Forcefield /
basis set
MMFF94s
OPLS2005
MMFF94s
MMFF94x
B3LYP/6-31G*
PDB
inhouse + public
vacuum
CSD
Electrostatics/solvent
constant dielectric
with/out GBSA water model
constant dielectric
distance dependent
dielectric
3119 entries
15528 entries
4592 for DFT

Local strain in the CSD
different methods suggest low overall strain energy
Package
Jaguar
MOE
Macromodel
Forcefield/basis
B3LYP/6-31G*
MMFF94s
MMFF94x
MMFF94s
MMFF94s_GBSA
OPLS2005
OPLS2005_GBSA
Median
0.1
1.8
1.7
0.5
0.4
0.4
0.4
sd
1.8
4.9
4.8
2.1
1.9
1.8
1.7

Local strain energy distribution
> 99% of DFT local strain less than 4 kcal/mol
100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
Local Strain > 4 kcal/mol
1 – 4 kcal/mol
< 1 kcal/mol
DFT MMFF94s MMFF94x MMFF94s MMFF94sw OPLS2005 OPLS2005w
Jaguar MOE Macromodel

B3LYP/6-31* can be used as a control
high local strain with DFT identifies problematic compounds
Gross Structural Errors
JIYGED
69 kcal/mol
Hydrogen Placement Errors
EKIWAW
53 kcal/mol
BEWWIJ
19 kcal/mol
BATYIE
24 kcal/mol
QUADWUO
13 kcal/mol
TAKNIC
22 kcal/mol

ΔE MMFF94s MOE
Identification of poor forcefield parameterization
large discrepancy of FF vs. DFT suggest problematic substructures in compound
70 n = 4500
60
50
40
30
20
10
0
0 0.5 1 1.5 2 2.5
ΔE B3LYP/G6-31*
> Median + sd
< Median + sd

ΔE MMFF94s MOE
Identification of poor forcefield parameterization
large discrepancy of FF vs. DFT suggest problematic substructures in compound
70
60
50
40
30
20
10
0
n = 4500
8
7
6
5
4
3
2
1
0
0 0.5 1 1.5 2
dE l l B3LYP/6 31G*
0 0.5 1 1.5 2 2.5
ΔE B3LYP/G6-31*
~3000 (of 4500)
> Median + sd
< Median + sd

Problematic substructures MMFF94s
fingerprint based learner identifies frequent substructures in problematic sets
N +
N +
N
O O
O
S R
S R
O
N
H O O
O
CSD 40-75°
MMFF94s 0°
S R
S R
O
N
O

Comparison Macromodel & MOE MMFF94s
MOE minimizer finds lower local min - same molecules as outliers
ΔE MMFF94s MM
10 n = 15500
8
6
4
2
0
0 2 4 6 8 10
ΔE MMFF94s MOE
94 O

Learnings from CSD strain energy calculations
• Different forcefields / methods different strain energies
• Strain energies are not transferable,
need to have strain energy analysis for each software setup
• The drug-like subset of the CSD is very clean – only few artifacts (gross errors)
• Some relevant substructures are poorly parameterized in major forcefields
• Although less accurate, forcefields can be used for strain energy assessment
• Solvation models, treatment of electrostatics is less important for
strain energy calculations in the absence of protein

Local strain energy in the PDB
only slightly higher local strain observed than in the CSD
Method
MOE / MMFF94x
Relaxation
0.2 box wall
0.2 box wall in protein
free in protein
Median Median (CSD)
2.5 2.5 (1.7)
7.8
7.3
sd (CSD) sd
6.8 6.8 (4.8)
7.2
5.5

Much more problems in the PDB
many protein- and protein unrelated errors public & inhouse
2g1n
8/18 kcal/mol
1ikv
961 kcal/mol
1bqm
17 kcal/mol
2pxk
4/474 kcal/mol
1ya4
93 kcal/mol
3da2
14/58 kcal/mol
2pmo
31 kcal/mol

Local strain energies (no protein)
no correlation of strain energy with resolution, dpi, MW or number of rotors
ΔE local MMFF94x MOE
ΔE local MMFF94x MOE
14
12
10
8
6
4
2
0
14
12
10
8
6
4
2
0
Resolution
1 1.5 2 2.5 3 3.5
dpi
0 0.2 0.4 0.6 0.8 1
ΔE local MMFF94x MOE
ΔE local MMFF94x MOE
14
12
10
8
6
4
2
0
14
12
10
8
6
4
2
0
Rotors
0 1 2 3 4 5 6 7 8 9
MW
200 300 400 500 600

Local strain energies (no protein)
no correlation of strain energy with resolution, dpi, MW or number of rotors
ΔE local MMFF94x MOE
ΔE local MMFF94x MOE
14
12
10
8
6
4
2
0
14
12
10
8
6
4
2
0
Resolution
1 1.5 2 2.5 3 3.5
dpi
0 0.2 0.4 0.6 0.8 1
ΔE local MMFF94x MOE
ΔE
14
12
10
8
ΔE local MMFF94x MOE
6
4
2
0
-2
-4
14
12
14
12
10
8
6
4
2
0
-2
-4
-6
-8
ΔE -6
10
8
6
4
2
0
14
12
10
8
6
4
2
0
RotorsRotatable
bonds
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
MW
Binned MW
0 1 2 3 4 5 6 7 8
200 300 400 500 600

Global strain similar distributed as in CSD
OMEGA finds >95% of global minima in the PDB (1/3 = local)
20% 20
15% 15
10% 10
5% 5
0
CSD
PDB
x = -0.5 -0.5 < x = 0.5 0.5 < x = 3 3 < x = 6 6 < x = 10 10 < x = 15 15 < x = 20 20 < x
1 2 3 4 5 6 7 8 9
Bi d dE l b l
Binned ΔE global MMFF94s MOE/OMEGA

CSD global strain energy (MOE/MMFF94s, OMEGA)
ΔE global MMFF94s MOE/OMEGA
spiro- and aliphatic bridgehead atoms explain 2/3 of wrong global minima
0
-100
-200
-300
-400
-500
n = 15500
0 2 4 6 8 10
ΔE local MMFF94s MOE
spiro-center
> 1 aliphatic bridgeheads
other
OMEGA 2.3.2 settings:
-searchff & buildff MMFF94s
-rms 0.8
-noestat false

Global strain similar distributed as in CSD
OMEGA finds >95% of global minima in the PDB (1/3 = local)
20% 20
15% 15
10% 10
5% 5
0
CSD
PDB
x = -0.5 -0.5 < x = 0.5 0.5 < x = 3 3 < x = 6 6 < x = 10 10 < x = 15 15 < x = 20 20 < x
1 2 3 4 5 6 7 8 9
Bi d dE l b l
Binned ΔE global MMFF94s MOE/OMEGA

Local strain for isolated proteins/series
within a series strain energies become meaningful
# compounds
Mean
Spread E
Spread MW
Kinase
target 1
Series 1
11
11.5
8 - 16
250
Kinase
target 1
Series 2
7
7.5
2 - 13
220
Kinase
target 2
7
1.4
0.3 – 3.8
100
all compound’s IC 50 < 500 nM
Viral
target
6
2.9
1.4 - 3.4
90
Serine
protease
6
2.1
0.6 - 4.3
115

Conclusions
• Calculating strain energies is not as straightforward as one might think
• Individual inspection of PDB structures necessary
• Strain energy does not correlate strongly with resolution, dpi, MW or number of
rotatable bonds
• PDB and CSD strain energy distributions surprisingly similar
• PDB structures are very heterogeneous no universal threshold for
tolerated strain energy
• Isolated series may provide guidance on what strain energies are acceptable