Tackling Receptor Flexibility in

Computer-Aided Drug Design

Rommie Amaro . NBCR Mini-Symposium . August 4, 2008

 Computer-aided drug design

Challenges:
• solvation effects, entropy,
rigorous thermodynamics
• prediction of
lead/candidate
pharmocokinetic properties
• networks /
polypharmacology
• receptor flexibility

Van Drie, J., J. Comp. Aid. Mol. Des., 21: 591-601 (2007)
 Tackling receptor flexibility:
relaxed complex scheme

Amaro, Baron, and McCammon, J. Comp. Aid. Mol. Des..,in press (2008)
 Developing new antivirals against
avian influenza
• Biological introduction
• Investigating the dynamics and flexibility of N1
(molecular dynamics)
• Extracting meaningful information and reducing
redundancy (clustering analysis)
• Finding new druggable hot spots (computational solvent
mapping)
• Identifying new drugs for experimental testing (virtual
screening)
• Summary & future work
 Influenza virus

Hemaggluttinin
(16 subtypes)

Neuraminidase
(9 subtypes)

Host-derived lipid
envelope

M2 ion channel

8 RNA segments

No proof reading during

replication - highly variable
~100 nm diameter
 Influenza

• Epidemics are normal, seasonal influenza outbreaks

• est. 300,000-500,000 people die each year due to
epidemic influenza
• deaths highest among > 65 yrs old, children < 2 yrs,
immunocompromised
• Pandemics are rare events that occur every 10-50 years.
• In the last 400 years, at least 31 pandemics have been
recorded
• Circulate around the globe in successive waves
• With global travel, est. a new pandemic would reach almost all
corners of the earth within 3-6 months, an estimated 2 billion
of the world’s 6.5 billion people will be infected
 Origin of pandemic viruses

antigenic
drift antigenic
shift

40 million
1-1.5 million 0.75 - 1 million
deaths
deaths deaths

Clercq, Nat Rev. Drug Disc.,5: 1015-1025 (2006)

 H5N1 influenza cases 2003-2008

• It is especially virulent (~ 50% mortality rate) & being spread by migratory birds
• Bird to mammal, bird to human transmission
• Like other influenza viruses, it continues to evolve.
Points of intervention in the viral replication
cycle

Clercq, Nat Rev. Drug Disc.,5: 1015-1025 (2006)

 Neuraminidase as drug target

HAs preferentially recognize sialic acid-gal linkage:

Human-type

avian-type

- Developed against N9,

not N1
- Effective for N1, but not
as good
sialic acid DANA zanamivir Oseltamivir - Resistance is a problem
(Relenza) (Tamiflu)

Clercq, Nat Rev. Drug Disc.,5: 1015-1025 (2006)

Moscana, New Eng. J. Med., 353: 1363-1373 (2006)
 Group 1 and 2 neuraminidases
2 phylogenetically distinct groups:
9 neuraminidase (NA) strains:

Group 1 ?

Russell et al, Nature, 443: 45-49 (2006).

 Group 1 and 2 neuraminidases
2 phylogenetically distinct groups:

Russell et al, Nature, 443: 45-49 (2006).

 Goals
• To develop a more effective, orally-available drug
against N1

• Methodological goal: to develop optimized scheme for

receptor flexibility in inhibitor discovery process

• Use the structural information from MD as a predictive

guide and to expand the receptor ensemble

• Improve final ranking of compounds and account for

induced-fit effects, as part of improved drug discovery
scheme
 Molecular dynamics to probe structure
& dynamics

! % + * .
12
+ * .
6
'
( )
qi q j
U R = 3 bond
k ( r ! ro ) 2
+ 3 " k (" ! " o ) 2
+ 3 dihed &
k %1 + cos ( n# + $ ) '
( + 3 ij 1-, r 0/ -, r 0/ 2 3 ) r
4 ) 1 ij
!
ij
2 +
bonds angles dihedrals nonbonded
paris & ij ij
( nonbonded ij
pairs

Van der Waals

& electrostatics

! 2! ! !
Classical dynamics
at 300K :
d ri
Fi = ma = mi 2 = !"U R
dt
( )

Δt Δt
...
 Molecular dynamics simulations
• 2HTY (open loop, apo)
• 2HU0 (open loop, holo)
• N1 tetramer, (ligands), ions
• Explicit solvent, 150mM NaCl
• 112,457 atoms

• NAMD2 on supercomputers
• 5 ns/day
• 40 ns for the tetramer
(eq. of 160 ns of monomer)

Amaro, R. E., Minh, D.D.L., Cheng, L.S., Lindstrom, Jr., W., Olson, A.J., Lin, J.-H., Li, W.W., and McCammon, J.A., JACS, 129: 7764 – 7765 (2007).
 Remarkable loop flexibility

• Molecular dynamics allows

sampling of receptor side
chains and larger local
motions (e.g. loop sampling)

• Can account for induced

effects of particular ligand
(e.g. Tamiflu)

• Changes in ligand binding

site can be exploited and
designed around
 150-loop dynamics
open, closed

Amaro, R. E., Minh, D.D.L., Cheng, L.S., Lindstrom, Jr., W., Olson, A.J., Lin, J.-H., Li, W.W., and McCammon, J.A., JACS, 129: 7764 – 7765 (2007).
 Implications for Antiviral Drug Design

430-loop and 150-loop

very flexible

Structural reorganization
reveals new pocket
topography

Goal:
To use these new
structural insights for
drug discovery/design
efforts

Amaro, R. E., Minh, D.D.L., Cheng, L.S., Lindstrom, Jr., W., Olson, A.J., Lin, J.-H., Li, W.W., and McCammon, J.A., JACS, 129: 7764 – 7765 (2007).
Clustering distills essential information
• Extracted snapshots from 4 chains explicit 40 ns
simulations (160 ns for both apo & holo)
• Alignment based on Cα atoms
• Then computed RMSD distance matrix using
subset of 62 residues (sidechains included) lining
the binding pocket
 Computational solvent mapping

• Assesses druggability of receptor surfaces using complementary physics-

based approach
• 14 organic probes to flood receptor surface
• Probes clustered and ranked by interaction energy with surface
• Hot spots indicate areas of high functional group affinity
 “Hot spots” predict areas of affinity

• Structures revealed by
MD have new high
affinity areas for ligands,
ligand-extensions to bind

• These hot spots vary in

size, number, and moiety

• Indicates which
residues in new areas
may be important to
optimize against

Landon, M., Amaro, R.E., Baron, R., Ngan, C.H., Ozonoff, D., McCammon, J.A., and Vadja, S., Chemical Biology & Drug Design (2008).
 Discovering new inhibitors:
virtual screen with molecular dynamics

• Typical virtual screens use only one crystal structure

• Virtual screen of 3 most dominant MD cluster
representative structures & crystal structures
• Rapid docking with AutoDock
• Against the NCI diversity set ~ 2000 compounds
• Top candidates filtered for druglikeness & clustering
• Identified 27 novel putative inhibitors, half of which would
not have been found based on crystal structures alone
• Ordering of known sialic acid analog inhibitors is correct
(positive controls: Tamiflu, Relenza, DANA)

Cheng, L.S., Amaro, R.E., Xu, D., Li, W.W., Arzberger, P.A., and McCammon, J.A., Journal of Medicinal Chemistry, in press (2008).
Ensemble-based virtual screening
Closed 150-loop Open 150-loop

Including full receptor flexibility opens new areas for ligand binding
Potential cross-cavity binders

Several compounds are

predicted to bind 2 or
more cavities

May provide addt’l

selectivity for N1

Many ligands predicted

to dock to the CS-map
hot spots
 Relaxed complex scheme rescoring
Apo Holo
Mean Predicted Chemical
Rank NSC Binding Site Crystal Crystal
Energy Ki ( M) Structure
Rank Rank

H
HN N

1 109836 -10.63 0.016 HN N O

N+
O-
SA-cavity 15 1
NH O

N
C
HN
2 211332 -10.34 0.026 SA-cavity 212 10
H2 N N NH 2
H

N
SA-cavity
3 45583 -10.09 0.040 N

OH
150-cavity 6 18
O

S S
O
430-cavity
HO OH
O O

HN

0.063 O NH2

- Oseltamivir -9.82 SA-cavity 238 5

(0.3 – 1.0) O O

NH2

0.133 N NH 2

- Zanamivir -9.38 OH
HN
SA-cavity 230 12
(0.5 – 2.5) HO
O
OH

• Redock top compounds into MD

OH O

snapshots (receptor flexibility)

O O O

OH N+ O-

4 106920 -9.20 0.180 N

OH
430-cavity 99 71
O

• Binding spectrum reorders

compounds
O
HO O

5 17245 -9.18 0.187 150-cavity 10 65

HO
O

NH2 • Final set of 27 recommended

6 350191 -9.14 0.200
O
NH2 SA-cavity 336 113 compounds being experimentally
tested
S
P
HO
OH
 Rescoring can be important!

African trypanosomiasis

RNA editing ligase required for survival of parasite

• Rescoring of top compounds provided important enrichment of

recommended set
• Limited experimental resources, best inhibitors would not have been
tested without rescoring
Amaro, R., Schnaufer, A., Interthal, H., Hol, W., Stuart, K., and McCammon, J.A., submitted (2008)
 Future methodological work:
relaxed complex scheme

Developing a
workflow tool
using Vision

Needs to be
flexible so new
modules can be
easily added

Developing
cyberinfrastructure
to launch jobs,
deal with &
manipulate data

Amaro, Baron, and McCammon, J. Comp. Aid. Mol. Des..,in press (2008)
 Avian Flu Grid: an international
collaborative effort
mpirun

• 32 institutions in 16
Job submission
(globusrun ) SDSC node
GRAM (in US A) Users
countries across the
File I/O

AIST cluster gfsd

Collaboration
Pacific Rim and USA
(in Japan)
USM node
(in Malaysia) Users • N1 project “science
Data, program driver” for technology
Gfarm filesystem
development
PRAGMA testbed
- Computational server
MGrid-CHARMM MD Simulation Job Preparing
- Storage server
AIST , ASGC, CNIC, CUHK, GUCAS,
IOIT-HCM, LZU, MIMOS, NECTEC,
NGO, SDSC, ThaiGrid , UZH, VPAC

• Developing computational environment

(infrastructure) and scientific
applications
• Portal for datasharing
http://www.pragma-grid.net/
 Training and Outreach

H5N1 projects serve as training projects for undergraduate

students through PRIME and high schoolers through the
Pinhead Institute

Thursday & Friday Track III sessions will teach YOU how to
set up an MD simulation, perform analysis, submit a virtual
screen and perform a relaxed complex scheme rescoring
 Acknowledgements

Professor Andy McCammon The SAFI & Avian Flu Grid Teams
The McCammon Group

Molecular Graphics Lab

H5N1: why so deadly?
• H5N1 seems to induce
hypercytokinemia, a.k.a. “cytokine storm”
• Overreaction of the innate immune
system, which is highly complex in its
interactions with other signaling
molecules, is suspected to play a role in
the virulence
• Preference for sialic acid receptors in
the lower respitory tract (as opposed to
upper) = delayed side effects (sneezing,
coughing, etc) = longer virus incubation
period, so when presents itself, higher
viral load, tougher on the body

• Onset of symptoms to death: 9 days

 Biomolecular simulations & the future of
computer-aided drug design
• Increased computing power, entering the petascale era
• Simulations of hundreds of ns already possible, microseconds
soon to follow
• Highly optimized parallel code allow building of complexity
(bigger systems), without sacrificing speed
• Enabling of grid-based technologies offer alternative computing
platforms for docking or other small-processor request jobs
• As compute power grows, so will the scope and level of CADD
modeling
• Good predictions cut time to positive experiment, assist in
understanding mechanism of action, drive discoveries!
 Generalized Born MD
- Projects with Xiaolin Cheng & Ivaylo Ivanov (McCammon group)
- GB: Represents the solvent implicitly as continuum with the dielectric
properties of water, and includes the charge screening effects of salt:
N1-apo, closed loop | N9-apo, closed loop
N1-tamiflu, closed loop | N9-tamiflu, closed loop
N1-tamiflu, open loop | N1-apo, open loop
Tetramer N1 system = HUGE! (20K+ atoms)...
- 16 ns for each system, Amber igb version 5, monomer only, with
Amber’s fast pmemd MD engine (~5500 atoms: big for GB)
- On new NCSA Abe machine, scales to 256 - 512 procs, ~ 8 ns/day
- Comparative dynamics analysis between N1 vs. N9, tamiflu bound and
apo systems, open & closed loops… possibly sample more open/closed
loop transitions
Manuscript in preparation… may use snapshots for CADD work
 GB-MD Preliminary Results
open, closed

N1-apo-closed

N1-tami-closed

N1-apo-open
N1-tami-open

N9-apo-closed N9-tami-closed
 GB-MD Preliminary Results

N1-closed N1-open N9-closed

 Grid Maps

• Fast energy evaluation is achieved by precalculating atomic

affinity potentials (grid maps), one for each atom type in the
ligand
– Calculated by autogrid & a .gpf file
– Affinity grid: each point stores the potential energy of a
probe atom due to all atoms• Also
in themakes
macromolecule
electrostatic maps
• Define receptor atom types,
ligand atom types
npts 60 60 60
spacing 0.375
gridcenter 1.602 18.973 4.55

AutoDock User’s Guide, v3.0.5, Morris et al.

 AutoDock4 force field

0 0
!G = V ( L"L
bound "V L"L
unbound ) + (V P"P
bound "V P"P
unbound ) + (V P"L
bound "VP"L
unbound + !Sconf )
Intramolecular energies Intermolecular energies

!Sconf = Wconf N tors Loss of torsional entropy upon binding

Huey, Morris, Olson & Goodsell, J. Comp. Chem, A Semi-empirical Free Energy
Force Field with Charge-based Desolvation, preprint (2006).
 AutoDock force field

Semi-empirical: combines traditional MM force fields

with empirical weights and an empirical approach for
entropic contributions

" Aij Bij % " Cij Dij % ( )

( )
qi q j
V = Wvdw * $ 12 ! 6 ' + Whbond * E(t) $ 12 ! 10 ' + Welec * + Wsol * SiVj + S jVi e
! rij2 2) 2

i, j # ij
r rij & i, j # ij
r rij & i, j ( (r )r
ij ij i, j

W’s are the weighting factors optimized to

calibrate the empirical free energy based on 188
experimentally characterized complexes

Huey, Morris, Olson & Goodsell, J. Comp. Chem, A Semi-empirical Free

Energy Force Field with Charge-based Desolvation, preprint (2006).
 AutoDock force field

Semi-empirical: combines traditional MM force fields

with empirical weights and an empirical approach for
entropic contributions

" Aij Bij % " Cij Dij % ( )

( )
qi q j
V = Wvdw * $ 12 ! 6 ' + Whbond * E(t) $ 12 ! 10 ' + Welec * + Wsol * SiVj + S jVi e
! rij2 2) 2

i, j # ij
r rij & i, j # ij
r rij & i, j ( (r )r
ij ij i, j

Normal Lennard-Jones potential describing

dispersion/repulsion interactions
Parameters A and B taken from the Amber force
field.
 AutoDock force field

Semi-empirical: combines traditional MM force fields with

empirical weights and an empirical approach for entropic
contributions

" Aij Bij % " Cij Dij % ( )

( )
qi q j
V = Wvdw * $ 12 ! 6 ' + Whbond * E(t) $ 12 ! 10 ' + Welec * + Wsol * SiVj + S jVi e
! rij2 2) 2

i, j # ij
r rij & i, j # ij
r rij & i, j ( (r )r
ij ij i, j

Directional H-bond term based on a 10/12 potential.

C and D give a maximal well depth of 5 kcal/mol at 1.9 Å for O—H and
N—H, and a depth of 1 kcal/mol at 2.5 Å for S—H.
Directionality of the hydrogen bond interaction E(t) is dependent on the angle
t away from ideal bonding geometry. Note that the directionality is only with
respect to the receptor:
 AutoDock force field

Semi-empirical: combines traditional MM force fields with

empirical weights and an empirical approach for entropic
contributions

" Aij Bij % " Cij Dij % ( )

( )
qi q j
V = Wvdw * $ 12 ! 6 ' + Whbond * E(t) $ 12 ! 10 ' + Welec * + Wsol * SiVj + S jVi e
! rij2 2) 2

i, j # ij
r rij & i, j # ij
r rij & i, j ( (r )r
ij ij i, j

Electrostatics described by a screened coulombic potential

 AutoDock force field

Semi-empirical: combines traditional MM force fields with

empirical weights and an empirical approach for entropic
contributions

" Aij Bij % " Cij Dij % ( )

( )
qi q j
V = Wvdw * $ 12 ! 6 ' + Whbond * E(t) $ 12 ! 10 ' + Welec * + Wsol * SiVj + S jVi e
! rij2 2) 2

i, j # ij
r rij & i, j # ij
r rij & i, j ( (r )r
ij ij i, j

Final term is a desolvation potential

S: atomic solvation parameter for each atom type (estimate of E needed to
transfer the atom between a fully hydrated state and fully buried state)
V: estimate of the amount of desolvation when the ligand is docked,
calculated with a volume-summing method similar to Stouten et al.

AD4 has solvation constants for 22 atom types

Stouten et al., Molecular Simulation, 10: 97-120 (1993). *on the wiki!
 AutoDock4

• Fast energy evaluation is achieved by

precalculating atomic affinity potentials
– Affinity grids: each point stores the potential
energy of a probe atom due to all atoms in
the macromolecule
– Each atom type in ligand gets a map

• Full ligand flexibility around all torsions

– Lamarckian genetic algorithm
– Very efficient global search

• Based on comprehensive thermodynamic model that allows incorporation of

intramolecular energies into the predicted free energy of binding

• Charge-based method for evaluation of desolvation for typical set of atom

types2

• Calibrated against 188 diverse protein ligand complexes

AutoDock User’s Guide, v3.0.5, Morris et al. 2Stouten et al., Molecular Simulation, 10: 97-120 (1993).
 AutoDock4 force field

0 0
L"L
!G = Vbound (L"L
" Vunbound P"P
+ Vbound P"P
" Vunbound P"L
+ Vbound ) (P"L
" Vunbound ) ( + !Sconf )
Intramolecular energies Intermolecular energies

!Sconf = Wconf N tors Loss of torsional entropy upon binding

" Aij Bij % " Cij Dij % ( )

( )
qi q j
V = Wvdw * $ 12 ! 6 ' + Whbond * E(t) $ 12 ! 10 ' + Welec * + Wsol * SiVj + S jVi e
! rij2 2) 2

i, j # ij
r rij & i, j # ij
r rij & i, j ( (r )r
ij ij i, j

Huey, Morris, Olson & Goodsell, J. Comp. Chem, A Semi-empirical Free Energy Force Field with Charge-based Desolvation, preprint (2006).