1. Ivelin Georgiev Bruce Donald Donald Lab Duke University
OSPREY Tutorial
Ivelin Georgiev
Bruce Donald
Donald Lab
Duke University

2. Distribution of Structures
min( )
Maximum Likelihood
(pick most probable)
Global Minimum Energy Conformation
Bayesian ò 1 Z
(average over all conformations)
Probability « Energy using Boltzmann distribution

3. Distribution of Structures
min( )
Maximum Likelihood
(pick most probable)
Global Minimum Energy Conformation
`Bayesian’ ò 1 Z
(weighted average over all conformations)
Probability « Energy using Boltzmann distribution

4. maximum likelihood GMEC BD traditional-DEE MinDEE
DEE-based predictions aim at identifying the single best solution, and in that, they are similar to a maximum-likelihood approach. In contrast, the computational predictions may be based on an ensemble of low-energy conformations, as is the case in our K* algorithm. K* is an enzyme-ligand binding prediction algorithm. K* computes partition functions over ensembles of conformations, where the contribution of each conformation to the partition function is weighed using Boltzmann probabilities. The ratio of the partition functions for the bound and unbound states is then used to compute a provably-accurate approximation to the binding constant for the given enzyme-ligand complex. As an example, here is what an ensemble of conformations for a given enzyme active site may look like when bound to a ligand.
In our computational experiments, we use a mix of the DEE-based and K* protein design algorithms.

5. via conformational ensembles
traditional-DEE
GMEC
maximum
likelihood
MinDEE BD ∫ 1 Z
weighted
average
K*: provably-accurate
approximation to the
binding constant
via conformational ensembles a
DEE-based predictions aim at identifying the single best solution, and in that, they are similar to a maximum-likelihood approach. In contrast, the computational predictions may be based on an ensemble of low-energy conformations, as is the case in our K* algorithm. K* is an enzyme-ligand binding prediction algorithm. K* computes partition functions over ensembles of conformations, where the contribution of each conformation to the partition function is weighed using Boltzmann probabilities. The ratio of the partition functions for the bound and unbound states is then used to compute a provably-accurate approximation to the binding constant for the given enzyme-ligand complex. As an example, here is what an ensemble of conformations for a given enzyme active site may look like when bound to a ligand.
In our computational experiments, we use a mix of the DEE-based and K* protein design algorithms.
sequence K*
TIAAIC
7.3
GIRMQM
3.1
TGIAIV
2.9
LMLAIS
1.7
TWAIGY
0.3
Application:
Enzyme-Ligand Binding

6. K* s1 s2 … si … sk thousands of sequences!!! MinDEE A* BD
J. Comp. Chem. (2008)
conformations
pruned
1 - ε
partition
function
ε approximation si
fraction evaluated confs
K* is fully compatible both with MinDEE and BD. K* uses MinDEE and BD as an initial pruning filter and the A* branch-and-bound search in the enumeration stage. Previously, partition functions could only be computed for a handful of sequences. With K*, however, we routinely compute partition functions in mutation searches for up to 20,000 sequences. One of the main reasons that allows us to perform computations of this size is the provable epsilon-approximation algorithm for partition functions. Since each conformation contributes exponentially in the negative of its energy to the partition function, only a very small subset of low-energy conformations must be evaluated by K* in order to guarantee that the computed function is at least (1-epsilon) from the full partition function where all conformations are included. As is seen in the figure on the left here, which represents the fraction of conformations evaluated by K* for a set of several hundred test sequences, K* is extremely efficient, evaluating at most 0.5% of all conformations for any given sequence. … sk

7. Cheng-Yu Chen Ivelin Georgiev Amy Anderson Bruce Donald
Example (PNAS, 2009)
Cheng-Yu Chen
Ivelin Georgiev
Amy Anderson
Bruce Donald

8. NonRibosomal Peptide Synthetases (NRPS)
NRPS enzymes found in some fungi and bacteria
NRPS enzymes make peptide-like products with pharmaceutical properties (antifungal, antineoplastic, antibacterial ) e.g. vancomycin, penicillin, gramicidin, bacitracin, cyclosporin, bleomycin, …
NRPS similar to PKS
Nonribosomal Peptide Synthetases (NRPS, for short) complement the traditional ribosomal peptide synthesis in a variety of bacteria and fungi. NRPS enzymes are big multidomain proteins that produce peptide-like natural products. Among the NRPS products are natural antibiotics, immunosuppressants and antineoplastics. Penicillin is probably the most well-known NRPS product (by the way, here is the original paper), but other NRPS products include gramicidin, bacitracin, vancomycin. NRPS are similar to the polyketide synthases.
As the famous story goes, Fleming went away for a weekend and when he returned he noticed that bacterial (staphylococcus aureus) growth was inhibited near a fungal colony. "that's funny..." and the rest is history.
Fleming and his colleagues did a series of experiments that demonstrated the ability of extracts of the fungus to kill various gram-positive bacteria, including Staphylococcus pyogenes, S. viridans, Micrococcus spp., and a few others. Effectiveness against gram-negative bacteria (such as Escherichia coli and Klebsiella pneumoniae) was limited, with very high concentrations of penicillin needed to kill those organisms. We now know that a membrane protects those gram-negative bacteria against penicillin.
Prevents cross-linking of elements of the bacterial cell wall causing bacterial cell lysis. Stops the last step in the wall synthesis, the cross linking of two peptidoglycan strands.
FPVOL

9. NRPS: GrsA-PheA Redesign
gramicidin S
Phe
Leu
The system in which our lab is mainly interested is the Phe adenylation domain of the NRPS enzyme GrsA, or GrsA-PheA for short. GrsA in concert with GrsB makes the antibiotic gramicidin S. Our goal is to redesign GrsA-PheA to switch its specificity towards different non-cognate substrates, which eventually may be incorporated into modified versions of gramicidin S.

10. Protein Redesign (NRPS)
Three-dimensional structure of GrsA PheA domain
[Conti et al., 1997]
Three-dimensional structure of GrsA PheA domain is known
[Conti, Stachelhaus, Marahiel, Brick. 1997]

11. Change specificity from Phe to Leu by allowing any 2 (of 9) mutations
Mutations to GAVLIFYWM
Appx Mutation Sequences = ,000,000 Conformations
(78,200 after pruning) -
+H3N
CO2
r = 9 s = 2
Leu

12. Crystal Structure: 1amu (1.9 Å) 563 a.a., 65 kD (K517) I330 C331
AMP
D235
A322
A301
A236
Here shows the binding pocket from the crystal structure. 9 resides that are important in binding of the substrate are shown here.
W239
T278
I299

13. Three-Step Enzyme Redesign
K*: active site mutations
Entropy step: mutatable positions
MinDEE: bolstering mutations
Ivelin Georgiev, Cheng-Yu Chen 1. 2. 3.
K* allows us to predict mutations to the active site of a protein in order to improve its specificity for a given target substrate. In our most recent work, we have extended K* to a three-step enzyme redesign algorithm that, in addition to the active site mutations, predicts mutations both close to and far away from the active site for further improvement in the substrate specificity. K* is applied in the first step of the three-step algorithm. The second-step uses an entropy-based approach to predict residue positions anywhere in the protein that may be tolerant to mutations – I will briefly describe this step shortly. In the third step, MinDEE is used to predict mutations to the residue positions from step 2 for additional improvement in the desired substrate specificity. Step 2 heuristically identifies additional residue positions for mutation. Steps 1 and 3 are provably with respect to the input model.
provable
heuristic
provable
Computational Structure-Based Redesign of Enzyme Activity. PNAS (2009)

14. T278L/A301G with Leu AMP K517 #1 3,000 sequences
6.8  108 rotameric conformations
PNAS (2009)

15. Mutations Outside the Active Site
rotamer
probabilities AA
probabilities
mutatable
positions
SCMF
residue entropy
Boltzmann
MinDEE
S447
I277
V187
F45
V238
I207
L210
The entropy step uses an approach based on work by Steve Mayo and co-workers published in Self-Consistent Mean Field is used to estimate AA probabilities, and consequently, the residue entropy for each residue position in the protein. Residues with high entropy imply they can accommodate multiple amino acid types, and so may be tolerant to mutation. In Mayo’s paper, this approach was used as a preprocessing step for directed evolution experiments. We have modified and extended this approach to predict high-entropy residue positions that are subsequently redesigned using our MinDEE algorithm.
PNAS (2009)

16. 6.8  108 rotameric conformations
All top 10
3,000 sequences
6.8  108 rotameric conformations
A301G/T278L
[L-Leu] mM
Rate (AU/s * 10^{-6})
Kd by stopped-flow for A301G/T278D:Leu is improved >1000X over WT:Leu.
if you know the Vmax and the total [enzyme], you can calculate Kcat as Kcat [Etotal] = Vmax. Also, Kcat/Km = v/[E][S].
Leu
Phe
Normalized kcat/ KM
PNAS (2009)

17. T278D/A301G with Arg AMP D235 K517 301G 278D W239 L-Arg WT L-Lys WT
Arg: #1 of 2511 sequences
Lys: #4 of 2511 sequences
>9  108 conformations WT
AMP
[L-arg] mM
D235
K517
D – Aspartic Acid (Asp)
Rate (AU/s * 10^{-6})
301G
L-Lys
W239
278D WT
PNAS (2009)

18. Tutorial

20. Installation
Setup
Running OSPREY

21. Installation √ 32-bit 64-bit Java mpiJava MPICH2 may require
special instructions

22. Setup
Compute Nodes
Input Structure
Rotamer Library
Energy Function

23. Compute Nodes Select MPI nodes: linux1 Select job-specific nodes:
mpdboot
mpdboot -n 5 -f mpd.hosts
Select job-specific nodes:
linux1
linux2
linux3
mpirun
java
OSPREY
mpirun -machinefile ./machines -np 5 java -Xmx1024M KStar mpi -c KStar.cfg

24. Input Structure KiNG missing atoms model delete possible
REMARK 470 MISSING ATOM
REMARK 470 THE FOLLOWING RESIDUES HAVE MISSING ATOMS (M=MODEL NUMBER;
REMARK 470 RES=RESIDUE NAME; C=CHAIN IDENTIFIER; SSEQ=SEQUENCE NUMBER;
REMARK 470 I=INSERTION CODE):
REMARK 470 M RES CSSEQI ATOMS
REMARK 470 GLU A 34 CG CD OE1 OE2
REMARK 470 GLU A 63 CD OE1 OE2
missing
atoms
KiNG
model
delete
possible
over-constraint
possible
under-constraint

25. Accelrys DS Visualizer
Input Structure
adding
hydrogens
proteins
general
compounds
recommended:
MolProbity
recommended:
Accelrys DS Visualizer
Check: protonation states
missing protons

26. Input Structure
His
residues
HIP
HIE
HID

27. Input Structure
steric
shell
close to design site
significant speedup

28. Input Structure Other considerations: protein, ligand, cofactor
ligand: natural AA, small molecule
water molecules
no chain ID’s
unique residue numbers
protein-peptide, protein-protein
connectivity (good input structures)

29. Check and double-check!!!
Input Structure
Check and double-check!!!

30. Rotamer Library rotamers Richardsons’ Penultimate general proteins
compounds
# dihed
# rot
name
TYR 2 4
N CA CB CG
CA CB CG CD1
62 90
TYR 2 5
N CA CB CG
CA CB CG CD1
62 90
FCL 2 4
N CA CB CG
CA CB CG CD1
62 90 1 2
one rotamer

31. Energy Function add params for new atom types antechamber typically
parm96a.dat
all_amino94X.in
all_nuc94_and_gr.in
atom types
dihedral parameters
vdW parameters
amino acids
partial charges
connectivity
general compounds
partial charges
connectivity
add params for
new atom types
antechamber
typically
no changes
add params for
new compounds
antechamber
can modify
partial charges
user control: distance-dependent dielectric, dielectric value,
vdW radii scaling, solvation energy scaling, dihedral energies switch

32. Running OSPREY
GMEC-based
Ensemble-based
Residue entropy

33. GMEC-based doDEE DACS input structure rotamer library energy function
mpirun -machinefile ./machines -np 5 java -Xmx1024M KStar mpi -c KStar.cfg doDEE System.cfg DEE.cfg
input structure
rotamer library
energy function
mutation search
parameters
doDEE
energy minimization
(MinDEE, BD, BRDEE)
DACS
1 MET GLY ASP ARG FCL unMinE: minE: bestE:
2 MET GLY ASP MET FCL unMinE: minE: bestE:
3 MET GLY ASP ARG FCL unMinE: minE: bestE:
1 MET GLY SER ARG FCL unMinE: minE: bestE:
2 MET GLY SER ARG FCL unMinE: minE: bestE:

34. GMEC-based genStructDEE rank input structure rotamer library
java -Xmx1024M KStar -c KStar.cfg genStructDEE System.cfg GenStruct.cfg
input structure
rotamer library
energy function
struct generation
parameters
genStructDEE
energy minimization
(MinDEE, BD, BRDEE)
1 MET GLY SER ARG FCL unMinE: minE: bestE:
2 MET GLY SER ARG FCL unMinE: minE: bestE:
3 MET GLY ASP ARG FCL unMinE: minE: bestE:
1 MET GLY ASP ARG FCL unMinE: minE: bestE:
2 MET GLY ASP MET FCL unMinE: minE: bestE:
3 MET GLY ASP ARG FCL unMinE: minE: bestE:
1 MET GLY SER ARG FCL unMinE: minE: bestE:
2 MET GLY SER ARG FCL unMinE: minE: bestE:
rank

35. Ensemble-based: Protein-ligand binding
mpirun -machinefile ./machines -np 5 java -Xmx1024M KStar mpi KSMaster System.cfg MutSearch.cfg
bound structure
rotamer library
energy function K*
mutation search
parameters
KSMaster
energy minimization
(MinDEE, BD, BRDEE)
doSinglePartFn
E+24 ILE TRP ILE ALA ALA ILE
E+24 TRP ASP ILE GLY ALA ILE
E+24 ILE THR ILE PHE ALA ILE
E+24 VAL THR ILE PHE ALA ILE
E+24 ILE THR ILE TYR ALA ILE

36. Residue entropy doResEntropy input structure rotamer library
mpirun -machinefile ./machines -np 5 java -Xmx1024M KStar mpi doResEntropy System.cfg ResEntropy.cfg
input structure
rotamer library
energy function
mutation search
parameters
doResEntropy
entropy
res ID
# prox res
AA probabilities

37. Some important parameters
mpirun -machinefile ./machines -np 5 java -Xmx1024M KStar mpi -c KStar.cfg doDEE System.cfg DEE.cfg
KStar.cfg:
hElect true
hVDW false
hSteric false
distDepDielect true
dielectConst 6.0
vdwMult 0.95
doDihedE true
doSolvationE true
solvScale 0.8
stericThresh 0.4
softStericThresh 1.5
rotFile LovellRotamer.dat
grotFile GenericRotamers.dat
volFile AAVolumes.dat
energy
function
steric filter
rotamer libraries
volume filter

38. Some important parameters
mpirun -machinefile ./machines -np 5 java -Xmx1024M KStar mpi -c KStar.cfg doDEE System.cfg DEE.cfg
System.cfg:
pdbName 1amuFH.pdb
numInAS 4
residueMap
pdbLigNum 566
ligAA false
numCofRes 1
cofMap 567
input pdb
design site
ligand
cofactor

39. Some important parameters
mpirun -machinefile ./machines -np 5 java -Xmx1024M KStar mpi -c KStar.cfg doDEE System.cfg DEE.cfg
DEE.cfg (partial):
doDACS true
distrDACS false
initDepth 2
subDepth 1
diffFact 6
doMinimize false
minimizeBB false
doBackrubs false
backrubFile none
useEref true
ligPresent false
ligType none
resAllowed0 gly ala val leu ile tyr phe trp met …
resAllowed3 gly ala val leu ile tyr phe trp met
resumeSearch false
resumeFilename runInfo.out.partial
DACS
minimization
reference energies
ligand in search
allowed mutations
resuming

40. Some important parameters
mpirun -machinefile ./machines -np 5 java -Xmx1024M KStar mpi KSMaster System.cfg MutSearch.cfg
MutSearch.cfg (partial):
mutFileName 1amuFCL_2MUT.mut
numMutations 2
targetVolume 620.0
volumeWindow
doMinimize false
minimizeBB false
doBackrubs false
backrubFile none
epsilon 0.03
gamma 0.01
repeatSearch true
useUnboundStruct false
unboundPdbName none
resAllowed0 gly ala val leu ile tyr phe trp met
resumeSearch false
resumeFilename 1amuFCL_MutSearch.partial
volume filter/
candidate mutants
minimization
(1-ε) accuracy
inter-mutation
at most 1 repeat
unbound struct
allowed mutations
resuming

41. Citing OSPREY General citation: K* and MinDEE: BD: BRDEE: DACS:
Original K* publication:

42. OSPREY is open source!!!

43. Acknowledgements Funding: Bruce Donald Ryan Lilien Faisal Reza
Kyle Roberts
Daniel Keedy
Pablo Gainza
Donald Lab
Funding:
NIH