MultipleActiveSite Corrections for Docking and Virtual Screening

1
Multiple-Active-Site Corrections for Docking and
Virtual Screening
Guy Vigers Array BioPharm OpenEye CUP V March
2nd, Santa Fe
2
The Aim To predict binding modes and affinities
of small molecules with proteins.
The Question How well do current docking
programs perform? 1) Dock individual ligands
into their binding sites. Can we predict the
binding conformation? 2) Dock multiple potential
ligands into each site. Can we pick out the
right one?
3
Test Proteins
4
Test Ligands
5
FlexX/Chemscore
Number correct 3/15. MRE 25.0
pvh_lig
6
Gold Score, 4/8 VdW potential
Number correct 4/15. MRE 32
poc_lig
7
Fred for 15 X-ray Structures
Number correct 5/15. MRE 17.2
epb_lig
8
Gold RMSDs
Run1 1x speedup, 6/12 VdW potential Run2 1x
speedup, 4/8 VdW potential Run3 3x speedup, 4/8
VdW potential
Glq_lig
9
Multiple Active Site Correction
B
A
B
gtgt
A
Molecule Dependent Biases
10
FlexX MASC
Number correct 11/15. MRE 6
11
Gold - MASC
Number correct 11/15. MRE 9
12
MASC-Fred
Number correct 11/15. MRE 6
13
How does MASC-scoring work?
A thermodynamic cycle for docking? Not
yet! A correction for molecule-dependent
biases? Additive terms corrected by m. Scale
factors corrected by s An indicator of
significance?
14
Why did we pick Gold?

• Very slow (8mins/molecule). No electrostatics.
• But on some classes of compounds it is the only
  program that works!
• Requires 4/8 internal VdW potential.
• Correct conformations of small molecules simply
  not generated by Omega.

15
An example of molecule-dependent biases
PTP1b P38 -31.24
-29.70 -13.9 -24.25

• A

FlexX Scores

• B

16
X-ray co-crystal structures
3 space groups P3121 P21 P212121
3.2-1.8A resolution
17
Can MASC Sort P38 Inhibtors over PTP1b Inhibitors?
B
A
B
A
30 Known PTP1b Inhibitors vs 30 Known P38
Inhibitors Added 600 Random Small Molecules
from Cleaned Version of the MDDR Preclinical to
Launched No peptides, metals, inorganics,
etc. Removed salt forms, hydrates, etc. Dock
all compounds vs P38a and PTP1b
18
P38a Active Site Using Gold Scoring
19
P38a Selectivity using Gold-MASC Scoring
20
Ingredients for MASC-scoring and Virtual Screening
Docking software Active site 7 control
sites 3D conformations for all
molecules (currently 330,000 in Array
database) 64-CPU Linux cluster Relational
database to store docking scores
21
A Relational Database for MASC-scoring
22
Code for MASC-scoring
! /usr/bin/perl -w Read data_hash my
data_hash () my r_data_hash \data_hash
GoldSubsread_hash(r_data_hash)
Establish database connections my dbh
ChemregSubsdb_connect(r_data_hash) Run
the programs ChemregProgsscore_target(dbh,r_da
ta_hash) ChemregProgsscore_controls(dbh,r_da
ta_hash) ChemregProgstop_to_hitlist(dbh,r_da
ta_hash) dbh-gtdisconnect exit
23
Does it work?
Test Case Enzyme structure determined by
X-ray crystallography ( no ligand in active
site). 10 Array database virtually-screened
with MASC correction. 3 libraries identified
for in vitro assays. -gt No hits. (But No hits
on this target at all so far, from any technique.)
24
Whats Next?
Weiner-filter correction. MASC (S m)s/(s2
f2) Generate m and s directly from molecular
structure. Use ?MWt as estimate of s Better
Estimates? Linear combinations of Scoring
Functions Combine scores, then MASC-correct,
or MASC-correct each score, then
combine. (Sybyl Cscore doesnt work right in
6.91)
25
Whats Next (2)?
Better scoring functions. Free energy DG
measurements? Solvent Electrostatics .. Prote
in Flexibility Use B-factor data? De-Novo
Design Evolutionary algorithms for SBDD
26
Conclusions
Multiple Active Site Correction Reduces molecule
biases for all algorithms tested Gives better
rank order across compounds Gives meaning to the
actual score More computationally
intensive Easily adapted to virtual
screening. Who cares how fast we get the wrong
answer. Many opportunities to obtain incorrect
results Protein flexibility Water
interactions Scoring functions. Lots more fun
to come.
27
Acknowledgements
Jim Blake Ken Brameld Barb Brandhuber Kevin
Condroski Ellen Laird Jim Rizzi