Optimizing Drug Design - PowerPoint PPT Presentation

1 / 46
About This Presentation
Title:

Optimizing Drug Design

Description:

Inhibition measured by Marta Rosin (Novartis' Excellon) , Hebrew University School of Pharmacy ... Prof. Marta Rosin HUJI, AChE inhibitors. Optimizing Drug ... – PowerPoint PPT presentation

Number of Views:443
Avg rating:3.0/5.0
Slides: 47
Provided by: david895
Category:

less

Transcript and Presenter's Notes

Title: Optimizing Drug Design


1
Forming focused libraries and discovering active
molecules with Iterative Stochastic Elimination
Amiram Goldblum, Anwar Rayan and David
Marcus Dept. of Medicinal Chemistry School of
Pharmacy Ein Kerem Campus http//www.md.huji.ac.il
/models
2
Iterative Stochastic Elimination (ISE)Our
Generic tool for optimizing highly complex
combinatorial problems
Problem type Systems with many variables, each
variable having many discrete values, the
variables interacting with each other, and each
state of the system can be evaluated and given a
score (transportation, communication, electronic
devices, life sciences) Method ISE finds optimal
system states (global and local minima/optima) by
iteratively eliminating values of variables that
contribute to worst results. Elimination is based
on careful statistics of randomly picked states
of the system Why ISE has been compared to
Genetic Algorithms, Monte Carlo, Simulated
annealing, Support Vector Machines and other
optimization methods on specific problems and
found to do as well or better
3
Iterative Stochastic Elimination publications
  • Glick, M. Goldblum, A. A novel energy-based
    stochastic method for positioning polar protons
    in protein structures from X-rays.
    Proteins-Structure Function and Genetics 38,
    273-287 (2000).
  • Glick, M., Rayan, A. Goldblum, A. A stochastic
    algorithm for global optimization and for best
    populations A test case of side chains in
    proteins. Proceedings of the National Academy of
    Sciences of the United States of America 99,
    703-708 (2002).
  • Noy, E., Gorelik, B., Rayan, A. Goldblum, A.
    Stochastic path to form ensembles and to quantify
    flexibility in proteins. Abstracts of Papers of
    the American Chemical Society 225, U781-U781
    (2003).
  • Rayan, A., Barasch, D., Brinker, G., Cycowitz,
    A., Geva-Dotan, I., Scaiewicz, A. Goldblum, A.
    New stochastic algorithm to determine
    drug-likeness. Abstracts of Papers of the
    American Chemical Society 226, U297-U297 (2003).
  • Rayan, A., Scaiewicz, A., Geva-Dotan, I.,
    Barasch, D. Goldblum, A. Screening molecules
    for their drug-like index. Abstracts of Papers of
    the American Chemical Society 228, U358-U358
    (2004).
  • Rayan, A., Senderowitz, H. Goldblum, A.
    Exploring the conformational space of cyclic
    peptides by a stochastic search method. Journal
    of Molecular Graphics Modelling 22, 319-333
    (2004).
  • Rayan, A., Noy, E., Chema, D., Levitzki, A.
    Goldblum, A. Stochastic algorithm for kinase
    homology model construction. Current Medicinal
    Chemistry 11, 675-692 (2004).
  • Rayan, A., Scaiewitz, A., Geva-Dotan, I., Marcus
    D., Barasch, D. Goldblum, A (2007). Determining
    the Drug Like character of molecules and
    prioritizing them by a drug like index, ACS
    presentations 2005-8.
  • Noy, E., Tabakman, T. Goldblbum A.
    Constructing ensembles of flexible fragments by
    ISE is relevant to protein-protein interfaces,
    Proteins (2007) 68, 702-711
  • Gorelik, B Goldblum, A. High Quality binding
    modes in docking ligands to proteins. Proteins
    (2008), 71, 1373-1386

4
General Model System
A1
A2
  • Variables
  • Values
  • Interactions

B5
B4
B6
A7
A
B
B7
A6
C7
B8
C
C6
The number of combinations 7(A)x8(B)x7(C)xn(D)xm(
E).. A very large number
D
C5
E
C4
  • An exhaustive calculation is not possible

5
(1) Randomly pick one value for each of the
variables
B4
A7
A
This determines a single conformation or
configuration of the system
B
(2) Employ the cost function to score the
current configuration
C
D
C5
E
6
(3) Repeat steps (1) and (2) for n
conformations (n103-106), and calculate the
total value of each
sample 2
2nd value
.
.
.
.
.
.
sample n
nth value
7
(4) Construct a histogram of the distribution
of values for all sampled conformations
low values region
high values region
8
(No Transcript)
9
(6) Evict values that contribute above
expectation to worst scores, and less than
expected to best
A3
B4
B8
B4
C6
D7
D2
D6
E8
E8
E2
F1
F2
F9
conformation 715
conformation 314
conformation 220
The total number of combinations is reduced
(7) Repeat the process iteratively until all
remaining combinations can be evaluated
exhaustively and sorted. We obtain a population
10
Acetylcholinesterase inhibitors with
ISE Inhibition measured by Marta Rosin (Novartis
Excellon) , Hebrew University School of Pharmacy
Molecular chemical properties
ISE engine
ISE Docking and scoring
11
Bcr-Abl dimerization inhibition by peptides 64aa
Synthesized and measured by Martin Ruthardt,
Goethe Univ. Frankfurt
Properties of amino acids
ISE engine
ISE protein design
12
Distinguishing between actives and inactives, on
a specific target Classification Drugs vs.
Non-drugs, Selectives vs. non Selectives Huge
combinatorial problem with more than 10100 options
Optimization problem find differences in
molecular properties to distinguish between
actives and inactives
13
Learning from known data
Actives Molecules with activity lt
100nm Selectives Molecules with selectivity gt
31
Inactives MDDR (randomly picked), or less
actives
Properties (descriptors, our variables) are
produced by computer programs (MOE) Molecular
weight, number of H-bond donors acceptors,
partial charges, topological, polar surface, Van
der Waals, Molar refraction etc
14
Optimization of property ranges by ISE to
distinguish between the two databases
Each property is separated into two sub
properties
0
1200
Lower Range 0
800 80 values at intervals
of 10
Overall there are 8070 5.6103 combinations
for ranges of this variable
15
Using properties to optimize the
difference between actives (selectives) and
inactives
2 lt HD ? 6 -2 lt logP ? 3 150 lt M.W ? 775
  • If we construct a RANGE for each property
  • Then we test each of the molecules in the
    Actives and each in the inactives

A FILTER
  • Determine if TP, TN, FP, FN
  • ( P N Pf Nf)
  • Compute the fraction of each category in the
    full DB
  • Use the Matthews Correlation to score

16
Scoring by the Matthews Correlation Each given
range is for ACTIVES, and actives can only be P
or Nf
For a fully correct prediction C 1 For a
completely erroneous prediction C - 1 For a
random prediction C 0.00
17
Applying ISE to discriminate between actives and
inactives by optimizing descriptor ranges
Construct filter i Pick randomly a value for
each of the variables, i.e., low range MW, high
range MW etc.
Pass all actives and inactives of the training
set through filter i
P, N, Pf, Nf
Get MCC value for filter i
Until i 106
Histogram, Elimination, Iteration, Exhaustive,
Test
18
Results of exhaustive step, before clustering
Best filter
19
Employing the best sets of filters to
construct a Molecular Bioactivity Index
With good data, the range of MBI is large and we
get a good resolution We have shown that we
can use MBI to fish a few active molecules out
of a sea of inactive ones http//www.md.huji.ac.
il/models (look for test MBI)
20
Employing the best sets of filters to
construct a Drug Likeness Index (DLI)
Drug Likeness is different than Lipinskis ROF !
21
MBI and DLI can make a difference in
  • High Throughput Screening
  • Combinatorial Synthesis
  • Hit to lead development
  • Lead optimization
  • Construction of Focused libraries
  • Molecular scaffold optimization
  • Selectivity optimization

22
Timeline for discovery, single processor One
target (enzyme, cells, organs)
23
Input VEGFR-2 KDR active inhibitors lt100nm
549 actives divided randomly into 412 training
and 137 test set Inactives are from MDDR
24
Output example of a filter with 6
descriptors One of the best (high MCC) there are
others with higher MCC but many desciptors
Number of descriptors 6 MCC of test set
0.79 TP - 98.9 TN - 78.6
Bcut_SMR_3 0.0 3.06 SMR_VSA4
0.1 - 100.6 Vsa_pol 0.1
102.4 Reactive 0.0 0.999
balabanJ 0.0 - 1.902 Q_RPC-
0.0 0.267
25
A 6-property filter
Bcut_SMR_3 Molar refraction SMR_VSA4 VdW
surface area Vsa_pol Approx VdW polar
surface Reactive Reactive fragments balabanJ T
opological variable Q_RPC- Relative Negative
partial charge
26
Enrichment in the training set of VEGFR2
27
Initial focused library from ZINC (2.1 million)
ZINC library screening gave 7826 molecules with
top MBI
28
Similarity of highest MBI to training set
29
BBB results
30
ER-MBI moving ensemble(normalized MBI values)
logRBA
ER-MBI
31
ER-MBI Combined high/low MBI
32
Molecular bioactivity index
33
Molecular Bioactivity Index (MBI) Fishing
actives from a bath of non-actives Mix 10 in
100,000 - find 9 in best 100, 5 in best 10
Enrichment of 5000
Enrichment of 900
34
Polypharmacology with our indexing method
  • We use several MBI (or MBI and DLI) to map
    activity into multiple targets. This may be used
    to extract potential new poly-active compounds or
    selective compounds depending on the behavior of
    the relevant disease

35
Docking Scoring
Do the molecules bind ?
How strong is the binding affinity ?
How does the complex look like ?
Binding mode
Score
X-ray, NMR, Homology model
Requirement 3D structure of the target
36
ISE-dock
  • A new docking program from our lab that uses the
    ISE algorithm in order to produce large sets of
    optimal results for docking of ligands to their
    targets

37
ISE-dock
  • Better than AutoDock the most cited docking
    program
  • Much better in the main docking criteria than
    other two popular programs Glide and GOLD
  • Produces large near optimal docking populations
    to study the nature of binding and to predict
    alternative binding modes
  • Accounts for ligand and protein flexibility
  • Correlation between ISE-dock populations and
    experimental multiple binding modes

38
Anti Alzheimer current main drug strategy
39
Based on 450 active molecules with IC50 lt 10
micromolar 8000 randomly picked molecules from
ZINC assumed to be inactives
40
Docking with ISE-dock/Autodock We used the
crystal structure of mouse AChE (1q84) for
docking. Compounds in protonated state were
docked to AChE by AutoDock3.0 and ISE-Dock. 751
out of 755 compounds were docked in the active
site by both methods
41
(No Transcript)
42
10 compounds from docking results (financial
limitation) The 10 compounds were picked by
direct examination of each of these molecules in
the active site, paying utmost attention to its
conformation, H-bonds and other interactions.
43
Experimental Results 9 out of the 10 compounds
were purchased 8 out of the 9 compounds reached
our lab with enough quantity 5 out of the 8
compounds are soluble 3 out of the 5 compounds
are active (IC503.25, 3.5, 3.75 µM) Similarity
to known active compounds is less than
0.35 molecules are novel AChE inhibitors (not a
single paper on any)
44
Conclusions
  • ISE is useful for solving extremely complex
    optimization problems
  • Provides large sets of graded results
  • Achieves high enrichments of actives vs.
    inactives by MBI, DLI, MSI etc.
  • Useful for developing multi-targeted drugs
  • Discovers new binders for known drug targets
  • Produces diverse sets of solutions

45
Molecular Modeling Group Partners http//www.md.h
uji.ac.il/models http//www.cancergrid.eu
Prof. Andrej Bohac Comenius U, Bratislava,
VEGFR2 (Angiokem) DAC company Milan, HDAC
and HSP90 inhibition Prof. Mart Sarma U.
Helsinki, RET Kinase inhibition Prof. Martin
Rhutardt U. Frankfurt, Bcr-Abl inhibition by
peptides Prof. Yousef Najajreh Al Quds
University, Bcr-Abl inhibitor synthesis Prof.
Yossi Schlessinger Yale, FGFR inhibitors Prof.
David Varon Hadassah, Jerusalem, ADAMTS-13
inhibition Prof. Angelo Carotti School of
Pharmacy, Univ. of Bari, MMP inhibitors Prof.
Marta Rosin HUJI, AChE inhibitors
46
Molecular Modeling Group, HUJI http//www.md.huji
.ac.il/models
Write a Comment
User Comments (0)
About PowerShow.com