Fragment Screening in LigandScout

Transcription

Pharmacophores and Their Applications
Thierry Langer
Prestwick Chemical, France
28 Aug 2013
Acknowledgements
Markus Böhler
Yasmin Aristei
Oliver Funk
Fabian Bendix
Johannes Kirchmair
Martin Biely
Eva Maria Krovat
Sharon D. Bryant
Daniela Ladstätter
Alois Dornhofer
Christian Laggner
Gökhan Ibis
Patrick Markt
Robert Kosara
Konstantin Poptodorov
Thua Huynh Buu Thi Hoang
Claudia Schieferer
Thomas Seidel
Daniela Schuster
Gerhard Wolber
Theodora Steindl
Hermann Stuppner
Rémy D. Hoffmann
Judith Rollinger
Nicolas Triballeau-Hugounencq
Caroline Bliem
Kareen Chuang ✝
T. Langer, KSSCI 2013
Acknowledgements
Markus Böhler
Yasmin Aristei
Oliver Funk
Fabian Bendix
Johannes Kirchmair
Martin Biely
Eva Maria Krovat
Sharon D. Bryant
Daniela Ladstätter
Alois Dornhofer
Christian Laggner
Gökhan Ibis
Patrick Markt
Robert Kosara
Konstantin Poptodorov
Thua Huynh Buu Thi Hoang
Claudia Schieferer
Thomas Seidel
Daniela Schuster
Gerhard Wolber
Theodora Steindl
Hermann Stuppner
Rémy D. Hoffmann
Judith Rollinger
Nicolas Triballeau-Hugounencq
Caroline Bliem
Kareen Chuang ✝
Prestwick Chemical’s Team
Outline
• Introduction
– The attrition problem in drug discovery
• Computational methods for medicinal chemists
– What are our challenges in medicinal chemistry ?
– What do we really need to make a good drug candidate ?
• Pharmacophore-based Ligand Activity Profiling
– High throughput model generation
– Parallel virtual screening & target fishing examples
Dramatic Situation in Big Pharma ?
• R&D expenditure raising
• New drug approvals decreasing
Billion $
120
100
80
60
40
2002 2003
2004 2006
2006
20
2007
2008
0
2009
2010
2011
Billion $
120
100
80
60
40
2002 2003
2004 2006
2006
20
2007
2008
0
2009
2010
2011
Billion $
A True Productivity Problem ?
• New drugs approved since 1950 ...
• 1103 Small molecules
• 119 Biologicals
• 119 Biologicals
• 119 Biologicals
Mio USD
Costs Per New Drug
Costs Per New Drug
Mio USD
Average (n=73): 4.34 billion USD (!!!)
Costs Per New Drug
Mio USD
Average (n=73): 4.34 billion USD (!!!)
An Interesting Article To Read ...
An Interesting Article To Read ...
Nat. Rev. Drug Discov., 8, 959-968 (2009)
Costs Attributed To Different Stages
Investment per phase of drug discovery and development
for one successful drug (USD millions)
G.M. Milne, Jr., Annual Reports in Medicinal Chemistry, 2003, 38, 383-396
Costs Attributed To Different Stages
Investment per phase of drug discovery and development
for one successful drug (USD millions)
G.M. Milne, Jr., Annual Reports in Medicinal Chemistry, 2003, 38, 383-396
What are we talking about today?
• How to use smart computational approaches
to support & direct medicinal chemistry efforts
• How to use smart computational approaches
to support & direct medicinal chemistry efforts
• Focus on
– Hit identification by virtual screening
– Decision support for med. chem. optimization
– Liability prediction for development
Expectations: The Medicinal Chemist
• Answer the most important question !
T. Langer,
Langer, KSSCI
2012 2013
T.
Expectations: The Medicinal Chemist
Tell me: Which
molecule shall I prepare
next ?
T. Langer,
Langer, KSSCI
2012 2013
T.
Expectations: The Biologist
T. Langer,
Langer, KSSCI
2012 2013
T.
Expectations: The Biologist
Tell me all about
the pathway & related
kinetics !
T. Langer,
Langer, KSSCI
2012 2013
T.
The Computational Chemist ...
• Wants to explain biological data
based on oversimplified models
• Wants to find the right answers
to all the medicinal chemist’s
questions
• Wants to find the right answers
to all the medicinal chemist’s
questions
• Wants to help the medicinal
chemist by designing great
molecules
DOOMED TO
?
Where to
begin ?
What Do Medicinal Chemists Really Need ?
• Reliable decision support tools
• Easy to use, easy to understand
• Need for speed
• Need for speed
• Solid science behind
Tanimoto Coefficient
Free Energy Calculations 3D QSAR
QM/MM 2D Fingerprints
Docking MEPs Shape Similarity
Pharmacophores
Self Organizing Neural Networks
Support Vector Machine QSAR
Similarity Index
CoMFA
What Is A Pharmacophore ?
“A pharmacophore is the ensemble of steric and electronic
features that is necessary to ensure the optimal supramolecular interactions with a specific biological target and to
trigger (or block) its biological response.”
C.-G. Wermuth et al., Pure Appl. Chem. 1998, 70: 1129-1143
Feature-based Pharmacophore Models
Totality of universal chemical features that represent a defined
binding mode of a ligand to a bio-molecular target
Features: Electrostatic interactions, H-bonding, aromatic interactions, hydrophobic regions, coordination to metal ions ...
The Usual Virtual Screening Protocol
10x molecules against one target
results in a hit list
Why Not Do This ?
Why Not Do This ?
x
10 molecules
against
x
10 targets
... needs a large number of models !
Our Aim: Predict Activity Pattern ...
• Modeling of all relevant targets
– responsible for drug action and side effects
– build feature-based pharmacophore models
• Compile all models (+ relevant info) into a database
– Activity profiling of leads / drug candidates
– Determination of side effects / bio-hazards
• Use this system for development of novel
interesting lead molecules and drug candidates
Why Use Pharmacophore Models ?
• Universal
- Pharmacophore models represent chemical functions that are
essential for binding to the target
• Computationally Efficient
- Due to their simplicity, they are suitable for large scale virtual
screening (>109 compounds, also in parallel settings)
• Comprehensive & Editable
- Selectivity-tuning by adding or omitting chemical feature
constraints, information can be easily traced back
- Ideal interface between modeling and medicinal chemistry
How To Build Pharmacophore Models ?
• Starting from ligand information
– Exploration of conformational space
– Multiple superimpositioning experiments
– DISCO, Catalyst, Phase, MOE, Galahad, LigandScout ...
• Starting from 3D target information
– GRID interaction fields: Convert regions of high interaction
energy into pharmacophore point locations & constraints
[S. Alcaro et al., Bioinformatics 22, 1456-1463, 2006]
– Start from target-ligand complex: Convert interaction
pattern into pharmacophore point locations & constraints
[G. Wolber et al., J. Chem. Inf. Model. 45, 160-169, 2005]
How To Build Pharmacophore Models ?
• Starting from ligand information
– Exploration of conformational space
– Multiple superimpositioning experiments
– DISCO, Catalyst, Phase, MOE, Galahad, LigandScout ...
• Starting from 3D target information
– GRID interaction fields: Convert regions of high interaction
energy into pharmacophore point locations & constraints
[S. Alcaro et al., Bioinformatics 22, 1456-1463, 2006]
– Start from target-ligand complex: Convert interaction
pattern into pharmacophore point locations & constraints
[G. Wolber et al., J. Chem. Inf. Model. 45, 160-169, 2005]
Let’s have a look ...
LigandScout 1.0: Implemented Procedure
•
Detect ligand and clean-up the binding site in the protein
(all amino acids within 7Å distance from the ligand)
•
•
Interpret hybridization status and bond types in the ligand
•
•
Search for corresponding chemical features of the protein
•
Add excluded volume spheres for opposite hydrophobic features
Perform chemical feature recognition for the ligand
(H-bond donor, H-bond acceptor, positive ionizable, negative
ionizable, hydrophobic, aromatic ring, metal ion coordination)
Add interaction features to the model only if a corresponding
feature pair is found in the complex
G. Wolber, T. Langer: J. Chem. Inf. Model. 45 , 160-169 (2005)
LigandScout 3.1 Graphical User Interface
LigandScout 3.1 Graphical User Interface
vHTS: We Need Speed & Accuracy
• Design a new alignment algorithm
• No upper limit for number of features acceptable
– high number of features will give good selectivity
• No exponential growth of search time with
growing number of features
• No graph matching necessary ...
vHTS: We Need Speed & Accuracy
• Design a new alignment algorithm
• No upper limit for number of features acceptable
– high number of features will give good selectivity
• No exponential growth of search time with
growing number of features
• No graph matching necessary ...
Work exclusively in the pharmacophore domain !
Alignment By Pharmacophore Points
Methotrexate
Dihydrofolate
Methotrexate
Dihydrofolate
Methotrexate
Dihydrofolate
Methotrexate
Dihydrofolate
Methotrexate
Dihydrofolate
Wrong
Correct
Methotrexate
Dihydrofolate
Wrong
Correct
Bystroff, Oatley, Kraut: Biochemistry. (1990) 29, 3263-77
Böhm, Klebe, Kubinyi: Wirkstoffdesign (1999) p. 320f
1RX2
1RB3
1RX2
1RB3
1RX2
1RB3
Pharmacophoric Alignment
molecule
best
3D rotation
pharmacophore
pairing
(Kabsch)
Superposition
Is pairing valid?
If not, remove invalid pairs and retry
Wolber G. et al., J. Comput.-Aided Mol. Des. 20: 773-388 (2006)
How To Find The Best Pairs ...
Hungarian Matcher (Marrying Problem)
Hungarian Matcher (Marrying Problem)
[Edmods 1965]
[Kuhn 1955]
[Richmond 2004]
Matching and a Polyhedron with 0-1 Vertices. J. Res. NBS 69B (1965), 125-30
[nonbipartite application]
The Hungarian method for the Assignment Problem.
Noval Research Quarterly, 2 (1955) [bipartite variant]
Application to chemistry: N. Richmond et al. Alignment of 3D molecules
using an image recognition algorithm. J. Mol. Graph. Model. 23 (2004) 199-209
Hungarian Matching
• How to define the pharmacophore feature
matching cost (similarity)?
– Use only few feature types
– Calculate similarity by defining geometric relations
=> Solution: Encode geometry and nature of the
neighborhood into each feature description
Typed Distance Shells
Acceptor
Donor
Lipophilic
Acceptor
Donor
Lipophilic
0|0|1
Acceptor
0|0|1
Donor
Lipophilic
0|1|1
Acceptor
0|0|1
Donor
Lipophilic
0|1|1
0|1|1
Acceptor
0|0|1
Donor
Lipophilic
0|1|1
0|1|1
Acceptor
0|0|1
Donor
Lipophilic
0|1|1
0|1|1
Distance Characteristics
Result: Best matching pairs for each feature
1
1
1
2
1
2
4
1
5
5
3
2
1
3
4
2
4
1
5
5
3
2
1
3
4
2
Final step: 3D rotation using Kabsch algorithm
Flexible Alignment
• Generation of conformer ensemble (OMEGA 2.0)
• Alignment experiment on bio-active conformation
1ke8
1rb3
1xp0
Wolber G. et al., J. Comput.-Aided Mol. Des. 20: 773-388 (2006)
vHTS Screening Strategies
• Align everything?
- 5 Mio comp. * 50 conf. * 50 ms= 144 CPU days -> too long
• Fingerprint only
- quick & dirty, loss of resolution & loss of predictivity
• Safe early failure strategy
- Bloom filtering based on fingerprints
- Accurate fitting of remaining compounds
T. Langer,
Langer, KSSCI
2012 2013
T.
Virtual Screening
Compound 1
Early
fail filter
Early
fail filter
Early
fail filter
3D alignment
Compound 2
Early
fail filter
Early
fail filter
Early
fail filter
3D alignment
Compound 3
Early
fail filter
Early
fail filter
Early
fail filter
3D alignment
Compound n
Early
fail filter
Early
fail filter
Early
fail filter
3D alignment
....
Fingerprinting strategy
• Pre-store conformations
• Pre-compute and store as much as
possible (2pt, 3pt)
• Store fingerprints for each
conformation and for the whole
molecule
...
T. Langer,
Langer, KSSCI
2012 2013
T.
Multiconformational Database Size
• 1 Mio compounds ca 7.5 Gb
• 13kB per compound or 273 byte per conformation
• 75% of the space is needed for fingerprints
• Overall very compact
What Did We Reach With LigandScout ?
• Higher accuracy than any other pharmacophorebased screening tool
• Higher speed than any other pharmacophorebased screening tool
• Higher speed than any other pharmacophorebased screening tool
• Let’s have a look …
Performance Analysis Using ROC Curves
T. Langer,
Langer, KSSCI
2012 2013
T.
T. Langer,
Langer, KSSCI
2012 2013
T.
T. Langer,
Langer, KSSCI
2012 2013
T.
T. Langer,
Langer, KSSCI
2012 2013
T.
T. Langer,
Langer, KSSCI
2012 2013
T.
T. Langer,
Langer, KSSCI
2012 2013
T.
T. Langer,
Langer, KSSCI
2012 2013
T.
T. Langer,
Langer, KSSCI
2012 2013
T.
T. Langer,
Langer, KSSCI
2012 2013
T.
Screening Metrics: PDE5
TP
FP
(of 51)
(of 1803)
Catalyst
21
239
2.94 0.632
Phase
10
74
4.33 0.560
MOE
29
1473
0.70 0.356
LigandScout
10
0
36.4 0.592
TP ... true positives
FP ... false positives
EF ... enrichment factor (TP/(TP+FP))/(A/(A+I))
AUC ... area under the (ROC) curve
T. Langer,
Langer, KSSCI
2012 2013
T.
EF
AUC
Ligand-based: Amine-binding GPCRs
• Study published by Evers et al. @ Sanofi-Aventis J. Med.
Chem. 48, 5448-65, (2005)
• Compared different VS techniques for
– Alpha1A receptor antagonists
– 5HT2A receptor antagonists
– D2 receptor antagonists
– M1 receptor antagonists
T. Langer,
Langer, KSSCI
2012 2013
T.
Amine-binding GPCRs
EF1%
5HT2A:
A1AA:
D2:
M1::
LigandScout
Phase
Catalyst
MOE
14
10
10
10
LigandScout
Phase
Catalyst
MOE
16
16
14
4
LigandScout
Phase
Catalyst 0
MOE
10
6
4
LigandScout
Phase
Catalyst
MOE
18
8
16
6
0
5
T. Langer,
Langer, KSSCI
2012 2013
T.
10
15
20
True and False Positives
LigandScout
100%
75%
50%
25%
0%
5HT2A
M1
D2
A1AA
True and False Positives
Phase
LigandScout
100%
100%
75%
75%
50%
50%
25%
25%
0%
0%
5HT2A
M1
D2
5HT2A
A1AA
M1
100%
100%
75%
75%
50%
50%
25%
25%
0%
0%
M1
A1AA
D2
A1AA
MOE
Catalyst
5HT2A
D2
D2
A1AA
5HT2A
M1
Average True Positive Rate
T. Langer,
Langer, KSSCI
2012 2013
T.
First Summary
• Universal method for accurate feature-based
pharmacophore generation
• Highly selective models will retrieve low number
of false positives / false negatives
• High enrichment factor can be obtained
• Application for in silico ligand profiling and
target fishing possible ?
Ligand Profiling Case Study
- 5 viral targets
- 50 pharmacophore models
- 100 antiviral compounds
Will their activity profiles be predicted correctly ?
Ligand Profiling: Targets
Target
Disease
Function
Mechanism
HIV protease
HIV infection,
Cleavage of gag and gag-pol
Inhibition at active site
AIDS
precursor polyproteins into
functional viral proteins
HIV reverse transcriptase HIV infection,
Synthesis of a virion DNA,
(RT)
integration into host DNA and
AIDS
Inhibition at allosteric site
transcription
Influenza virus
Influenza
neuraminidase (NA)
Viral envelope glycoprotein,
Inhibition at active site
cleave sialic acid residues for viral
release
Human rhinovirus (HRV)
Common cold
coat protein
Hepatitis C virus (HCV)
RNA polymerase
Hepatitis C
Attachment to host cell receptor,
Binding in hydrophobic pocket
viral entry, and uncoating
(capsid stabilization)
Viral replication, transcription of
Inhibition at various allosteric
genomic RNA
sites
Results Matrix
T. Steindl et al., J. Chem. Inf. Model., 46, 2146-2157 (2006)
Ligand-directed Analysis
Ratio ≥ 1
90% of the compounds
correctly predicted
Ratio < 1
8% more often
predicted for one
specific false target
than for correct one
for 2% of the
compounds no activity
prediction possible
Pharmacophore-directed Analysis
HIV protease
HIV RT
Influenza NA
HIV protease
HIV RT
Influenza NA
HRV coat
protein
HCV
polymerase
123
HRV coat protein
HCV polymerase 1 2 3
Pharmacophore-directed Analysis
HIV protease
HIV RT
Influenza NA
HRV coat protein
HCV polymerase 1 2 3
HIV protease
HIV RT
Influenza NA
HRV coat
protein
HCV
polymerase
123
Model with lowest selectivity:
70% of actives (HIV RT), but 75% from one specific
false target (HRV coat protein)
40% active and 60% inactive compounds in hit list
Models retrieving no false positives
First Conclusion ...
• First published examples of applications of extensive
parallel screening approach based on pharmacophores
• Multitude of pharmacophore models (up to several thousand …)
• Large set of molecules (up to several million …)
• Results indicate
• Correct assignment of selectivity in most cases
• Independent of search algorithms used
• Fast, scalable in silico activity profiling is now possible !
First Conclusion ...
• First published examples of applications of extensive
parallel screening approach based on pharmacophores
• Multitude of pharmacophore models (up to several thousand …)
• Large set of molecules (up to several million …)
• Results indicate
• Correct assignment of selectivity in most cases
• Independent of search algorithms used
• Fast, scalable in silico activity profiling is now possible !
Inte:Ligand’s Pharmacophore Database
~ 320 unique targets ready to use*
• Represented in
~ 210 ligand-based pharmacophore models
~ 2500 structure-based pharmacophore models
• Covering a selection of all major therapeutic classes
• Contains anti-target models for finding adverse effects
• Categorized according to the pharmacological target
* out of ~650 categorized by August 2013
Other Application Examples
• Pharmacophores for
– drug repositioning
– target fishing for natural products
– predicting ligands for protein-protein interfaces
– in silico fragment-based screening
Drug Repositioning
LigandScout Model of ABCG2 Inhibitors
Yinxiang Wei, Yuanfang Ma, Qing Zhao, et al., Mol Cancer Ther 2012, 11, 1693-1702.
Inhibiting ABCG2 With Sorafenib
… at a concentration up to 2,5µM/L
no cytotoxic effect was observed ...
…. led us to conclude that sorafenib
behaves like ABCG2 degradationinduced inhibitor. Sorafenib may,
therefore, be a good candidate for
MDR chemosensitizing agent.
Yinxiang Wei, Yuanfang Ma, Qing Zhao, et al., Mol Cancer Ther 2012, 11, 1693-1702.
Target Fishing With Natural Products
Small molecule new chemical entities organized by source/year (N =974).
David J. Newman, and Gordon M. Cragg, J. Nat. Prod., 2007, 70, 461-477
• Target identification for leoligin, the
major lignan from the plant Edelweiss
(Leontopodium alpinum Cass.),
• Construction of a 3D multiconformational
molecular structure of this compound
• Pharmacophore-based ligand affinity profiling
• Isolation and biological testing
K. Duwensee, et al.,
Atherosclerosis (2011) 219, 109-115
Pharmacophore Profiling of Leoligin
Leoligin matches the pharmacophore model encoding for the
interaction site of cholesteryl ester transfer protein (CETP)
Biological Testing
Leoligin enhances the activity of
human and rabbit CETP in vitro
when applied in subnanomolar
Leoligin activates CETP in vivo
(7 days test with CETP
transgenic mice, leoligin dosed
orally, 1 mg/kg)
concentration
(control compound: Probucol)
Pharmacophores To Identify PPIs
A. Voet et al., Curr Pharm Design, 2013, May 29
L. De Luca et al., ChemMedChem, 2009, 4, 1311 – 1316
V. Corradi et al., BMCL, 2010, 20, 6133 – 6137
A Recent Prestwick Success Story
Protein Protein Interfaces
Interesting targets for drug discovery ?
– Problem: Large contact surfaces with limited number of hot
spots for specific interactions
– Can small molecules disturb such an interface ?
Protein Protein Interfaces
Interesting targets for drug discovery ?
– Problem: Large contact surfaces with limited number of hot
spots for specific interactions
– Can small molecules disturb such an interface ?
<
Dömling, A. Curr. Opin. Chem. Biol. 2008, 12, 281-291.
Described p53/mdm2 Inhibitors
Dömling, A. Curr. Opin. Chem. Biol. 2008, 12, 281-291.
Structure-based Pharmacophore
• p53/mdm2 complex structure (PDB: 1YCR)
Comparison with JNJ-26854165
Pharmacophores for FBDD ?
A Recent FBDD Success Story
FBDD Strategy @ Prestwick
• Use set of smart, recombinable fragments
• Perform pharmacophore-based screening
• Recombine fragments in silico
• Synthesize the highest ranked solutions
– IP situation
– Fit for the target
– Chemical tractability
– Physicochemical properties
The Prestwick Chemical Fragments
• Privileged structure concept
• Smart fragmentation of drug molecules
FBD Strategy @ Prestwick
Months
1
2
3
4
5
6
PhD 1 (CADD)
PhD 2
Technician
7
8
9
10
11
12
13
Months
1
2
3
4
5
6
PhD 1 (CADD)
PhD 2
Technician
7
8
9
10
11
12
13
1
Months
2
3
4
5
6
PhD 1 (CADD)
PhD 2
Technician
d
e
er
iv
t
is
l
t
l
de
hi
7
8
9
10
11
12
13
1
Months
2
3
4
5
6
PhD 1 (CADD)
PhD 2
Technician
er
liv
ed
siz
he
fir
st
hi
ts
sy
nt
hi
&
t
is
l
t
de
iv
l
de
ed
d
e
er
7
8
9
10
11
12
13
1
Months
2
3
4
5
6
PhD 1 (CADD)
PhD 2
Technician
ity
fin
er
af
µM
ow
:l
ed
lid
at
nt
va
sy
ts
fir
s
th
its
hi
st
fir
liv
ed
siz
he
hi
&
t
is
l
t
de
iv
l
de
ed
d
e
er
7
8
9
10
11
12
13
1
Months
2
3
4
5
6
PhD 1 (CADD)
PhD 2
Technician
ity
fin
er
af
µM
ow
:l
ed
lid
at
nt
va
sy
ts
fir
s
th
its
hi
st
fir
liv
ed
siz
he
hi
&
t
is
l
t
de
iv
l
de
ed
d
e
er
hit to lead
7
8
9
10
11
12
13
1
Months
2
3
4
5
6
7
8
9
10
PhD 1 (CADD)
PhD 2
Technician
ad
ity
in
at
io
µM
n
af
of
le
fin
er
m
ow
:l
no
ed
lid
at
nt
va
sy
ts
fir
s
th
its
hi
st
fir
liv
ed
siz
he
hi
&
t
is
l
t
de
iv
l
de
ed
d
e
er
hit to lead
lead optimization
11
12
13
1
Months
2
3
4
5
6
7
8
9
10
11
12
PhD 1 (CADD)
PhD 2
Technician
ad
ity
in
at
io
µM
n
af
of
le
fin
er
m
ow
:l
no
ed
lid
at
nt
va
sy
ts
First hits had already novel structures, were patentable
fir
s
th
its
hi
st
fir
liv
ed
siz
he
hi
&
t
is
l
t
de
iv
l
de
ed
d
e
er
hit to lead
lead optimization
13
Fragment Screening in LigandScout
• Normal virtual screening mode
– Match all requested query features
– Useful for ‘fragment linking’ approach
• Fragment screening mode
– Match a defined minimum number of features
– Optimized algorithm for partial matching
– Useful for ‘growing fragments’ strategy
Fragment Screening: ‘Linking’
Fragment Screening: ‘Growing’
Conclusions
• Pharmacophore-based virtual screening and ligand
activity profiling are straightforward and rapid
methods for the identification of new and better
lead compounds
• Combined with informatics-based molecular
building tools, optimized design of novel and
promising compounds becomes feasible
• Assessment of risks in later development stages
becomes possible on a rational & transparent basis
Thank you for your attention
Thank you for your attention
www.prestwickchemical.com
www.inteligand.com

Fragment Screening in LigandScout

Transcription

Similar documents

Press release

January 2015 - Boulder County Bar Association

The Logical and the Empirical Form of Feeling

Beating the market, and the the drum

Check Up From The Neck Up

Thursday, April 24 @ 7:30pm AMC BarryWoods 24

mti photoscreener - Ruidoso Valley Noon Lions Club

February 8th at 12:45 pm CINEPLANET 16

segmentation of cytological images using color and