Large Scale Virtual Screening of Drug Design on the Grid Fighting against Avian Flu

1
Large Scale Virtual Screening of Drug Design on
the GridFighting against Avian Flu

• Yun-Ta Wu and Hurng-Chun Lee
• ISGC 2006, Taiwan

2
Credit

• Docking workflow preparation
• Contact point Y.T. Wu
• E. Rovida
• P. D'Ursi
• N. Jacq
• Grid resource management
• Contact point J. Salzemann
• TWGrid H.C. Lee, H. Y. Chen
• AuverGrid E. Medernach
• EGEE Y. Legré

• Platform deployment on the Grid
• Contact point H.C. Lee, J. Salzemann
• M. Reichstadt
• N. Jacq
• Users (deputy)
• J. Salzemann (N. Jacq)
• M. Reichstadt (E. Medernach)
• L. Y. Ho (H. C. Lee)
• I. Merelli, C. Arlandini (L. Milanesi)
• J. Montagnat (T. Glatard)
• R. Mollon (C. Blanchet)
• I. Blanque (D. Segrelles)
• D. Garcia

Grid operational supports from sites and
operation centers DIANE technical supports from
CERN-ARDA group
3
Outline

• The avian flu
• EGEE biomed data challenge II
• Conclusion

4
Influenza A pandemic
NA
HA
H1N1
H1N1
H2N2
H3N2
H1N1
2006
2005
Apr 21, 2006
113 deaths/204 cases
http//www.who.int/csr/disease/avian_influenza
5
A closer look at bird flu

• The bird flu virus is named H5N1. H5 and N1
  correspond to the name of proteins
  (Hemagglutinins and Neuraminidases) on the virus
  surface.
• Neuraminidases play a major role in the virus
  multiplication
• Present drugs such as Tamiflu inhibit the action
  of neuraminidases and stop the virus
  proliferation
• The N1 protein is known to evolve into variants
  if it comes under drug stress
• To free-up medicinal chemists time to better
  response to instant and large scale threats, a
  large scale in-silico screening was set for
  initial investment of the design of new drug

6
In-silico (virtual) screening of drug design

• Computer-based in-silico screening can help to
  identify the most promising leads for biological
  tests
• systematic and productive
• reduces the cost of trail-and-error approach
• The requirement of CPU power and storage space
  increases proportional to the number of compounds
  and target receptors involved in the screening
• massive virtual screening is time consuming

7
The computing challenge of large scale in-silico
screening

• Molecular docking engine
• Autodock
• FlexX
• Problem size
• 8 predicted possible variants of Influenza A
  neuraminidase N1 as targets
• around 300 K compounds from ZINC database and a
  chemical combinatorial library
• Computing challenge (a rough measurement based on
  Xeon 2.8 GHz)
• Each docking requires 30 mins CPU time
• Required computing power in total 137 CPU years
• Storage requirement
• Each docking produces results with the size of
  130 KByte
• Required storage space in total 600 GByte (with
  1 back-up)
• To speed-up and reduce the cost to develop new
  drugs, high-throughput screening is demanded
• Thats the Grid can help !!

8
EGEE Biomed DC II objectives

• Biological goal
• finding potential compounds that can inhibit the
  activities of Influenza A neuraminidase N1
  subtype variants
• Biomedical goal
• accelerating the discovery of novel potent
  inhibitors thru minimizing non-productive
  trial-and-error approaches
• Grid goal
• aspect of massive throughput reproducing a
  grid-enabled in silico process (exercised in DC
  I) with a shorter time of preparation
• aspect of interactive feedback evaluating an
  alternative light-weight grid application
  framework (DIANE) in terms of stability,
  scalability and efficiency

9
EGEE Biomed DC II grid resources

• AuverGrid
• BioinfoGrid
• EGEE-II
• Embrace
• TWGrid
• a world-wide infrastructure providing over than
  5,000 CPUs

10
EGEE biomed DC II current status

• The first DC job was submitted at 10 Apr, 2006
• It is scheduled to be finished in the mid of May
• As of today, we have completed 1500K dockings
• 60 of the whole challenge (i.e. 82 CPU years)
• Grid efficiency 80

11
EGEE Biomed DCII the Grid tools

• WISDOM
• has succeeded to handle the first EGEE biomed DC
• a workflow of Grid job handling automated job
  submission, status check and report, error
  recovery
• push model job scheduling
• batch mode job handling
• DIANE
• a framework for applications with master-worker
  model
• pull mode job scheduling
• interactive mode job handling with flexible
  failure recovery feature
• we will focus on this framework in the following
  discussions

12
The WISDOM workflow in DC2

• Developed for 1st data challenge fighting against
  Malaria
• 40 millian dockings (80 CPU years) were done in 6
  weeks
• 1700 CPUs in 15 countries were used
  simultaneously
• Reproducing a grid-enabled in silico processing
  with a shorter time of preparation (lt 1 month
  preparation time has been achieved)
• Testing new submission strategy to improve the
  Grid efficiency

Use AutoDock in DC2
http//wisdom.eu-egee.fr
13
The DIANE framework

• DIANE Distributed Analysis Environment
• A lightweight framework for parallel scientific
  applications in master-worker model
• ideal for applications without communications
  between parallel tasks (e.g. for most of the
  Bioinformatics applications in analyzing huge
  amount of independent dataset)
• The framework takes care of all synchronization,
  communication and workflow management details on
  behalf of application

http//cern.ch/diane
14
The DIANE Autodock adapter for DC2
15
The DIANE exercise in DC2

• Taking care of the dockings of 1 variant
• the mission is to complete 300 K dockings
• Taking a small subset of the resources
• the mission is to handle several hundred
  concurrent DIANE workers by one DIANE master for
  a long period
• Testing the stability of the framework
• Evaluating the deployment efforts and the
  usability of the framework
• Demonstrating efficient computing resource
  integration and usage

16
Statistics of one of the DIANE runs

• Submitted Grid jobs 300
• Healthy jobs 261 (87)
• Total number of dockings 40210
• Total CPU time 55684848 sec (1.76 year)
• Job duration 249746 sec (2.9 days)

9.24 CPU years ? 250 CPUs x two week
17
Development and deployment efforts of DIANE

• Development efforts
• The Autodock adapter for DC2 is around 500 lines
  of python codes

• Deployment efforts
• The DIANE framework and Autodock adaptor are
  installed on-the-fly on the Grid nodes
• Targets and compound databases can be prepared on
  the UI or pre-stored on the Grid storages
• Output are returned to the UI interactively

18
Intuitive user interface of DIANE

• Start the DIANE job and allocate 64 workers from
  LCG and local cluster
• Allocate more workers from LCG if resources are
  available

diane.startjob job autodock.job ganga w
32_at_lcg,32_at_pbs
diane.ganga.submitworkers job autodock.job
nw100 bklcg

• -- python --
• Application 'Autodock'
• JobInitData 'macro_repos' file///home/hclee/
  diane_demo/autodock/macro',
• 'ligand_repos'file///home/hclee/
  diane_demo/autodock/ligand',
• 'ligand_list''/home/hclee/diane_de
  mo/biomed_dc2/ligand/ligands.list',
• 'dpf_parafile''/home/hclee/diane_d
  emo/biomed_dc2/parameters/dpf3gen.awk',
• 'output_prefix''autodock_test'
• The input files will be staged in to workers
• InputFiles JobInitDatadpf_parafile

19
The profile of DIANE job
good load balance
A simple test on local cluster
a DIANE/Autodock Task 1 docking
20
The profile of realistic DIANE job

• Each horizontal line segment one task one
  docking
• Unhealthy workers are removed from the worker
  list
• Failed tasks are rescheduled to healthy workers

21
Efficiency and throughput of DIANE

• 280 DIANE worker agents were submitted as LCG
  jobs
• 200 jobs (71) were healthy
• 16 failures related to middleware errors
• 12 failures related to application errors

22
Logging and bookkeeping Thanks to GANGA
In 1 print jobsDIANE_6
Statistics 325 jobs slice("DIANE_6") -----------
--- id status name subjobs
application backend
backend.actualCE 1610 running
DIANE_6 Executable
LCG melon.ngpp.ngp.org.sg2119/jobmanager-lcgpbs-
1611 running DIANE_6
Executable LCG node001.grid.auth.gr
2119/jobmanager-lcgpbs-b 1612 running
DIANE_6 Executable
LCG polgrid1.in2p3.fr2119/jobmanager-lcgpbs-biom
1613 failed DIANE_6
Executable LCG polgrid1.in2p3.fr21
19/jobmanager-lcgpbs-sdj 1614 submitted
DIANE_6 Executable
LCG ce01.ariagni.hellasgrid.gr2119/jobmanager-pb
1615 running DIANE_6
Executable LCG
ce01.pic.es2119/jobmanager-lcgpbs-biomed 1616
running DIANE_6 Executable
LCG ce01.tier2.hep.manchester.ac.uk
2119/jobmanag 1617 running DIANE_6
Executable LCG
clrlcgce03.in2p3.fr2119/jobmanager-lcgpbs-bi

• Helpful for tracing the execution progress and
  Grid job errors
• Fairly easy to visualize the job statistics

23

• The in-silico screening provides not only the
  docking poses of a compound against the target
  but also the docking energy
• By ranking the information, chemist can select
  the promising compounds to go on the
  structure-based drug design for potential drugs

24
Conclusion

• From biological point of view
• We managed to shorten the molecular docking
  process of structure-based drug design from 137
  year to 4 weeks
• A large set of complexes has been produced on the
  Grid for further analysis
• From Grid point of view
• The DC has demonstrated that large-scale
  scientific challenge can be effortlessly tackled
  on the Grid
• The WISDOM system has successfully reproduced the
  massive throughput of in-silico screening with
  minimized deployment effort
• The DIANE framework which can take control of
  Grid failures and isolate Grid system latency
  does benefit the Grid application in terms of
  efficiency, stability and usability
• Moving toward a service
• Stability and reliability of Grid has been tested
  through the DC activity and the result encourages
  the movement from prototype to real service
• Friendly graphic user interfaces for up-coming
  analysis among the large set of outputs is needed

25
Thank you for your attention!!