Speaker
Dr
Ying-Ta Wu
(Academia Sinica Genomic Research Center)
Description
The potential for re-emergence of influenza pandemics has been a great threat since
the report of that the avian influenza A virus (H5N1) having acquired the ability to
be transmitted to humans. An increase of transmission incidents suggests the risk of
human-to-human transmission, and the report of development of drug resistance
variants is another potential concern. At present, there are two effective antiviral
drugs available, oseltamivir (Tamiflu) and zanamivir (Relenza). Both drugs were
discovered through structure-based drug design targeting influenza neuraminidase
(NA), a viral enzyme that cleaves terminal sialic acid residue from glycoconjugates.
The action of NA is essential for virus proliferation and infectivity; therefore,
blocking the actives would generate antivirus effects. To minimize non-productive
trial-and-error approaches and to accelerate the discovery of novel potent
inhibitors, medicinal chemists can take advantage of using modeled NA variant
structures and doing structure-based design.
A key work in structure-based design is to model complexes of candidate compounds to
structures of receptor binding sites. The computational tools for the work are based
on docking tools, such as AutoDock, to carry out quick conformation search of small
compounds in the binding sites, fast calculation of binding energies of possible
binding poses, prompt selection for the probable binding modes, and precise ranking
and filtering for good binders. Although docking tools can be run automatically, one
should control the dynamic conformation of the macromolecular binding site (rigid or
flexible) and the spectrum of the screening small organics (building blocks and/or
scaffolds; natural and/or synthetic compounds, diversified and/or “drug-like”
filtered libraries). This process is characterized by computational and storage load
which pose a great challenge to resources that a single institute can afford (For
example, using AutoDock to evaluate one compound structure for 10 poses within the
target enzyme would take 200 Kilobyte storage and 15 minutes on an average PC). The
task to evaluate 1 million compound structures 100 poses each would cost 2 Terabyte
and more than hundred years). To support such kind of computing demands, this project
was initiated to develop a service prototype for distributing huge amount of
computational docking requests by taking the advantages of the LCG/EGEE Grid
infrastructure.
According to what we have learned from both the High-Energy Physics experiments and
the Biomedical community, an effective use of large scale computing offered by the
Grid is very promising but calls for a robust infrastructure and careful preparation.
Important points are the distributed job handling, data collection and error
tracking: in many cases this might be a limitation due to the need of grid-expert
personnel effort. Our final goal is to deliver an effective service to academic
researchers who for the most part are not Grid experts, therefore we adopted a
light-weight and easy-to-use framework for distributing docking jobs on the Grid. We
expect that this decision will benefit future deployment efforts and improve
application usability.
Introducing the DIANE framework in building the service is aimed at handling the Grid
applications in master-worker model, a native computing model of distributing docking
jobs on the Grid. With the skeletal parallelism, applications plugged into the
framework inherit the intrinsic DIANE features of distributed job handling such as
automatic load balancing, and failure recovery. The python-based implementation also
lowers the development effort of controlling application jobs on the Grid. With the
hiding of composing JDL and of submitting jobs, users can even easily distribute
their application jobs on the Grid without having Grid knowledge. In addition, this
system can be used to seamlessly merge local guaranteed resources (like a dedicated
cluster) with on-demand power provided by the Grid, allowing researches to
concentrate on setting up of their application without facing a heavy entry barrier
to move in production mode where more resources are needed.
In a preliminary study, we arranged the work into six tasks: (1) target 3D structure
preparation; (2) compound 3D structure preparation and refinement, (3) compound
properties and filter, (4) Autodock run (5) probable hits analysis and selection, and
(6) complex optimization and affinity re-calculation. The DIANE framework has been
applied to distribute about 75000 time-consuming AutoDock processes on LCG for
screening possible inhibitor candidates against neuraminidases. In addition to show
the distribution efficiency, advantages of adopting DIANE framework in the AutoDock
application are also discussed in terms of usability, stability and scalability.
Authors
Mr
Hurng-Chun Lee
(Academia Sinica Computing Center)
Dr
Ying-Ta Wu
(Academia Sinica Genomic Research Center)
Co-authors
Mr
Hsin-Yen Chen
(Academia Sinica Computing Center)
Mr
Li-Yung Ho
(Academia Sinica Grid Computing Center)
