Transcript The folding problem

Massively Distributed Computing
and
An NRPGM Project
on
Protein Structure and Function
Computation Biology Lab
Physics Dept & Life Science Dept
National Central University
From Gene to Protein
About Protein
• Function
– Storage, Transport, Messengers, Regulation…
Everything that sustains life
– Structure: shell, silk, spider-silk, etc.
• Structure
– String of amino acid with 3D structure
– Homology and Topology
• Importance
– Science, Health & Medicine
– Industry – enzyme, detergent, etc.
• An example – 3hvt.pdb
Protein Structure & Function
• Primary sequence  Native state
with 3D structure
– Structure
function
– Expensive and time consuming
• Misfolding means malfunction
– Mad cow disease (“prion” misfolds)
The Folding Problem
• Complexity of mechanism & pathway is
huge challenge to science and
computation technology
Molecular Dynamics (MD)
• Molecular’s behavior determined by
– Ensemble statistics
– Newtonian mechanics
• Experiment in silico
• All-atom w. water
– Huge number of particles
• Super-heavyduty computation
• Software for macromolecular MD
available
– CHARMm, AMBER, GROMACS
Simple Statistics on
MD Simulation
• Atoms in a typical protein and water
simulation
32000
• Approximate number of interactions in force
calculation
109
• Machine instructions per force calculation 1000
• Total number of machine instructions
1023
• Typical time-step size
10–15 s
• Number of MD time steps
1011 steps
• Physical time for simulation
10–4 s
• Total calculation time (CPU: P4-3.0G ) days 10,000
Protein Studies by
Massively Distributed Computing
A Project in National Research Program on Genomic
Medicine
• Scientific
– Protein folding, structure, function, proteinmolecule interaction
– Algorithm, force-field
• Computing
– Massive distributive computing
• Education
– Everyone and Anyone with a personal PC
can take part
• Industry – collaborative development
Distributed Computing
• Concept
– Computation through internet
– Utilize idle PC power (through screen-saver)
• Advantage
– Inexpensive way to acquire huge computation power
– Perfectly suited to task
• Huge number of runs needed to beat statistics
• Parallel computation not ALWAYS needed
• Massive data - good management necessary
• Public education – anyone w/ PC can take
part
Hardware Strategies
• Parallel computation (we are not this)
– PC cluster
– IBM (The blue gene), 106 CPU
• Massive distributive computing
– Grid computing (formal and in the future)
– Server to individual client (now in inexpensive)
• Examples: SETI, folding@home, genome@home
• Our project: protein@CBL
Software Components
• Dynamics of macromolecules
– Molecular dynamics, all atomistic or meanfield solvent
– Computer codes
• GROMACS (for distributive comp; freeware)
• AMBER and others (for in-house comp; licensed)
• Distributed Computing
– COSM - a stable, reliable, and secure system
for large scale distributed processing
(freeware)
COSM’s Structure
Client
Server
System tests
System test
(test all Cosm functions)
Self-tests
Open Multithread
( Working Channel)
Connect to server
Connect to client
Send Request
Packet Request
Recv Request
Recv Assignment
Packet Assignment
Send Assignment
Put Result
Packet Result
Get Result
Get Accept
Packet Accept
Put Accept
Running Simulation
Structure at Server end
Protein
database
•Temporary
databank
•Job analysis
•Automatic
temperature
swaps by
parallel tempering
Databank
Jobs
Send(COSM)
Receive
Human
intervention
Exceptions
clients
Structure at Client end
Server
Receive
MD Run
Return result
Delete files
If crash
Restart
Multi-temperature Annealing
• Project suited for multi-temperature runs –
Parallel Tempering
• Two configurations with energy and temperature
(E1, T1) and (E2, T2)
Temperature swapped with probability
P = min{1, exp[-(E2-E1)(1/kT1 – 1/kT2)]}
• Mode of operation
– Send same peptide at different temperature to many
clients; let run; collect; swap T’s by multiple parallel
tempering; randomly redistribute peptides with new T’s
to clients
Databank
Multi-temperature Annealing
Server
Old temperatures
client
Swap temps by
client
Multiple “peptide”
parallel
tempering
client
New temperatures
client
client
client
client
Potential of
Massive Distributive Computing
• Simulation of folding a small peptide for 100ns
– Each run (105 simulation steps; 100 ps) ~100 min PC time
– 1000 runs (100 ns) per “fold”
~105 min
– Approx. 70 days on single PC running 24h/day
• Ideal client contribute 8h/day
– 100 clients
– 10,000 clients
70x3/100 = 2 days per fold
50 folds/day (small peptide)
• Mid-sized protein needs > 1 ms to fold
– 7x105 days on single PC
– 10,000 clients
210 days
– 106 clients (!!)
2~3 days
Learning curve
• Launched –August 2002
• Small PC-cluster – October 2002
– In-house runs to learn codes
• Infrastructure for Distributive Computation
– Installation Gromacs & COSM – January-March 2003
• Test runs and debugging
– IntraLaboratory test run – March-October 2003
– NCU test run – July-October 2003
• Launched on WWW – 20 November 2003
– Traffic jam – multiple server (see next slide)
• Scientific studies > November 2003
In-House Test Runs
• Time – Began March 2003
• Clients
– About 25 PCs in CBL and outsiders ( MSWindow )
• Goal – test and debug
– Test server-client communication
• Lots of debugging
– Test data distribution, collection and
management
– Test parallel tempering
Multi-Server Architecture
Client
Font-End
Client
Server
Backend
Server 1
Client
Step 1: Client sends request to Front-End Server
Step 2: FES assigns IP of a Back-End Server i to client
Step 3: Client requests job from BESi
Step 4: BESi sends job to client
Step 5: Client sends result to BESi
Repeat cycle.
Backend
Server N
Current status and Plans for
immediate future
• Last beta version Pac v0.9
– Released on July 15
– To lab CBL members & physics dept
– About 25 clients
• First alpha version Pac v1.0 released October 1 2003
• Current version Pac v1.2
– Releases for distributed computing on 20 November
2003
– In search of clients
• Portal in “Educities” http://www.educities.edu.tw/
~2,500 downloads, ~500 real clients
• PC’s in university administrative units
• City halls and county government offices
• Talks and visits to universities and high schools
Current Simulations
1SOL: (20 res.)
A Pip2 and FActin-Binding
Site Of Gelsolin,
Residue 150169. One helix.
1ZDD: (35 res.)
Disulfide-Stabilized Mini
Protein A
Domain.
Two helices.
1L2Y: (20 res.)
NMR Structure
Of Trp-Cage
Miniprotein
Construct Tc5B;
synthetic.
A small test case – 1SOL
• Target peptide – 1SOL.pdb
– 20 amino acids; 3-loop helix
and 1 hairpin; 352 atoms;
~4000 bonds interaction
– Unit time step= 1 fs
• Compare constant temperature and
parallel-tempering
– Constant T @ 300K
– Parallel-tempering with about 20 peptides,
results returned to server for swapping after
each “run”, or 105 time steps (100 ps)
Parallel-tempering (1SOL)
P = min{1, exp[-(E2-E1)(1/kT1 – 1/kT2)]}
Temperature (K)
550
500
450
400
350
300
250
200
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Number of runs (in units of 100 ps)
273
285
300
315
333
348
366
384
405
426
447
471
498
Preliminary
result on
1SOL
Initial structure
Parallel-temp. (1.6ns)
Native conformation
Const temp. (20ns)
A second test case – 1L2Y
• Simulation target
– Trp-Cage
• 20 amino acids,
2 helical loops
• A short, artificial and fold-by-itself peptide
• Have been simulated with AMBER
• Folding mechanism not well understood
Swap History (1L2Y)
500
270
290
310
400
330
350
370
350
390
410
430
300
450
470
Number of runs (in units of 100 ps)
69
65
61
57
53
49
45
41
37
33
29
25
21
17
13
9
5
250
1
Temperature (K)
450
Preliminary
result on
1L2Y
(11 peptides)
Initial state
Native state
PAC 6ns
Modifications needed
• Reduce size of water box
– Save computation time
• Rewrite the energy function
– Ignore the water-water interaction
• Increase cut-off radius
• Try different simulation algorithms for
changing pressure and temperature
• Others…
Looking ahead
•
•
•
•
Better understanding of annealing procedure
Better understanding of energetics
Expand client community
Develop serious collaboration with biologists
– Structure biologists, e.g., NMR people
– Protein function people
– Drug designers
• “…investigation of motions that have
particular functional implications and to
obtain information that is not accessible to
experiment.”
Karplus and McCammon, Nature Strct. Biol. 2002
The Team
• Funded by NRPGM/NSC
• Computational Biology Laboratory
Physics Dept & Life Sciences Dept
National Central University
–
–
–
–
–
–
–
–
PI: Professor HC Lee (Phys & LS/NCU)
Co-PI: Professor Hsuen-Yi Chen (Phys/NCU)
Jia-Lin Lo (PhD student)
Jun-Ping Yiu (MSc Res. Assistant)
Chien-Hao Wei (MSc RA)
Engin Lee ( MSc student )
PDF (TBA)
We are looking for collaborators, research
associates, programmers, students, etc.
http://protein.ncu.edu.tw
Thank you for your
attention