looking for a Ph.D. position in Protein Folding

Eugene Leitl eugene.leitl at lrz.uni-muenchen.de
Sat Oct 7 01:34:45 PDT 2000

Dear CCLers/Beowulfers,

I'm looking for a Ph.D. position in brute force MD approach to the
Protein Folding Problem, especially using massive parallelism (spatial 
decomposition of the simulation box over a large number (10^3..10^6)
of computational nodes) and novel algorithms (integer gases for
forcefield rendering, including long-range forces).

Some of the research topics I intend to focus on:

 * Development of algorithms for efficient search of configuration

 * Simulation of long-time scale events

 * Testing and improvement of molecular models and force fields by
   comparison of simulation results to experimental data

 * Simplification of force fields

 * application of the above to protein engineering and de novo design

I think that integer lattice gas algorithms simultaneously fill
several of these slots, since scaling in ~O(const) with number of 3d
lattice locally interconnected Beowulf-type nodes while being message
passing matrix latency tolerant. They can also profit from SIMD in a
register (MMX-type) parallelism of recent CPU architectures. Also they
tend to reduce cache misses both for code and data and spend most of
the time streaming through memory in sequential order, thus profiting
from burst mode of modern SDRAM memories.

This performance advantage should allow to investigate mesoscale or
long time scale phenomena on Beowulfish COTS-type hardware without the
Blue Gene price tag. (Of course, such algorithms can also profit from
massively parallel custom hardware).

Lattice gas data structures are suitable for realtime volume
visualization via the voxel paradigm [1], allowing to create fully
interactive simulations (dynamically adding constraints) and are very
useful for identifying and interactive segmentation of interesting
features in a potentially very large volume simulation box [2].

Apart from visualization and speedup, integer lattice gases appear
useful for fitting forcefields to empirical data (using evolutionary
algorithms to select the forcefield which can restore native fold of
randomly distorted Brookhaven database structures from a population of
forcefield individua). Learning from QM results should work via the
same route.

And these algorithms are, well, fundamentally simple, since shifting
forcefield complexity from code to the number arrays in the
(interpolating) lookup tables. This also makes them GA-friendly.

Since those early lattice gas CFD results, there has been some slow,
but steady progress in lattice gas codes [3]. Also, there are theoretical
reasons to believe that their perceived limitations are not
fundamental. Furthermore, there is a trend in MD to converge toward
the cellular automata way of doing things [4]

I'm looking for a Ph.D. position (preferably in Europe) exploring some
of these ideas, currently writing a research proposal further
explaining the details of the project. I would very welcome your
critical comments on this work in progress (not yet online), and, of
course, would be very interested if there is a Ph.D. position vacancy
in your research group. I would love to drop by from an interview. My
CV and my resume can be found on my web site:


Finally, I would like to apologize for this lengthy, and somwhat
crammed message which was necessary to explain this somewhat strange
approach to the PFP.


Eugene Leitl

[1] http://www-graphics.stanford.edu/software/volpack/

[2] http://linux-green.lanl.gov/~pxl/papers/sc96/INDEX.HTM

[3] http://physics.bu.edu/~bruceb/MolSim/

[4] http://linux-green.lanl.gov/~pxl/papers/par_md.ps

More information about the Beowulf mailing list