Parallel Evolutionary Computations in Discerning Protein Structures - Parallel Computing for Bioinformatics and Computational Biology - page 462

TABLE 19.1 Classification of Methodologies Used in Solving the PSP Problem

Experimental Energy

Methods Simulation Minimization

X-ray crystallography 2-4 months

after crystal

NMR spectroscopy 2-3 years

Molecular dynamics 1 ps of the folding process

can be simulated

Monte Carlo analysis Highly computational

Lattice simulations Highly computational

Off-lattice simulation Highly computational

Evolutionary algorithm Stochastic search

and nonatomistic lattice simulation, off-lattice simulation, and energy minimization.
Various methodologies are classified in Table 19.1.
Fundamentally, there are two main reasons for having many alternate techniques.
The first reason would be a feasibility issue. Some protein conformations are easily
found using empirical methods, such as X-ray crystallography, because they crystal-
lize easily, yet others are easily found in solution using nuclear magnetic resonance
spectroscopy [12]. Still others cannot be crystallized at all. The second reason would
be a compromise between dihedral angle resolution and time to calculate a solution.
It is known that the time involved to find structures using empirical measurement
techniques can sometimes be cost prohibitive. Additionally, each approach has reso-
lution limitations. The trade-off between resolution and time is apparent in Table 19.1.
Furthermore, having alternate search methods permits their use in complementing one
another.
Corporations like International Business Machines (IBM) have recognized the
need for extensive computational resources to run experiments and is currently build-
ing a petaflop computer specifically designed for simulating the protein folding
process [16, 17].
In examination of IBM's petaflop computer chances of solving the PSP problem,
we turn to what is known about the computational requirements for a simulation. It is
known that the time steps required to accurately account for thermal oscillations
of the protein are on the order of 1 fs (10 − 15 s) [18, 19]. Therefore, if a single
calculation between two atoms must be computed within a femtosecond, the number
of calculations required for a single pair combinatorially rises as the number of atoms
is increased. For example, if it takes 4000 floating point operation per second (flops)
to calculate one quantum mechanical function between two atoms, the number of
flops required would grow exponentially 4000 number of atoms per femtosecond for a
real-time simulation. IBM's petaflop computer is scheduled to be finished in 2006;
however, this computer though built for massive simulations may not meet the sim-
ulation requirement for protein folding. To date, these protein folding methods were
used in finding 2% of the total number of proteins found in the Protein Data Bank [20].