Mathematical Modeling and Optimization Methods for De Novo Protein Design - Systems Biology: Networks, Models, and Applications - page 46

proteins, RosettaDesign yielded sequences of 70-80% identity as the
final results of energy optimization when multiple runs were started
with different random sequences [41]. Originated in genetics and
evolution, genetic algorithms generate a multitude of random amino
acid sequences and exchange for a fixed template. Sequences with low
energies form hybrids with other sequences while those with high
energies are eliminated in an iterative process which only terminates
when a converged solution is attained [42]. Desjarlais and Handel [20]
have applied a two-stage combination of Monte Carlo and genetic
algorithms to design the hydrophobic core of protein 434cro. Both
Monte Carlo methods and genetic algorithms can search larger combi-
natorial space compared to deterministic methods, but they share
the common disadvantage of lacking consistency in finding the global
minimum in energy.
Recent methods attempt to avoid the problem of optimizing residue
interactions by manipulation of the shapes of free energy landscapes [43].
Another class of methods focus on a statistical theory for combinatorial
protein libraries which provides probabilities for the selection of amino
acids in each sequence position [44-46]. The set of site-specific amino acid
probabilities obtained at the end actually represents the sequence with
the maximum entropy subject to all of the constraints imposed [44,45,47].
This statistical computationally assisted design strategy ( scads ) has been
employed to characterize the structure and functions of membrane
protein KcsA and to enhance the catalytic activity of a protein with a
dinuclear metal center [47]. It has also been used to calculate the iden-
tity probabilities of the varied positions in the immunoglobulin light
chain-binding domain of protein L [45]. Scads serves as a useful frame-
work for interpreting and designing protein combinatorial libraries, as
it provides clues about the regions of the sequence space that are most
likely to produce well-folded structures [48].
Several sequence selection approaches have been tested and validated
by experiment, thereby firmly establishing the feasibility of computa-
tional protein design. The first computational design of a full sequence
to be experimentally characterized was that of a stable zinc-finger fold
(bba) using a combination of a backbone-dependent rotamer library
with atomistic level modeling and a dead-end elimination-based
algorithm [49]. Recently, Kuhlman et al. [50] introduced a computa-
tional framework that iterates between sequence design and structure
prediction, designed a new fold for a 93-residue a/b protein, and validated
its fold and stability experimentally. Despite these accomplishments, the
development of a computational protein design technique to rigorously
address the problems of fold stability and functional design remains a
challenge. As mentioned earlier, one important reason for this is either
the almost universal specification of a fixed backbone, or the use of a