Optimal neural networks for protein-structure prediction

Energy optimization for off-lattice protein folding

Two three-dimensional AB off-lattice protein models consisting of hydrophobic and hydrophilic monomers are studied in this paper. By incorporating an extra energy contribution into the original energy function, the protein folding is converted from a constraint optimization problem into an unconstrained one which can be solved by the well-known gradient method. From the initial configurations randomly generated by the heuristic strategy proposed in this paper, our algorithm can find better results than those by nPERM for the four Fibonacci sequences. Based on the initial configurations obtained by energy landscape paving (ELP) routine, some of our results for the lowest energies are better than the best values reported in the literature.

Annealing contour Monte Carlo algorithm for structure optimization in an off-lattice protein model

We present a space annealing version for a contour Monte Carlo algorithm and show that it can be applied successfully to finding the ground states for an off-lattice protein model. The comparison shows that the algorithm has made a significant improvement over the pruned-enriched-Rosenbluth method and the Metropolis Monte Carlo method in finding the ground states for AB models. For all sequences, the algorithm has renewed the putative ground energy values in the two-dimensional AB model and set the putative ground energy values in the three-dimensional AB model.

Structure optimization by conformational space annealing in an off-lattice protein model

The optimization results by conformational space annealing are presented for an off-lattice protein model consisting of hydrophobic and hydrophilic residues in Fibonacci sequences. The ground-state energies found are lower than those reported in the literature. In addition, the ground-state conformations in three dimensions exhibit the important aspect of forming a single hydrophobic core in real proteins. The energy landscape for the population of local minima is also investigated.

Structure optimization in an off-lattice protein model

We study an off-lattice protein toy model with two species of monomers interacting through modified Lennard-Jones interactions. Low energy configurations are optimized using the pruned-enriched-Rosenbluth method (PERM), hitherto employed to native state searches only for off-lattice models. For two dimensions we found states with lower energy than previously proposed putative ground states for all chain lengths >/=13. This indicates that PERM has the potential to produce native states also for more realistic protein models. For d=3, where no published ground states exist, we present some putative lowest energy states for future comparison with other methods.

Application of a chaperone-based refolding method to two- and three-dimensional off-lattice protein models

A model of protein-chaperone interaction as a two-phase (unfolding/refolding) iterative annealing mechanism able to promote structural segregation of hydrophobic and hydrophilic monomers and thereby facilitate access to nativelike states has recently been applied successfully to two 22-mers of the Honeycutt and Thirumalai BLN (hydrophobic, hydrophilic, neutral) heteropolymer model. This technique is here applied to a much wider data set: 94 8-mers of the off-lattice protein model originally presented in two dimensions by Stillinger and Head-Gordon, and later extended into three dimensions by Irbäck and Potthast; the model chaperone is shown to be equally successful, and by progressive elaboration of the chaperone model as in the earlier BLN model work, to be utilizing very similar underlying mechanisms. It is demonstrated that on average, contacts with the model chaperone give rise to a consistent movement in structure space in the direction of more nativelike structures; this method of global minimization does not therefore rely fundamentally on random search. Insofar as the responses to the chaperone of the two- and three-dimensional forms of the substrate model do differ, this can be interpreted as reflecting the different handling of hydrophilic monomers in the models-in particular, whether there is active repulsion between these and monomers of hydrophobic character. The chaperone-induced refolding method is also tested on a set of 220 9-mer chains of each version of the substrate model, where it is seen that the two-dimensional model, with its more clearly distinguished roles for the hydrophobic and hydrophilic monomers, shows a more favorable scaling behavior.

A global optimization strategy for predicting alpha-helical protein tertiary structure

We present a global optimization strategy that incorporates predicted restraints in both a local optimization context and as directives for global optimization approaches, to predict protein tertiary structure for alpha-helical proteins. Specifically, neural networks are used to predict the secondary structure of a protein, restraints are defined as manifestations of the network with a predicted secondary structure and the secondary structure is formed using local minimizations on a protein energy surface, in the presence of the restraints. Those residues predicted to be coil, by the network, define a conformational sub-space that is subject to optimization using a global approach known as stochastic perturbation that has been found to be effective for Lennard-Jones clusters and homo-polypeptides. Our energy surface is an all-atom 'gas phase' molecular mechanics force field, that is combined with a new solvation energy function that penalizes hydrophobic group exposure. This energy function gives the crystal structure of four different alpha-helical proteins as the lowest energy structure relative to other conformations, with correct secondary structure but incorrect tertiary structure. We demonstrate this global optimization strategy by determining the tertiary structure of the A-chain of the alpha-helical protein, uteroglobin and of a four-helix bundle, DNA binding protein.

Evidence for nonrandom hydrophobicity structures in protein chains

The question of whether proteins originate from random sequences of amino acids is addressed. A statistical analysis is performed in terms of blocked and random walk values formed by binary hydrophobic assignments of the amino acids along the protein chains. Theoretical expectations of these variables from random distributions of hydrophobicities are compared with those obtained from functional proteins. The results, which are based upon proteins in the SWISS-PROT data base, convincingly show that the amino acid sequences in proteins differ from what is expected from random sequences in a statistically significant way. By performing Fourier transforms on the random walks, one obtains additional evidence for nonrandomness of the distributions. We have also analyzed results from a synthetic model containing only two amino acid types, hydrophobic and hydrophilic. With reasonable criteria on good folding properties in terms of thermodynamical and kinetic behavior, sequences that fold well are isolated. Performing the same statistical analysis on the sequences that fold well indicates similar deviations from randomness as for the functional proteins. The deviations from randomness can be interpreted as originating from anticorrelations in terms of an Ising spin model for the hydrophobicities. Our results, which differ from some previous investigations using other methods, might have impact on how permissive with respect to sequence specificity protein folding process is-only sequences with nonrandom hydrophobicity distributions fold well. Other distributions give rise to energy landscapes with poor folding properties and hence did not survive the evolution.