Teresa Head-Gordon and Silvia Crivelli,
Lawrence Berkeley National Laboratory
Betty Eskow, Richard Byrd, and Robert Schnabel,
University of Colorado, Boulder

Research Objectives

We have developed a joint global optimization approach to protein fold prediction based on sampling, perturbation, smoothing, and biasing, which has been successful working directly on the potential energy surfaces of small homopolymers, homopeptides, and -helical proteins. We currently have DOE funding to extend the ab initio approach to -sheet and mixed / proteins as well.

Computational Approach

As the starting point, secondary structure is predicted by a neural network algorithm. Our approach then consists of two phases. The first phase starts with a completely extended conformer, with no secondary or tertiary structure, and performs local minimizations using the constrained energy function first and then the unconstrained potential energy function. The second phase starts with the outcome of the previous phase as the first member of a list of local minimizers. From the set of dihedral angles predicted to be coil, the algorithm randomly selects a subset, then performs a small-scale global optimization using the selected dihedral angles as variables while keeping the rest temporarily fixed at their current values. This optimization produces a number of local minimizers to the unbiased energy function on the subspace of dihedral angles chosen. A number of those conformations with low energy values are considered for further refinement, which is done by performing local minimizations (using the unconstrained energy function) on the full variable space. The new minimizers are merged into the current list, the lowest energy conformation is selected from this list, and the second phase repeats for a number of iterations.

Accomplishments

We have included with the AMBER energy function a potential of mean force between hydrophobic carbons like that learned from our experimental and simulation studies of amino acid monomers in water. Using our global optimization approach with this new energy function, starting with the crystal structure of four -helical proteins, we generated no new conformation that was lower in energy than the crystal/NMR structure.

We tested the stochastic/perturbation with soft constraints algorithm on the prediction of the -chain of uteroglobin and a DNA binding protein (referred to as 2utg and 1pou, respectively). Both proteins were ~70 amino acids long and contained ~1200 atomic centers. We also have preliminary results on another 70-amino-acid protein, 3icb, and a 143-amino-acid protein, 3cln.

A comparison between the NMR structure of a four helix bundle DNA binding protein, 1pou (right), and the outcome from our global optimization algorithm (left). Purple indicates helical regions, while green indicates residues that are coil.

Significance

During the next decade, the experimental acquisition of structural data for proteins will be driven by the need to define a "basis set" of fold topologies that will allow for generalization to the fold and function of all protein sequences from any genome. This experimental underpinning will allow the computational effort in structural genomics to establish a robust framework for reliably predicting the three dimensional architecture of proteins in order to gain insight into their function, and to integrate this functional annotation into systems level understanding.

Publications

S. Crivelli et al., "A hierarchical approach for parallelization of a global optimization method for protein structure prediction," in Proc. Euro-Par '99 (in press).

A. Azmi et al., "Predicting protein tertiary structure using a global optimization algorithm with smoothing," in Proc. International Conference on Optimization in Computational Chemistry and Molecular Biology: Local and Global Approaches (in press).

S. Crivelli et al., "A global optimization strategy for predicting protein tertiary structure: -helical proteins," Computers & Chemistry (conference proceedings for New Trends in Computational Methods for Large Molecular Systems, submitted).