Application of statistical geometry to protein folding

Motivation. Statistical geometry based methods have been used previously in ab initio protein folding simulations, fold recognition calculations, and protein structure analysis. Earlier applications of a chain growth folding method have difficulties in identifying native-like structures in the ensemble of generated structures. In addition, earlier simulations employed only two-body statistical potentials since the use of higher order potentials was computationally prohibitive. We have proposed to implement four-body statistical potentials to guide simulations thanks to the improved algorithms developed by our collaborator Prof. Jack Snoeyink (UNC-Computer Science). We have also proposed to employ off-lattice simulations hoping that it would generate a more realistic predicted structure ensemble than with the on-lattice method.; Methods. Delaunay tessellation uniquely partitions a set of points representing amino acid residues in the folded protein structure into an aggregate of space-filling, irregular tetrahedra that define uniquely four nearest neighbor amino acids. The four-body statistical potential derived by the means of tessellation was used in on-lattice chain growth simulations. Partial tessellation was employed during chain growth to significantly speed up the calculations of the four body potentials. A novel off-lattice chain growth algorithm was applied to folding simulations as well. Average Distance Matrix (ADM) calculated form the ensemble of predicted structures was used to reconstruct the native-like protein chain geometry. Four body scoring function has been applied to the analysis of folding trajectories obtained in independent calculations using Go potentials.; Results. The new implementation of the chain growth algorithm is substantially more computationally efficient than the previous one allowing, for the first time, the use of four body potentials. Comparable structure ensembles have been obtained using both on- and off lattice approaches for several proteins. ADM was found to be more similar to the distance matrix of the native structure then most decoy structures for most of the studied proteins. Distance geometry algorithms have been implemented to restore 3D coordinates of the structure corresponding to ADM. Application of four body potentials to the analysis of folding trajectories revealed strong correlation between the average structure score and the probability of folding as assessed in independent folding simulations.