We have been investigating techniques to improve an ab initio chain growing algorithm for protein folding. The current version of the algorithm is at least three orders of magnitude faster than its predecessor. Recent work has concentrated on using a distance geometry matrix to represent a protein. Pande et al. reported that, for energy-based folding, averaging the distance matrices of many backbone decoys gave a more native-like distance matrix, so we have been exploring this with Shuquan Zong, a postdoc in Medicinal Chemistry. Several thousand decoys for a given sequence are rapidly generated and the best are chosen using his implementation for fast multi-body potential scoring. The distance geometry of each is calculated and averaged. The average distance geometry is significantly closer to the distance geometry of the native than most or all of the individual decoys. For most small proteins (under 100 residues) freely available distance geometry reconstruction software can produce a fold that closely resembles the native backbone. Refinements to the average distance geometry to correct inaccuracies in the shorter distances and the those in suspected secondary structure locations will improve the folds generated by the reconstruction software.

The average distance geometry also provides information on long range residue contacts. Subsets of residues from different parts of the sequence that are relatively nearby define a core, and several conformations of this core are created that meet the constraints in the average distance geometry. A modification of the chain growing algorithm will use a new implementation of a molecular chain inverse kinematics solver to build chain segments between residues in the core. This will significantly reduce the size of the fold conformation space and force the algorithm to build decoys that meet predicted long range residue contacts. This will greatly increase the percent of decoys that are native-like and will lead to better folds.