Tropsha and Cammer used Delaunay tetrahedralizations to score protein conformations by Simplicial Neighborhood Analysis of Protein Packing (SNAPP). The vertices of the tetrahedralizations are chosen at the mass centers of the residues. Frequencies of the four-tuples of different residue types are compared with frequencies occurring in a representative database, and the score is computed by summing the log-likelihoods of the four-tuples. We have extended this to use separation along the chain in addition to four-tuple frequency, as that appears to have good discrimination potential. We are extending the analysis published already for five proteins to additional examples where sufficient numbers of hydrophobic core mutants have been analyzed, in order to assess the generality of the scaling relationship between the statistical four-body potentials and protein stability. In collaboration with Marshall Edgell we pursue additional experimental details of the implied relationships by constructing targeted point mutants and by making combinatorial mutant libraries in a high throughput mode, to improve the database on which we base regression analysis of stabilities.

Gan, Tropsha, and Schlick proposed to use four-body potentials to fold protein sequences and derive candidate structures. They implemented a Monte Carlo chain-growing algorithm, but found that the four-body potentials were too time consuming to compute. By incorporating calls to the incremental Delaunay triangulation of the CGAL library directly into their lattice-folding code, we achieved two to three orders of magnitude speed up, and were able to use the SNAPP potentials during protein folding instead of as post-processing. This improved the SNAPP scores of the candidate folds, although the RMSD deviations are still larger than hoped for. Part of the reason for this shortcoming is the restriction to the lattice, and current efforts include off-lattice folding by Rosetta-type analysis.