Fig. 2
Work flow of the structural descriptor optimization used to parameterize mutual information. The backbone structure is characterized by a pair of virtual alpha carbon dihedral angles, whose two-dimensional space can be discretized by the Voronoi partition into k states or seeds. The number of seeds k and their locations in the dihedral angle space can be optimized by a Monte Carlo search using mutual information as objective function, as illustrated on the left side of the Figure. The tertiary contact structure is characterized by the contact distances between pairs of non-adjacent residues, with the parameters dmax that describe the maximum distance of interaction and m that dictates the number of discrete bins by which the length dmax is partitioned. An exhaustive search is made across various dmax and m with mutual information as objective function. The two sets of information-optimized descriptors (for the backbone and for tertiary contacts) are used to compute the mutual information Ibb and Itotal used in the virtual mutagenesis procedure (see Fig. 1), and also used to parameterize the energy function ΔU employed in the threading experiment