Stochastic algorithm for kinase homology model construction

A stochastic algorithm for constructing multiple loops in homology modeling of proteins is presented. The algorithm discards variable values in iterations based on a cost function and on statistical analysis of results. Values that remain are used for constructing an ensemble of best solutions. In test cases, the stochastic algorithm retains all the best solutions, compared to an exhaustive scan of the full set of conformations. Individual loops are constructed by adding dipeptide units. Dipeptide conformations are extracted from a database of proteins and their conformations include bond lengths and all angles. Single loops are constructed from both N- and C- terminals to the center, and loop closure is evaluated by a combination of penalties for the peptide closure and Miyazawa-Jernigan (MJ) [1]. residue-residue interactions with the rest of the protein. Large ensembles of each loop are clustered and re-evaluated with a refined [2]. energy term. The reduced, clustered set of each loop is then employed to construct simultaneously all the loops. The algorithm was applied to construct simultaneously six loops in c-Src kinase family proteins, incorporating a total of 37-40 residues. The best RMSD for reconstructing the loops is 1.45 A for Lck (structure 1QPE in the Protein Data Bank) and 2.54 A for human c-Src (structure 1FMK). The multiple loop conformations with lowest energy have higher RMSD values, of 2.06 A and 3.09 A, respectively. The average RMSD values for the first 1000 conformations are 3.00 A and 3.46 A, respectively. Models for the "open" structures of c-Src and of Jak-2 were constructed on the basis of 1QPE. The Jak-2 model is found to be more flexible in the loops region than its c-Src counterpart.