Towards accurate modeling for protein rigidity analysis

We evaluate the performance of the modeling in KINARI, our pebble-game rigidity analysis system, on a benchmark data set of 32 PDB files used by the Gerstein Lab to validate the RigidFinder algorithm. For this we use a novel evaluation method, called the cluster decomposition score. This is adapted from the B-cubed score from the information retrieval literature, which is used as a comparative score on two clusterings of the same data. We focus on the the most extensively used tuning parameter in rigidity analysis, the hydrogen bond energy cutoff. We also propose a new modeling methodology for hydrogen bonds, where those with energies lower than the cutoff, instead of being discarded, are modeled with a weaker constraint. For the 5 largest proteins (501 residues or more), the KINARI decompositions scored significantly better than the crude baseline, indicating that the parameterization used is reliable for larger proteins. For most of the proteins including all hydrogen bonds leads to the best scores. For the others, where a cutoff improved the score, our new modeling methodology matched or improved results for 9 out of 10 PDBs. We investigate the sensitivity of the cluster decomposition score with case studies on pyruvate phosphate dikinase and phosphotransferase.