|
We have conducted a large-scale comparison of the performance of protein structural alignment methods, relying on the observation that protein structural alignment is a geometric optimization problem. We considered a set of 2930 CATH v.2.4 domains and align all structure pairs using six publicly available protein structure alignment programs, SSAP, STRUCTAL, DALI, LSQMAN, CE and SSM. As a first step, we evaluated the performance of these programs by comparing their ROC curves, using a “gold standard” derived from the structures' CATH classification. Although ROC curves are the standard way used in the structural biology community to evaluate structural alignment methods, we argue that ROC curve based evaluations have their limitations, which can be overcome by using a direct, geometric comparison of the alignments. We think about a structural alignment program as an optimizer: given two protein structures it should find the best alignment. In this context, a good alignment is defined as being long (i.e. many pairs of aligned residues), geometrically similar (i.e. the Euclidean distance between aligned pairs is small) and with few gaps. We use four different geometric match measures that combine these properties into a single number, allowing direct comparison of the alignments found by the different programs under study. Using this novel perspective, we combine the different methods of protein structure alignment into a new, improved method that we call the “Best-of-All”.