Assessment of ab initio protein structure prediction

The assessment of predictions of protein structures from amino acid sequences is related to the question of analyzing the differences between similar natural protein structures to study protein evolution and conformational change. However, assessments present certain special features: the alignment is fixed, unlike evolving proteins which are subject to insertions and deletions, and the character of the differences between predicted and target structures is different in kind from the differences between related known proteins. This report describes the assessment methods used for CASP-2, the recent organized `blind test' of protein structure prediction. This is set in the context of general methods for extraction of similar substructures. It is emphasized that the problem of finding one common similar substructure is much simpler than that of guaranteeing the identification of all common substructures with similarity better than a specified threshold. Software based on the results presented here will make it possible to extract from two protein structures all substructures that have r.m.s. deviations upon optimal superposition no greater than a prespecified value. Applied to a predicted structure and the experimentally-determined coordinates of the target, this will identify any and all three-dimensionally correct features of the prediction.