Meta prediction of protein crystallization propensity

Abstract

Production of high-quality crystals is one of the main bottlenecks in the X-ray crystallography driven protein structure determination. Availability of structure determination data repositories, such as TargetDB and PepcDB, and flexibility in target selection in structural genomics motivate development of methods that predict crystallization propensity from a given protein sequence. We introduce a novel linear model tree-based meta-predictor, MetaPPCP, which takes advantage of the complementarity of state-of-the-art protein crystallization propensity predictors to provide predictions with about 80% accuracy. Our method combines predictions of XtalPred and CRYSTALP2 with information concerning isoelectric point, hydropathy and number of solved structures for similar sequences. Empirical comparison shows that MetaPPCP outperforms current predictors including OB-Score, XtalPred, ParCrys, and CRYSTALP2. MetaPPCP obtains over 92% accuracy for over a half of its predictions that have probability (propensity to be predicted as crystallizable or crystallization resistant) of above 0.8. The proposed method could provide useful input for target selection procedures of current structural genomics efforts.