Quality assessment of protein structure models using evolutionary information

PALOPOLI, N.; JURITZ, E.; FERNANDEZ ALBERTI, S.; DIEGO FABIAN GOMEZ CASATI; PARISI, G.

Montevideo

Congreso; Congreso de la . Int Soc for Computacional Biology (ISCB) Latin-America 2010; 2010

ISCB

The study of protein structure plays a central role in understanding biological function. In thissense, the prediction and selection of best structural models are current major topics instructural bioinformatics. Here we present a novel method that use evolutionary information forthis assessment.Our method, called BEEP (BEst Evolutionary Pattern), considers how the modulation ofsequence divergence by structural features produces a structurally constrained residuesubstitution pattern. To recover this information we use the Structurally Constrained ProteinEvolution model (SCPE) [1]. The SCPE simulates protein evolution by introducing randommutations into the evolving sequences and selecting them against too much structuralperturbation. The output for each protein model is a set of structurally constrained, site-specificsubstitution matrices. Then, the models could be compared under a maximum likelihoodframework. For a given set of homologous sequences and a fixed topology, the best structuralmodel would be the one that better explains the substitution pattern found in the homologoussequences.We have applied this method to the assessment of 55 protein targets from CASP8 [2]. For 82%of the targets our top ranked decoy is among the lower Calpha-RMSD decoys (see Figure 1 asan example), showing that the use of evolutionary information through maximum likelihood (ML)calculations is able to select native-like conformations. Moreover, in 89% of the targets the scoreof the reference solved structure is similar to those of many low Calpha-RMSD decoys,indicating that we should be able to recover not just one, but a complete set of native-likestructures. This could be expected taking into account that the native state of a protein is notunique and it comprises an ensemble of conformers. Each of these conformers could haveimportant functional implications for the protein, imposing specific structurally derived restrictionson the divergence of its sequence, so any prediction method would benefit from taking intoconsideration this pattern of evolution. In particular, we have found that 25% of the targets in ourdataset could comprise a highly populated and heterogeneous native state, as some of theirbest ranked decoys display a difference in their Calpha-RMSD values against the solved PDB ofup to 10 RMSD units, much higher than expected.