Protein Structure and Function: Application of Bioinformatics Methods - John Rigden 2014

Ab Initio Protein Structure Prediction
Model Selection
Clustering of Structural Models

To determine the state with the minimum Free energy value, many ab initio modeling Methods employ adapted Structure clustering techniques. Shortle et al. (1998) demonstrated that for all 12 cases studied, the conformation at the center of the largest cluster was closer to the native structure than the majority of templates. The cluster-center structures belonged to the 1-5% of structures most closely resembling the native ones.

Zhang and Skolnick developed SPICKER, an iterative structural clustering method (Zhang and Skolnick 2004c). They utilized 1,489 representative test Proteins, generating up to 280,000 structural decoys for each. The top models among the five highest-ranked models for each protein fell within the top 1.4% of all decoys used. For 78% of these 1,489 proteins, the RMSD between the best of the top five ranked models and the decoy structure closest to the native one was less than 1 Å.

In the ROSETTA ab initio modeling method (Bradley et al. 2005), structural decoy clustering is used to select low-resolution models. The model structures are subsequently refined through full-atom modeling, yielding the final models. In the TASSER/I-TASSER methods (Zhang and Skolnick 2004a; Wu et al. 2007), thousands of model decoys obtained via Monte Carlo calculations are clustered using the SPICKER method (Zhang and Skolnick 2004c), with the cluster centers serving as the final models. In the method developed by Scheraga and coworkers, structures are first clustered, followed by the Selection of those with the lowest energy values (Oldziej et al. 2005).