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

Comparative Protein Structure Modeling
Efficiency of Comparative Modeling Methods
Accuracy of Methods

An informative evaluation of Cell/13.html">Protein Structure modeling Methods, including comparative modeling, can be obtained from the biennial CASP (Critical Assessment of Protein Structure Prediction) meeting (Moult 2005). Protein modelers are challenged to generate a model for a sequence of unknown 3D structure and submit it to the organizers prior to the meeting. Concurrently, the actual 3D structures of the target Proteins chosen for prediction are determined by X-ray crystallography or NMR spectroscopy. These structures are made publicly available only after the models have been calculated and submitted for evaluation. This ensures a bona fide Assessment of Protein structure modeling methods, although it is often difficult to determine whether the success of a given method stems from the software features or the expertise of the modelers.

An alternative large-scale, continuous automated evaluation method is employed in the EVA (Evaluation of Automatic protein structure prediction) project (Eyrich et al. 2001). Every week, EVA feeds sequences soon to be released in the PDB to the prediction servers participating in the competition. After collecting and Processing the modeling results, the platform provides detailed statistics on Secondary structure prediction, Fold Recognition, comparative modeling, and contact prediction. The LiveBench initiative implements a similar approach using unique evaluation methods (Bujnicki et al. 2001).

Rigorous statistical assessment (Marti-Renom et al. 2002) of blind prediction experiments has demonstrated that the accuracy of various modeling approaches—such as segment matching, rigid-body assembly, spatial restraint satisfaction, or any combination thereof—is roughly comparable when the methods are applied correctly (Dalton and Jackson 2007; Wallner and Elofsson 2005b). This also reflects the fact that critical modeling steps, such as Template Selection and alignment accuracy, heavily dictate overall model precision, given that the protein core is highly conserved. From a practical standpoint, models should be evaluated by their utility—that is, how well they advance our understanding of protein function. Unique functional roles are typically tied to distinct structural features, which are more frequently found in variable loop regions rather than in the conserved core. However, descriptions of functional sites are not only curated manually but are also increasingly missing or incomplete. This is particularly true for high-throughput structural Genomics outputs, which often specifically target proteins of unknown function. Consequently, large-scale benchmarking of modeling Methods based on the accuracy of functional annotations remains a desirable yet challenging objective in practice (Chakravarty and Sanchez 2004; Chakravarty et al. 2005).