[Tutorial Index]


What accuracy can I expect for a model build by the automated mode of SWISS-MODEL?

Evaluation of template structure and model quality is a crucial step in homology modelling. The reliability of different protein modeling methods can be assessed by evaluating the results of blind predictions after the corresponding protein structures have been determined experimentally. The overall performance of the SWISS-MODEL pipeline is evaluated by the EVA project. SWISS-MODEL was the first comparative modelling server to join the EVA project in May 2000, and has since then been continuously evaluated. As of Summer 2005, EVA-CM is based on the assessment of 261 weekly releases of the PDB database, resulting in 48098 protein models for 19698 protein target chains for five different prediction servers, among these 18314 from SWISS-MODEL. All models generated by SWISS-MODEL server, evaluation results, score definitions and detailed statistics are available from the EVA project website.

The C-alpha atoms RMSD after global superimposition of the model and the experimental target structures was computed and plotted vs. % of sequence identity between target and best template to give an estimation of the overall accuracy of the different modelling servers with regards to different sequence identities between target and template:

In general, major differences between the individual prediction methods are only observed for target-template pairs sharing sequence identities of less than 40 %, where methods favouring higher coverage of the target sequences are more likely to generate models with a higher RMSD. As expected, model RMSD is increasing with decreasing alignment accuracy as defined by the percentage of equivalent C-alpha positions (within 3.5 Angstroms) between the optimally superimposed target and model structures: