Every two years, the CASP ("Critical Assessment of Structure Prediction") experiment attempts to assess how well people can make structural predictions based on sequence data (http://predictioncenter.llnl.gov). One section of this is dedicated to comparative modelling. The following data come from the assessment of the CASP2 experiment (Martin et al.,1997 "Assessment of Comparative Modelling in CASP2" Proteins: Struct., Funct., Genet. Suppl 1, 14-28).
The overall CASP2 results are summarized by the following graph
As can be seen from the graph as the sequence identify gets higher than 80% models of <1.0Å RMSD can be generated, equivalent to the differences between two crystal structures of the same protein. As the sequence identity drops below 40%, the model quality falls rapidly.
As the sequence identity drops it becomes more difficult to obtain the structurally correct alignment as shown in the following graph:
When one makes mistakes in the alignment, then the RMSD increases. Given that the distance between adjacent C-alpha atoms in a protein is approximately 3.8Å, then an average shift of 1 residue in the alignment will lead to a minimum possible RMSD of 3.8Å
As the graph shows, in practice an average shift in alignment of 1 residue leads to an RMSD of ~5.0Å. An average shift of <0.1residues is required to get a high quality model.
The final graph shows how well we would have expected people to do in CASP2 and how well they could have done if they got the sequence alignment correct:
For each target, the square represents the RMSD between the parent and the target over the SCRs if the structurally correct alignment is used. The diamond represents the RMSD if a simple automatic sequence alignment is used. One would thus hope that manual modification to the alignment would reduce thr RMSD from the diamond towards the square. In practice, this is generally the case, but in some cases the manual correction of the alignment made the model worse.