We have begun to use and study I-Tasser and other softwares eg Modeller. We wish to look for quite small differences between variant protein target sequences which have high homology.
One general concern is whether the precision of the models is restricted to the quoted resolution in PDB of the template determined by physical techniques eg X-ray crystallography. For example if the resolution of the template structure is 1-2A, how do we interpret an average RMSD of say 0.8A between the models of two targets?
We notice that if we submit the same target sequence twice in succession to I-Tasser the Cartesian coordinates are slightly different even after superposition in Chimera. The correlation across residues between the successive models is high but there are these slight differences. Looking at Zhang (2008, I-TASSER server for protein 3D structure prediction) we see reference to Monte-Carlo approaches. We assume that the slight differences we observe with the same sequence submitted twice are the result of this random element seeded differently. Is this more or less correct ?
Big question. Can we use these differences between different submissions of the same sequence as a statistical error for determining significant differences in structure between our different but very similar target models? This would we assume be within the context of the I-Tasser approach and be related mainly to the precision of this approach.