Hello, I have 10,000 models of a 5 chain protein assembly (1SAC.pdb). I need to rank them based on their nativeness (RMSD vs. 1sac.pdb). I have tried multiple tools for this, profitV3.1, maxcluster etc but in all cases I have a problem where models that superimpose very well with the native structure by eye (and which should be native) give very poor rmsd and will therefore be classified as non-native. Conversely, some models which poorly align by eye have a better rmsd with the native structure. Accordingly, I can run the programs but I do not trust their rmsd rankings.
I have tried joining the chains, specifying zones, and iterative fitting but in all cases I have many good models with low scores and many poor models with higher scores. I have been using these tools for some time and never had any problems with monomers and heterodimers. This protein is a homopentamer, I think this is significant in some way.
Any clues out there??
1 answer
I think there are a few options here.
Firstly, RMSD has been shown to not be a very good metric overall (it can be easily 'tricked'), so you can try your fitting using other programs/metrics like TMalign.
Secondly, try removing 4 of the chains, so that your reference structure only has a single chain. If its a homopentamer it won't matter, and may help your alignments.
That said, I've done a similar thing, whereby all my matching is done using UCSF chimera, and for the most part the models match nicely even to reference structures with 10+ chains (e.g. 3J9Q which contains 2 distinct homohexamers). You're welcome to try this script too if it helps(https://github.com/jrjhealey/bioinfo-tools/blob/master/ModelMatcher.py)
Disclaimer: it can be slow, and you'll need to set up pychimera and Chimera but there are instructions at the pychimera github page if needed.
Log in to answer this question.