1
$\begingroup$

How do I perform protein model quality assessment?

I will obtain predicted structures from DMPfold, I-Tasser, and Rosetta, what should be the next approach to select a structure from among this set of models?

Note, I have been looking at some, appropriate methods such as MetaMQAP, QA-RecobineIt, but their servers are not working.

$\endgroup$
  • 1
    $\begingroup$ Please don't ask "what is the best". The best will always be a matter of opinion and the best method for your data will depend on exactly what kind of data you have. Please edit your question and rephrase it so it isn't asking for opinion (best is an opinion). $\endgroup$ – terdon Sep 24 '19 at 14:54
  • $\begingroup$ Edited to be compliant $\endgroup$ – Michael Sep 24 '19 at 21:57
2
$\begingroup$

Have a read of the Estimation of model accuracy in CASP13 paper to get the latest rankings.

Generally you would run the model quality software on the models you have generated and choose the best. Also have a look at the structures in PyMol, you learn a lot that way. See, for instance, how much variety there is, and where the models vary. Is there a conserved core in many of the models? Or is there just a blob of "spaghetti" that doesn't look like an ordered protein?

If you have many models (e.g. from a Rosetta run) you might want to look into clustering the models first to reduce the number you have to work with.

| improve this answer | |
$\endgroup$
0
$\begingroup$

Co-variance with convolution is an extremely hot method, so DMPFold via psicov. Checking the reality of the folding is obviously important, convolutions can be missed, inaccuracies can happen.

| improve this answer | |
$\endgroup$
  • $\begingroup$ DMPfold uses PSICOV in its pipeline. Does this mean that DMPfold does quality assessment in its pipeline? $\endgroup$ – DanielSebas Sep 25 '19 at 10:51
  • 1
    $\begingroup$ It would be good for you to read the publications on psicov, to understand it. I believe @jgreener is part of the group that developed it $\endgroup$ – Michael Sep 25 '19 at 11:07
  • $\begingroup$ DMPfold has a model quality predictor that is specific to DMPfold. It can only score DMPfold models because it uses some internal measures, such as how well the final model matches the initial distance prediction. DMPfold shouldn't therefore be thought of as a model quality assessment method. PSICOV is part of DMPfold but isn't related to model quality assessment at all - it predicts residue-residue contacts by inverting covariance matrices. $\endgroup$ – jgreener Sep 25 '19 at 14:50
  • $\begingroup$ Straight covariance results in convolution in the non- deep learning sense, ie pure statistical sense (I haven't explained that very well,but hope it makes some sense) $\endgroup$ – Michael Sep 25 '19 at 15:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.