Secondary Structure Prediction

Question

[Elisa]

Hi all,

I'm trying to predict secondary structure of a transmembrane protein. I've used the following tools/servers : PsiPred, Jpred, ProfSec (PhdSec), porter and spritz. I've got different results for predictions (in some case enough marked). How can optimize results and find a consensus?

I'm wondering if prepare a dataset using only homologs and perfom an analysis by a support vector machine.

Any suggestion?

Thanks

[Jarretinha]

Do you have a good reason to predict only TMHs instead the whole domain/protein? Anyway, you'll not find a consensus. Too many different methods. Besides that, only I-TASSER can possibly get better results than PsiPred server.

[Nicojo]

@Jarretinha: could you give us a reference for your statement about I-TASSER?

Answer 1

[Lars Juhl Jensen]

Since you write that your protein is transmembrane, the first thing I would do would be to use a tool such as TMHMM to predict the transmembrane segments of the protein (there may be slightly better tools than TMHMM for this, I am not an expert on this).

Having masked those, I would use SignalP to predict if there is a signal peptide. The reason for this is that transmembrane proteins often contain signal peptides. If there is one, you will want to chop it off, as it is not part of the mature protein.

I would then use protein domain databases such as Pfam, SMART, InterPro, and SUPERFAMILY. If there are any predicted domains, I would check if there are known 3D structures for any of the family members, in which case I would infer the secondary structure from those.

Only after having done all of these steps would I use a secondary structure and intrinsic protein disorder prediction tools to predict the structure (or lack thereof) of the remaining segments. For this I would refer to CASP and pick the best method for each task rather than attempt to make my own consensus method.

[Hamish]

Phobius (http://phobius.sbc.su.se/, http://phobius.binf.ku.dk/ and http://www.ebi.ac.uk/Tools/pfa/phobius/) is worth a mention as an alternative to TMHMM and SignalP which provides both signal peptide and transmembrane predictions. My understanding is that transmembrane prediction is still one of those areas where using multiple methods and comparing the results is recommended. If the methods only disagree on the boundaries the prediction is pretty solid, if they disagree on the number of transmembrane domains further evidence may be required.

Answer 2

[Shigeta]

The typical SS programs are not usually well tuned for TM helix detection, some of them exclude membrane proteins from their training sets, so TMHMM is your best shot. I agree with Lars and I wouldn't bother with the other ones.

Psipred will include psiblast profile information so it does have an alignment backing it up. It might be interesting to compare the two, if you can get the likelihood values out of all the programs.

I'm guessing you already know what family your protein is in, so PFam, etc might not be so useful. If you have a family alignment the easiest thing to do is to overlay the TMHMM predictions for each one and

If you wanted to do this with a set of MPs of known structure you could train an SVM - it would be great, but it might take some time/effort. eyeballing the alignments and prediction of the family can be surprisingly useful if you don't have the time to put that together.

Answer 3

[Rm]

HMMTOP, SOSUI, TMHMM, TMPRED, MEMSAT, are specialised for TM Helix prediction. you can select the consensus regions predicted by multiple tools.