Anthony,

The protein structure prediction problem is still an open question. If you have a template, it is still possible to obtain moderate to very good models that may be representative of some state of the protein structure in solution. Unfortunately, without similarity with any other protein in protein data bank (PDB) it is not amenable to the present predictions methods to obtain even the right fold prediction.

You may be interested in look some of the results from CASP 11, for example, in which several groups submit predictions to sequences that may not have similarities with any one in PDB (FM  Free Modeling category)  and you will find out the conclusions I have adressed above.

http://www.predictioncenter.org/casp11/results.cgi

I have been trying to develop a project to derive from experimental data a way to predict with more confidence the tertiary structure of proteins, specially to those that are difficult to obtain from template based modeling. I believe I have some good results I will publish in a while.

I could discuss more with you if you are interested.

As for your question regarding to secondary structure prediction, it has been shown that the algorithms developed up to now are quite robust and we can mostly trust to them. Anyway, you may find your self dealing to a protein that is a exception. It a matter of check the statistics.

Best regards,

Állan.

ab-initio and threading method u can used for this purpose

The threading method is the first stage to I-TASSER, but it is still only as useful as the likeness to the templates recovered from a PDB database. 

Hello Allan, would you tell more about your project? I have been trying something similar, I made a database with all protein resolution < 2A and I made a resume of the distance between pair combinations to get the average distance between the neighborhood, at this moment its not really folding but if I start from a folded protein its not denaturing any more. Btw I also from Brazil.

Hi Anthony,

it really varies highly from case to case. Even though template-free modeling (or free modeling as Allan pointed out) made considerable progress, it is still very hard to get a good model if sequence similarity to the template is low.

However, many groups in CASP11 (our group included) obtained very good results by using evolutionary contacts. I would recommend to check how many sequences are available for your protein if you search with PSIBlast or HHBlits. If the number fo sequences is high (around 5L, where L is the number of residues of your proteins), you can use of many evolutionary contact prediction methods (PSICOV, plm-DCA, Gremlin) to obtain the contacts and then fold the protein with free modeling servers (Pcons-Fold, RBO Aleph, the last one is self-advertising).

If you don't have enough sequences, try to get predictions from the top 10 CASP 11 or CAMEO servers (http://cameo3d.org/sp/3-months/). Superimpose the structures from the different servers and look whether they are similar. Similarity between methods is a good indication that the prediction is right.

Otherwise, you will just have to gamble :-).

Many thanks, Michael. That information was extremely useful. 

Dear Michael, after reading you response and acting on some of your suggestions I do have a very simple question regarding 5L sequences. 

1) Am I correct in understanding that by 5L is 635 if your protein sequence is 127 residues long? 

2) Why is it important to have greater than 5xL number of returned sequences (unique DB:ID from the results of PSI-BLAST of one of my sequences)? Surely it is more to do with quality of templates than quantity?

BTW, the process you've described is very helping. I will take a look at RBO Aleph.

Thanks

Hi Anthony,

To 1)

Yes, this is correct.

To 2)

There is no fixed number of sequences that you need. This number is just currently the general consensus that you can be quite sure that evolutionary contacts will be sufficiently accurate. But this is really a rule of thumb. It might vary from case to case and will surely go down as better contact prediction methods are developed. I would say that the "critical zone" is below 1xL sequences for evolutionary contact prediction.

You are also correct that the quality of the template matters and not the quantity.  For a simple modeling problem you can easily identify a single templae and you are done. But the challenge is to identify the correct template in hard modeling problems, where sequence similarity is very low.

Virtually all protein structure prediction methods (with the exception of pure physics-based methods) leverage the sequence profile information of related sequences. This is especially true for template identification methods like HH-search. Thus, the number of sequences is a pretty good indicator whether the algorithm will give good results. The higher, the better. At least this is true in "hard" modeling cases where a template is not easily identified.

It is worth mentioning that the lack of a good "how good is my model" criterion for hard modeling cases is a hot research problem in the field for which, unfortunately, no universally adopted approach exist. What I'm laying out is merely a couple of best practices.

Hope that helps!

Michael

Hi Michael, again, thanks for all of that information it is extremely informative. I'm very well versed in Molecular Dynamics and Quantum Chemistry, however, that has always been from crystal structure or first principle work. This study is fairly new to me but seems quite straight forward. Thanks.