Fold recognition using web-servers

Morten Nielsen (


Fold recognition (FR) is the name given to the process of assigning a known structure (a template) to a sequence of unknown structure (the query). The methodologies used include:

  • Straight sequence methods (like BLAST and PSI-BLAST)
  • Sequence methods incorporating structure, including:
    • predicted and known secondary structure
    • structural environments
    • structural alignments
  • 'True' threading approaches

The difference between "sequence" based methods and methods using threading is not always clear. In principle the sequence based method defines the "fitness" of the query onto the template from on the primary structure of the query and template sequences, respectively. Threading methods on the other hand defines the "fitness" of the query from the structural environment of the template structure. However as you saw from the list above some sequence based methods also incorporates structural information of the template in the alignment so the borderline is not very clear. The most powerful method are neither "true" sequence based nor "true" threading method, but some mixture of the two.

Many fold recognition programs are available over the web, and today we're going to get some experience of how to use them most effectively.

Finding information

Note: If links are not given on this page, it's assumed that you'll use Google to find it

The exercise

Below we give you a list of three protein sequences. You shall now try to use some tools to assign a fold and structure to each sequence. The sequences are placed in the categories CM (comparative modeling), CM/FR (comparative modeling/fold recognition), and NF (new fold). As the categories suggest, the CM is the easy class, the CM/FR the hard, and the NF the difficult (close to impossible) class. (Why is this?)

In the exercise you shall find out which of the three sequences belong to which of the three categories, and for the two sequences belonging to the CM and the CM/FR categories you shall find which template you should use to build an homology model (template recognition).

On to the BLAST web-site. Select Protein BLAST. Blast the three sequences Query1, Query2, and Query3 against PDB. Note you do this by pasting your sequence (including FASTA header) into the Query Sequence window. Then under Database select Protein Data Bank proteins(pdb). Then press BLAST.

  • Q1 What are the E-values for the three searches?
  • Q2 Are any of the hits significant (Eval < 0.001)?

Next use the PSI-BLAST version of Blast. On to the BLAST web-site and choose Protein BLAST once more. Paste in the Query1 sequence. This time under Algorithm select PSI-BLAST. Set the database to Non-redundant protein sequences (nr). Under Algorithm parameters set Max target sequences to 1000. Then press BLAST.

  • Q3 How many significant hits does blast find (E-value < 0.001)?
  • Q4 How large a fraction of the query sequence does the significant hits match?
  • Q5 Do you find any PDB hits among the significant hits (search for the colored S to the right of the E-value))?

Now run a second Blast iteration. Press Run PSI-Blast iteration 2.

Note. You might get a Blast error for one of the sequences. Can you make sense of this error? If you get an error, this just means that Psi-Blast has failed for this query sequence and you cannot answer the questions Q6-Q10 for this Query.

  • Q6 How many significant hits does Blast find (E-value < 0.001)?
  • Q7 How large a fraction of the query sequence does the significant hits match?
  • Q8 Make use you understand what is going on. Why does Blast come up with more significant hits in the second iteration?
  • Q9 Do you find any PDB hits among the significant hits (search for red colored S to the right of the E-value)?
  • Q10 What is the PDB identifier for the best PDB hit?

If you did not find any PDB hits, try a third iteration.

Repeat the PSI-BLAST search for the other two query sequences (up to three iterations).

Now you have probably found that one of the three protein targets could be modeled using sequence searches only, and this query is hence the easy one (the CM query).

Identifying conserved residues

You have now (hopefully) identified a structural relationship between the Query sequence and a protein sequence in the PDB database of protein structures. Say you would like to validate this relationship. This one could do by mutating (substituting) essential residues in the query sequence and test if the protein function (or structure) is affected by these mutations.

The protein sequence of the CM query (Query1) is large (more than 400 amino acids) and a complete mutation study including all residues would be extremely costly. Instead one can use PSI-BLAST and sequence profiles to identify conserved residues that are likely to be essential for the protein structure and/or protein function.

Below you find a set of 8 residues from the Query protein sequence. You shall use the PSI-BLAST and Blast2logo programs to select four of the eight residues for a mutagenesis study (you shall select the four residues based on sequence conservation only).

  • (a): H271
  • (b): R287
  • (c): E290
  • (d): Y334
  • (e): F371
  • (f): R379
  • (g): R400
  • (h): Y436

You shall use the Blast2logo server to identify which residues are conserved in the Query protein sequence. Go to the Blast2logo server and upload the Query sequence. The program allows you to select the sequence database for the Blast search. You shall NOT submit the query here. To save time, we have submitted the job, and you can find the output following this link Blast2logo output.

When the job is completed you should see the logo-plot on the website. If the logo does not display, you can download the image file (click on the Download logo file) and open it from your desktop.

  • Q10.1 Spend a little time looking at the logo plot. Can you understand why the logo is so flat for the first 100 residues (how large a fraction of the query section did the Blast search cover)?
  • Q10.2 Which four of the eight residues listed above are most conserved and hence most likely to be essential for the protein stability and/or function?

You shall use the Phyre protein homology program to validate if the structural properties of the four most conserved residues from question Q10.2 indeed could form an active site. Go to the Phyre web-site and upload the Query sequence. Note it might take some (10-20) minutes before your job is completed. To save you time, we have run the calculation for you. Yoy can find the output here Phyre output.

Find the PDB hit identified by PSI-BLAST (you can click on the SCOP code to get to the PDB template for each model).

  • Q10.3 Does Phyre agree that this hit is significant?

Download the highest scoring Phyre model (click on the View Model image for the first model), and open the model file in Pymol. If you do not have Pymol installed on your computer, you can find a free download here Pymol 099 downloads. Show the location of the four essential residues from question Q10.2 on the structure.

  • Q10.4 Could the residues form an active site?

Remote homology modeling

You shall now use some more advanced tools to try to model the last two sequences. There exist a large series of web-based protein model programs. Here we cannot go through them all, thus we will focus on just two servers. First Phyre, not because it performs better than the other programs, but because it has a very nice and informative web-interface. Next HHpred and CPHmodels.

To save you time, we have submitted the two sequences (Query2 and Query3) to the three servers To see the output click on the following links




Spend some time looking at the results and make sure you understand what is going on. The output from CPHModels is fare from intuitive, so you might what to focus of the two other methods when trying to understand what is going on. Note that the CPHmodels and Phyre servers provide full atom model(s) for the query sequence based on single template modeling. HHpred does only provide templates and alignments.

  • Q11 Try to classify the queries into CM/FR (hard), and NF (difficult/close to impossible)? As a rule of thumb, a CPHModels Z-score 10 will indicate a correct fold, the other servers provide E and P-values.
  • Q12 What template do the Phyre and CPHmodels servers find for the hard (CM/FR) query?

Save the top scoring model from the CPHmodels (it only gives you one model) and Phyre servers. Try to superimpose the two models using Pymol. This you can do by uploading the two files to Pymol, and use the align command

align CPHmodel.pdb,Phyremodel.pdb

where CPHmodel.pdb, Phyremodel.pdb are the structures containing the model predicted by CPHmodels and Phyre, respectively.

As you can see the two models are different but clearly structural imposable.

  • Q12.1 Where on the HHpred template list do you find the best scoring Phyre template?

As you probably found, the HHpred server does not agree with the Phyre and CPHmodel servers in what is the best scoring template. This lead to the concept of meta-servers and multiple template modeling. Like in all other prediction games, you can often get a better idea of an answer to a question by asking the question to many different prediction servers. The list of public protein modeling servers is long. You can submit a query to a list of servers using the META-server. On the server you can submit a query sequence to a list of protein model prediction server simultaneously. The calculation takes some time, so we have submitted the sequences to the server beforehand. You find the output by clicking on the links Query2 and, Query3.

Check the Meta-server output for each of the two targets (we have left out Query 1). A powerful ways to combine the output from many prediction servers, is to extract a consensus prediction. The 3D-jury program does this. The 3d-jury calculates a score for the models predicted by the different servers, and reports a jury score. A value above 50 means a significant model. On the Meta server output webside you can select the set of servers you would like display (use the right selection window, and use Ctrl to select multiple methods). If you have time, play a bit with the settings and compare the predictive results.

  • Q13 Can any of the two sequences (Query2 and Query3) be modeled with a significant hit (Jscore > 40)?
  • Q14 Does the best hit found by the 3D-JURY, CPHModels, Phyre and HHpred methods agree (i.e. same name or function)?
  • Q15 How do the top scoring jury hits share SCOP class?
  • Q16 Can you classify the two sequences into hard (CM/FR) and difficult (NF)?
  • Q17 Can you come up with a good guess for a template, SCOP class and function for the CM/FR query?

Now you probably found that the fold of one to the difficult query sequences could be found using the jury approach where many different protein structure prediction servers are combined. The significance of hit you found and the corresponding SCOP class was very high even though none of the individual prediction servers could come up with (significant) hits.

And now a final philosophical question. What do you do if you cannot find a template?


Now you have seen a real life protein fold recognition experiment. All the servers included in the Mete-server believe them self to be among the best in the world. Read about the Livebench project. Why is it important to assess fold recognition servers in this way? Click on the Set Livebench-2008.2, and see which methods performed the best.

Further reading

For more practical advice on FR and structure prediction in general, including the issue of domain assignment and non-globular regions which we have not had time to cover here, see Rob Russell's comprehensive Guide to Structure Prediction.

That's it for now!