I have some protein sequences and I need to predict the physical distance between each two residues in the protein when folded correctly in 3 dimensions; I need to predict this distances just by having the protein sequence and not the structure.

I have a group of pdb IDs and I have got homologs for them and made a multiple sequence alignment for each query sequence; But later I need to imagine that I don't have any structure for being able to predict the distance between residues in query sequences without any known structure.

There are some articles for solving this issue using machine learning methods but I don't want to train with a large numbers of structures. Any one knows any way to solve this problem by using mathematical or statistical methods?

  • $\begingroup$ Are there any similar sequences with known structure? $\endgroup$
    – Devon Ryan
    Apr 28, 2018 at 17:40
  • 2
    $\begingroup$ Please add that detail into your question. Also, pelase clarify that you are referring to the physical distance between the residues in the protein when folded correctly in 3 dimensions (at first I thought you meant in the linear sequence). $\endgroup$
    – terdon
    Apr 29, 2018 at 12:51
  • $\begingroup$ Between ALL pairs of residue-residue combinations? You can't use statistics without a large enough dataset and expect reliable results... BTW, I might be outdated but I don't remember having seen de novo folding success after Brian Kuhlman's, 2003 science paper with David Baker. If you only need to compute the distance within short fragments, you could estimate the secondary structure with psipred and use this information to estimate, for example, an average distance between residue i and residue i+4 $\endgroup$
    – aerijman
    Jul 3, 2019 at 16:13
  • $\begingroup$ @aerijman I think you are mixing up de novo protein design and de novo protein structure prediction. Design of novel folds hasn't had much success as you say, but structure prediction of novel folds is getting to a decent level now. See the CASP website for some nice examples. $\endgroup$
    – jgreener
    Jul 4, 2019 at 10:19

1 Answer 1


Developments in this area over the last year are worth pointing out. All use residue-residue covariation information from a protein family so require a multiple sequence alignment to work, though it sounds like you have that.

The most accurate way to do this would be to use pre-trained machine learning methods, which would not require you to train your own version. Examples include:

  • DMPfold from our lab, which is freely-available open source.
  • The RaptorX method from the Xu lab.
  • The state of the art is AlphaFold from DeepMind, though this has not been published yet.

These predict the probabilities of a residue-residue distance being in various distance 'bins', e.g. the probability of being 5 - 6 Angstrom apart. They don't predict beyond ~20 Angstrom.

If you specifically wanted a statistical, non-trained method, then consider:

All of which are available in some form. The problem here is that they all predict binary contacts at 8 Angstrom, whereas it sounds like you need more fine-grained distances.

In a general sense, you can use any 3D structure prediction method and then read the distances back from the predicted structure.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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