Question: How do reseek deals with discontinuous domains during protein alignment?

[camilogarciabotero]

Hi Robert.

Thank you for developing reseek.

I am working with a set of protein domain structures, some of which are sequence-discontinuous (i.e., composed of multiple non-contiguous sequence segments that form a single compact structural domain).

When using reseek to search these structures against a reference set of domain structures, I was wondering:
Does reseek impose sequence-like constraints during alignment (i.e., direct mapping of residue indices), or can it align structurally coherent regions even if they correspond to disjoint sequence intervals?

In practice, would a discontinuous domain be recovered as a single aligned region if structurally compact, or could it be fragmented due to sequence discontinuity?

Thank you very much in advance.

[rcedgar]

Reseek makes a local pair-wise alignment between contiguous segments of the primary chains. TM-align and Foldseek also do this, DALI (I think, not sure...) can align multiple segments out of order. If a domain is discontinuous due to a reasonably short insertion, it's possible that Reseek might continue the alignment but there is nothing in the algorithm which is designed to account for discontinuous domains. I believe discontinuous domains are very rare, more common are circular permutations. Currently, Reseek is not designed to handle circular permutations either, but this is on my list for the next major release.

You could hack both discontinuous and circular permutations by re-aligning -- delete the segment aligned in the first round and try again.

Do you have a particular case in mind? Then the easiest thing to do is to try with and without the hack and see if it works.

If you are interested in investigating discontinuous domains in general, then I'd be interested in working with you on this; the first step would be for you to collect some examples, as many as you can find with reasonable effort.

[camilogarciabotero]

Thanks for the response:

I do not have a particular case in mind yet. I am analyzing a set of domains predicted by a segmentation method and found a considerable proportion of discontinuous ones:

I think its a reasonable sample size to test the hack you suggested to see reseek's behaviour. In a preliminary all-vs-all alignment analysis after doing some quality filtering to the above domains (~13K passed), I observed an approximate 10:1:1 ratio of clusters composed exclusively of continuous domains, exclusively of discontinuous domains, and mixed clusters containing both.

I am not specifically focused on discontinuous domains as a primary research question at this stage, but I would definitely be open to discussing it further if it turns out to have methodological implications.

[rcedgar]

The null hypothesis here is a 11,911/(73,038+11,911) = 14% false-discovery rate of non-continuous predictions by the segmentation method, but it's not clear to me exactly how you can define a "non-contiguous domain" except by alignment to a pre-defined contiguous domain which is a bit of a circular problem. What is exactly is a "domain" in this context? The usual definition is an independently folding unit which is found in enough proteins to infer homology. If it's not contiguous, then how do you adjust the "independent folding" requirement? How could you verify that these are valid predictions in principle? Happy to discuss further, if so I suggest email or slack would be better than GitHub issue, you can contact me via https://drive5.com/contact.html.