I have a question related to the evaluation of variant pathogenicity in the context of human genetics. There are many predictors available, however, is there a specific predictor that takes into account the 3D protein structure as well?

Imagine there is a mutational hotspot, i.e. a sequence of amino acids of a protein region that are pathogenic (assessed by functional studies, for example). And then, let's say several hundred residues far from this region there is a variant whose pathogenicity status is unclear. Now, what can happen here is that this variant, although far from the mutational hotspot (considering the protein secondary structure) can actually be quite close to the mutational hotspot in the 3D space. I mean, this variant may actually form a part of a functional site, together with residues from the mutational hotspot. This could indicate that this single variant is pathogenic as well, although there is no direct evidence.

And so, is there is a predictor that takes 3D structure into account?

In other words, I'd like to ask a question: Are there any pathogenic variants (defined by clinvar) within a sphere with a radius of 100A centered on a particular residue of interest?

Or, are there, for example, any glycine residues within such a sphere?

It's not really a prediction, it would be just a search algorithm, taking 3 parameters:

  • protein 3D shape
  • search sphere [center residue; radius]
  • needle - the element I'm searching for, such as AA residue type, for example
  • $\begingroup$ Please edit the question to limit it to a specific problem with enough detail to identify an adequate answer. $\endgroup$
    – Community Bot
    Feb 7, 2022 at 10:38

1 Answer 1



I am the developer of venus.cmd.ox.ac.uk, which is a site that given a mutation makes the variant (and gives the ∆∆G) and shows what residues are nearby and whether these appear in clinVar or gnomAD (and scores them too) or have some annotation such as disulfide or post-translational modifications. The aim is to allow the user to see and judge for themselves what is going on as there are multiple possible mechanisms which may make a variant different (and then saving and sharing the results via Michelanglo). In other words it is not a predictor but a data aggregator of soughts. It shows, in addition to the structure in the NGL viewport, a table where the nearby residues are listed (and if clicked focuses on in the protein)


An issue according to non-users is that it does not give a statistical relevance score for the features, but I really did not want that as these can be completely off and easy to manipulate and just encourage users to blindly believe the calculations.

An issue according to me, is that it is based on empirical data: if nothing is known to bind or where it binds, it cannot say much, unlike MutPred2, a linear sequence deep-learning tool that assigns a prediction of whether there's a binding site etc. without prior empirical evident.

Very very few users provide a PDB file and this aspect is limited. Venus is that it is centred around canonical Uniprot numbering, so providing PDB files with differing numbering will not work. Furthermore, the app is coded to "give up easily" on a structure and pick the next option, so does not try to parameterise unusual ligands in a provided structure (I notice now a job failure due to the ACE C-terminal cap, for example).

Other options

There are other tools that show the variant in a structural context, such as Missense3D (worth a gander), but do not show other variants or annotations.

However, for a specific query, such as the closest proline to the residue of interest in a structure with no relevant Uniprot, it may be best to simply use a viewer, such as Mol* or a local one, such PyMOL.


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.