This is a test version of Biostars. For the public version, visit https://www.biostars.org.
Protein structure to classify domain sequences into full Vs. partial lengths and false positives

Research problem - Using protein structure to verify and classify protein domain predictions from HMMER.

Dataset: I have a test dataset of ~ 800 protein domain sequences that are 30-110aa in length. Most of them are ~ 40-45aa long. The shortest (~30aa) and longest sequences (~110aa) are extreme variations, but there are many many predicted sequences that are either shorter or longer than expected, and which look quite doubtful. The full dataset is much larger at ~17K domain predictions.

Proposed Solution: Therefore, I want to use (2D or 3D) protein structure to "verify" and "classify" these domain sequence predictions into

  • full-length domains,
  • partial-length domains, i.e. truncations, and
  • false positive predictions (that do NOT match the canonical 2D and/or 3D details for this protein domain)

Structure information: This domain is found in 7 PDBs at RCSB, and when trimmed to just the domain boundaries and superimposed in 3D space, they overlap quite well, even though pairwise sequence identity can be as low as ~ 17%. This is not surprising since structure is much more conserved than sequence. But it reinforces my motivation to explore using "protein structure" to verify and classify my domain predictions based on "protein sequence" methods.

Questions: to forum members are as follows:

1. Is 2D prediction for my domain sequences sufficient for the domain verification and classification I want to carry out?

  • If yes, then which SS prediction tool is suggested (PsiPred, DSSP, something else?)
  • How do I parse those results, and
  • What do I compare such parsed results to, would it be to SS prediction for the 7 solved PDBs? All of them, one of them, some consensus SS?

2. If 2D info is necessary but not sufficient, and 3D info is required, then what tool(s) should I use?

  • Is there any Deep Learning or Machine Learning software than can do this? Al Quiraishi @ Harvard said his very recent RGN/ProteinNet is not suitable since I do have templates to check against, and
  • The I-TASSER people at U-Mich said this is computationally too intensive, even though my inputs are not full-length proteins, but just shorter domain sequences -to be fair my inquiry was to analyze the full dataset, not just the test dataset.

Based on responses from RGN and I-TASSER groups, I've started thinking whether 2D rather than 3D prediction would provide an acceptable solution to my problem?

I look forward to your answers to these questions, as well as orthogonal thinking to solve this research problem. Thanks, in advance!

structure domain 3d sequence protein
• 224 views • link
updated by Mensur Dlakic Mensur Dlakic • written by Anand Rao Anand Rao

2 answers

I'm not completely clear what you are trying to do, but I would consider using the alignment with sequences or HMMs to check the domain boundaries rather than using structure prediction. This will be more sensitive than secondary structure prediction, which uses local sequence features and is less accurate in single sequence mode (I guess how you would use it). Side note, DSSP assigns secondary structure from tertiary structure but does not predict it.

Tertiary structure prediction is also probably not suitable - if templates are available, which it sounds like they are, then the best current methods use HMM searches anyway to find templates so that is effectively what you would be doing. If templates are not available then structure prediction is less accurate and effectively requires an alignment as input, so you end up predicting a sort of 'average' for the family which is not what you want here.

Remember there are also databases of structural domains (e.g. CATH) and sequence families with HMMs (Pfam) that could be useful here.

• 0 views • link
written by jgreener jgreener

Thanks for your response. I'll reiterate what I want to do in very simplistic language - "find out if and how prediction of structure for a protein domain can verify if its sequence is really the domain that HMMER3 says it is"

One suggestion was to use hhblits to identify best aligned PDB chain to each domain sequence, and use that as template for MODELLER - I can perform those hhblits steps.

SPECIFIC HELP I SEEK But before embarking on the MODELLER step, I am curious to know how I would use MODELLER output to help "verify" and "classify" my predicted domains into true positives vs. false positives. And how would I use MODELLER results to further sub-classify the predicted domains into

  • full-length and functional,
  • partial-length and functional, and
  • too truncated/degenerated to be functional.

Bottomline - Does MODELLER produce relevant details in the output file(s) and in a parsable format, so I can parse it to help validate / classify my domain predictions. If yes, then what file do I parse, and what criteria do I use for my classification?

Another suggestion, along these same lines, was to use LOMETS, which is part of I-TASSER This approach also identifies the best PDB template for threading. But it uses 11 different, independent threading programs, which is probably overkill for me.

Any follow-up thoughts / advice for me? Thanks!

• 0 views • link
written by Anand Rao Anand Rao

A short answer to your question, especially as it relates to 17K proteins, is that no amount of sequence or structure analysis will conclusively answer everything you want to know. For example, I don't know of any predictive method that can conclusively answer for every single protein domain whether they are full-length and functional or full-length but non-functional, or partial but functional or partial but non-functional. Too much domain expertise and case-by-case studies is required for such an undertaking, and it still would not work 100% without some kind of experimental verification.

The best one can do is provide a reasonable argument, based on sequence and structure considerations, that a given sequence contains the F-box domain or not. Though there will be exceptions, it is usually safe to assume that HMMer-identified sequences that match the model with E-values of 1e-5 or lower are safe bets to have the domain. That will hopefully take care of most of your sequences. Borderline cases can be submitted to HHpred (https://toolkit.tuebingen.mpg.de/#/tools/hhpred) or other servers for additional verification.

I doubt that you have the capacity or the inclination to create 3D models for 17K proteins, and it is probably an overkill anyway. But again, it may be worth modeling some borderline cases. It is widely accepted that one can't get a good comparative model if a protein sequence has a different fold from the template. By that I mean the model will be flawed energetically, even though its fold may look "pretty" and it may seem correct stereochemically. [EDIT: Therefore, if you can't build a good model for a given sequences, there is some argument that your sequence does not have the same fold as your template structure. One should not take this literally: poor models can also be caused by poor sequence-template alignments, which happens in cases of low sequence-template identity.]

I am partial to ProSA, but there are many ways to test the quality of models:

https://prosa.services.came.sbg.ac.at/prosa.php

http://servicesn.mbi.ucla.edu/SAVES/

http://servicesn.mbi.ucla.edu/Verify3D/

http://molprobity.biochem.duke.edu/

https://proq.bioinfo.se/ProQ/ProQ.html

https://swissmodel.expasy.org/qmean/help

http://melolab.org/anolea/

Again, I would recommend 3D modeling only for borderline cases and not for 17K proteins, as many modeling and evaluation steps are difficult to automate.

• 0 views • link
written by Mensur Dlakic Mensur Dlakic

Log in to answer this question.

More posts like this