Accuracies of protein structure predictors have improved quite a lot in recent years. Algorithms such as Rosetta have gotten robust enough to predict structures of large number of proteins. However, I could't find any initiative to make a database for proteome-wide predictions of protein structures or may be integrate the predicted structures in existing databases (PDB or Uniprot).

If any such exists, please let me know.

  • $\begingroup$ So, your question is if there is a database of predicted structures? (What would be the utility of this database? How would it be up to date/updated?) $\endgroup$
    – llrs
    Jan 25 '19 at 10:05
  • 1
    $\begingroup$ Yes. There could be many uses of predicted structures of all the proteins in the proteome, given that the predictions are accurate enough. One could use that resource to make proteome wide inferences about protein structures, which we can not do now with structures of fewer than ~30% of the proteins available. In terms of advantage, the users would not have to run the prediction model themselves, so it would save time and effort for them. $\endgroup$
    – Ramirez
    Jan 25 '19 at 14:23
  • 1
    $\begingroup$ Regarding updation of the database, I would imagine that would a minor task if there is a prediction model available. Its just about rerunning the updated prediction model again. $\endgroup$
    – Ramirez
    Jan 25 '19 at 14:23
  • 1
    $\begingroup$ I'm not sure that the accuracy is "good enough". Enough for what? Also the folding of a protein changes with the environment. If it seems so easy and no one have done it there might be some "hidden" problems on it. Perhaps the computational cost of those predictions is too expensive for any single (or a consortium) group to try. $\endgroup$
    – llrs
    Jan 25 '19 at 15:05

There are three initiatives I know of to have a go at this:

  1. PConsFam, which collects data from this paper.
  2. The Baker group's metagenomic study which you mention in the question.
  3. The recent DMPfold work from our lab. This compares its results to the above two studies and discusses the effect on various model organisms.

The second and third of these don't provide a database per se, but the models are available for download and are linked to the Pfam database.

  • $\begingroup$ DMPfold looks very promising. Hoping that the work gets published soon so we could get to use the data. $\endgroup$
    – Ramirez
    May 24 '19 at 19:58
  • 1
    $\begingroup$ Thanks, it's currently under review. Though all the data is already available, see github.com/psipred/DMPfold#data. $\endgroup$
    – jgreener
    May 25 '19 at 21:01
  • 1
    $\begingroup$ The paper is now published, see nature.com/articles/s41467-019-11994-0. $\endgroup$
    – jgreener
    Sep 10 '19 at 11:10

By far the most accurate model right now should be AlphaFold, and there are open structures across several proteomes available at https://alphafold.ebi.ac.uk/.


I am using DMPfold to predict protein structure of a 382 residue protein sequence. All goes wel, and even the files are produced but during the "seq2maps.csh", I get this error in between:

Running PSI-BLAST with sequence c600m2-tail-before-lysin-sequence.temp.fasta ...
Segmentation fault (core dumped)
[blastpgp] WARNING: Unable to open BLOSUM62
[blastpgp] WARNING: BlastScoreBlkMatFill returned non-zero status
[blastpgp] WARNING: SetUpBlastSearch failed.
BLASTP 2.2.17 [Aug-26-2007]

Reference: Altschul, Stephen F., Thomas L. Madden, Alejandro A. Schaffer, 
Jinghui Zhang, Zheng Zhang, Webb Miller, and David J. Lipman (1997), 
"Gapped BLAST and PSI-BLAST: a new generation of protein database search
programs",  Nucleic Acids Res. 25:3389-3402.

Reference for composition-based statistics:
Schaffer, Alejandro A., L. Aravind, Thomas L. Madden,
Sergei Shavirin, John L. Spouge, Yuri I. Wolf,  
Eugene V. Koonin, and Stephen F. Altschul (2001), 
"Improving the accuracy of PSI-BLAST protein database searches with 
composition-based statistics and other refinements",  Nucleic Acids Res. 29:2994-3005.

Query= C600M2_00100 tail protein [Salmonella phage BPS17S6]
         (377 letters)

Database: c600m2-tail-before-lysin-sequence.a3m 
           46,978 sequences; 11,323,713 total letters

SearchingFATAL: Error whilst running blastpgp - script terminated!
Running PSICOV
Running FreeContact
Running CCMpred
 _____ _____ _____               _ 
|     |     |     |___ ___ ___ _| |
|   --|   --| | | | . |  _| -_| . |
|_____|_____|_|_|_|  _|_| |___|___|

using CPU (10 thread(s))

Reweighted 30404 sequences with threshold 0.8 to Beff=22917.9 weight mean=0.753779, min=0.0232558, max=1

Will optimize 62686429 32-bit variables

iter    eval    f(x)        ║x║         ║g║         step
1       1       1.50338e+07 152357      3.021629e+11    1.55e-06
2       1       1.47369e+07 152356      2.1447226e+11   1.18e-06
3       1       1.44292e+07 152355      1.4646483e+11   1.22e-06
4       1       1.41268e+07 152353      1.0388224e+11   1.5e-06
5       1       1.38295e+07 152358      7.337667e+10    1.92e-06

Can you please let me know if I need to rerun seq2maps.csh because of this error? And, how this error can be solved? Thank you!



As of 2021, the EBI alphaFold2 is the best resource. There are some caveats, which I have outlined in this blog post.

But majorly, it contains only monomeric predictions and not oligomers and complexes. As a result, for protein with a close homologue with a solved structure, it may be more handy to use a threaded structure as found in the SwissModel database (see original answer below) because it may be oligomeric and/or one can copy over ligands or binding partners if deemed suitable.

One can make one's own predictions with AlphaFold2, who have released a handy notebook. Also one can make complexes and oligomers with some tweaks or with the heavily improved notebooks found in ColabFold GitHub repo.

Original Answer

It is worth saying that for model organisms, Expasy Swissmodel is the "classical" database of predicted proteome structures. It has several model organisms and gets updates regularly. The models are threaded only and the algorithm fell behind some time ago —their website however is modern and sleek. If using Rosetta afterwards, one should do a few cycles of FastRelax with a movemap preventing backbone movements or CA restrained, followed by a couple of cartesian bond/angle fixing cycles: the structure tends to blow up if energy minimised otherwise.

However, if one had a specific protein in mind predicting with RosettaCM, the I-Tasser server or Phyre server is by the recommended option. And actually, the Phyre server has a dataset of the precomputed human proteome.

About the PDB, the database has been purged of all in silico models and contains only empirical structures (NMR, X-ray, neutron diffraction and cryo-EM).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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