Models of structures deposited in the Protein Data Bank vary in the quality, depending both on the data quality and expertise and patience of the person who built the model. Is there a well-accepted subset of the PDB entries that has only "high quality" structures? Ideally these structures would be representative for classes of proteins in the whole PDB.

based on a real question from biology.SE


There is a very nice database, pdbcull (also known as the PISCES server in the literature). It filters the PDB for high resolution and reduced sequence identity. It also seems to be updated regularly. Depending on the cut-offs, you get between 3000 and 35000 structures.

If you are specifically interested in rotamers, you may want to look at top8000 instead, where they have checked for high resolution, and good MolProbity scores. They also provide a rotamer database.

PDB also provides their own clustering. They first cluster the sequences, and then extract a representative structure for each one, based on the quality factor (1/resolution - R_value). This has the advantage of being comprehensive, but you will have bad structures when no good ones were ever obtained.


If you choose to perform your own culling of the PDB, resolution is probably the first thing you'll want to look at, which as Davidmh mentions is the main selection criteria for PISCES. High quality structures will also have better R-factor values. You can also give preference based on experimental technique, in descending order of quality:

Neutron diffraction, X-ray diffraction, solution/solid state NMR, electron microscopy/crystallography, fiber diffraction, solution scattering.

  • $\begingroup$ yes, these are criteria from PISCES, although they look to be quite simplistic. R-factor is used but the gap between R-factor and Rfree is ignored. Resolution (pressumably d_min) is the only criterium of data quality (data completeness is ignored). No geometry validation. $\endgroup$
    – marcin
    Jun 15 '17 at 8:45
  • $\begingroup$ R-factor and R-free discrepancy is a great thing to look at, but in my experience the number of structures reporting both in a standardized way is rather small. BioJava only recently made them both available, IIRC, and that's typically the tool I use. $\endgroup$ Jun 15 '17 at 20:24
  • $\begingroup$ 80.7% of PDB structures reports R-free, slightly less than 86.8% that reports resolution. $\endgroup$
    – marcin
    Jun 16 '17 at 13:49

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.