Predictions in peptide-protein interfaces are problematic

[josemduarte]

Issue by josemduarte
Monday Jun 23, 2014 at 10:29 GMT
Originally opened as https://github.com/eppic-team/eppic-cli/issues/9

---

[issue transferred from JIRA]

In a peptide-to-protein interface usually the prediction is alright for the protein side, but for the peptide side there are problems:

• homologs are difficult to catch with blast (even when using PAM30 matrix as we do): many times we end up with 0 homologs
• in core-surface score: there's not enough free surface to have proper sampling (this commit 593f669 should have helped)
• in core-rim: many times there's not even a rim: all residues are fully buried

Some examples:

• 3bfq: this is FimG-FimF complex, good prediction based on the protein side only
• 1qsn: a case of 0 homologs for peptide side, even using PAM30 matrix
• 1f4v: quite good signal from peptide side but number of core residues fall below cutoff for nopred. The peptide-protein interface is repeated 3 times in this structure (interface cluster 2, composed of interfaces 2,3,4) and for only 1 of the 3 there is a prediction
• 1w0w: a case of 0 homogs for peptide side, the protein side is not well predicted either because of separate issue with MHC sequences
• 1sdx: too few core residues in both sides
• 1w9q: the UniProt reference was not even found

A possible idea is to use some kind of geometric approach to detect these kind of interfaces, for instance the BSA/total ASA ratio should be quite high. What to do after to provide a meaningful evolutionary decision is not very clear anyway...

[sbliven]

Can we just accept these as weak predictions?

[josemduarte]

Realistically accepting them as weak predictions is about the best we can do for now. Let's see if we can do something in 3.1 or later.