Extreme amount of clashes in some multimer predictions

[dennissv]

Sometimes when using the multimer models I get extreme amount of clashes in the predictions to the point where the proteins are just a blob. At first I thought it was a problem with my local installation but I tried running some problematic proteins on the Colab notebook and get similar results there. An example would be the following protein sequence as a heptamer;

MEENVTEDSILSTELRGQEIRIEEVPNEDTLVVYLSRGGRNAKESQSANSNNTGDIDKHGRHRADNCICLSKVHYEIDGDILCIREMKPFVYINKVSILRRRSTNKTKEDTETQKSP

The complete disregard for stereochemistry only seem to appear in regions with very low confidence. The only modification I have made is to disable the Amber relaxation step but this never seemed to cause large issues like these in the monomer version nor in the ColabFold multimer version. I also doubt that Amber can resolve these clashes in a meaningful way?

Is this expected behaviour without Amber and for proteins that are hard to model or is something wrong?

[arashnh11]

Which model is this from? And do you have an alignment RMSD w.r.t the experimental?
We are working on an octamer complex and see clashes too. In our case model_3 had the best outcome though still off...

[tony9664]

I'm also seeing some similar serious clashes when predicting multimers. Colabfold was good though.

[arashnh11]

Aren't the colabfold and v2.1.0 from the same source code? Not sure about clashes but an octamer we tried resulted in poor alignment with pdb > 15A

…

On Sat, Nov 6, 2021, 2:30 PM tony9664 ***@***.***> wrote: I'm also seeing some similar serious clashes when predicting multimers. Colabfold was good though. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#236 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AELBF5IUPJNBIHFJT5QOGBDUKV65LANCNFSM5HPBJSGA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[arashnh11]

OK, here is the comparison of clashes for 7KEU: magenta: v2.1.0 green: RCSB.

[dennissv]

@arashnh11 2.1.0 included new multimer models for protein complexes which is where I started observing these issues. ColabFold used a flexible linker solution which didn't provide the same accuracy but also never displayed these clashes.

Some proteins work very well and it's been great seeing these models. But whenever a protein has some region with low confidence it quickly becomes a tangled mess. I've heard several people having the same issues, some examples can be found here; https://twitter.com/RolandDunbrack/status/1457538160561377283

Is there any update on this?

[arashnh11]

@dennissv Do you know of some sample multimers with good alignment with reference PDBs?
Haven't heard any update yet. I can tweet our failed Octamer testcase above, but I believe this is more of a bug...

[tfgg]

We think we may have identified the issue with overlapping domains but need to test a bit more. It will take us a few days at least to be sure.

[arashnh11]

@tfgg Thanks for the update. Please let us know if there is any information we can provide or any additional tests we can do to help. One thing we tested was the Amber relaxation on multiple cases and it seems that it actually increased the clashes for the relaxed model vs unrelaxed, though the general structure was about the same, i.e., no improvements.

[dennissv]

This seems largely fixed with the new v2 multimer models. Thank you!