search

Graylab / IgFold

Fast, accurate antibody structure prediction from deep learning on massive set of natural antibodies
Other
319 stars 60 forks source link

Corrupt structures #61

Closed lhatsk closed 10 months ago

lhatsk commented 10 months ago

Hi,

I am currently testing IgFold on newer targets from SAbDab, for example, 7ox1 A/B following the instructions in: https://github.com/Graylab/IgFold#antibody-structure-prediction-from-sequence using IgFold 0.4 (also tested 0.3.1).

For most of the targets, I end up with broken predictions, ie, no secondary structures and big clashes (see attachment). The example on GitHub works. Any idea what's going wrong?

Thanks!

sequences = {'H': 'QLQLVESGPGLVKPSQTLSLTCTVSGGSITTGYYAWSWIRQPPGKGLEWMGFIARDGSTSYSPSLKSRTSISRDTSKNQFSLQLSSVTPEDTAVYYCARAGEGRSWYPGYYYGMDYWGKGTLVTVSSASTKGPSVFPLAPSSKSTSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSSSLGTQTYICNVNHKPSNTKVDKKVEPKSC', 'L': 'QAVVTQEPSLSVSPGGTVTLTCGLSSGSVTSSNYPGWYQQTPGQAPRVLIYNTNSRHSGVPSRYSGFISGNKAALTITGAEPEDEADYYCHLHKGSTGVFGGGTHLTVLSQPKAAPSVTLFPPSSEELQANKATLVCLISDFYPGAVTVAWKADSSPVKAGVETTTPSKQSNNKYAASSYLSLTPEQWKSHRSYSCQVTHEGSTVEKTVAPTECS'}

7ox1_igfold

lhatsk commented 10 months ago

I just noticed that the problem is caused by using the full heavy / light chain from the PDB.

jeffreyruffolo commented 10 months ago

Hi, yes that's correct. IgFold expects to receive just the Fv portions of the heavy and light chains.