How to handle input pdb file with [UNK] token ?

[Eikor]

Dear researchers,

Thank you for your amazing work!

I have encountered a small problem when running on my custom dataset.
The PDB file is as follows:

ATOM     15  CB  SER A   4      50.214  -9.038  18.043  1.00  0.00           C  
ATOM     16  OG  SER A   4      49.838 -10.302  18.781  1.00  0.00           O  
ATOM     17  N   UNK A   5      49.409  -6.273  19.161  1.00  0.00           N  
ATOM     18  CA  UNK A   5      49.233  -5.330  20.255  1.00  0.00           C  
ATOM     19  C   UNK A   5      48.099  -4.517  19.988  1.00  0.00           C  
ATOM     20  O   UNK A   5      47.425  -4.219  20.898  1.00  0.00           O  
ATOM     21  CB  UNK A   5      50.344  -4.350  20.407  1.00  0.00           C  
ATOM     22  CG  UNK A   5      49.845  -3.031  21.133  1.00  0.00           C  
ATOM     23  N   ASP A   6      48.053  -4.242  18.702  1.00  0.00           N  
ATOM     24  CA  ASP A   6      47.037  -3.448  18.051  1.00  0.00           C  

There is a [UNK] residue type in my PDB file, and this residue will be encoded as id 20. However, this will trigger an error of ResidueTypeSeqFeat.

How should I process the input if there are unknown residues in my PDB?
Am I supposed to set the unknown residue to a padding token?

Thanks!

Best,
JD

[tgeffnerNVIDIA]

Unfortunately, La-Proteina does not support UNK tokens yet. We are working on extending La-Proteina to more tasks, and handling residue types outside the 20 standard ones (or UNK) is on the list. For the time being, to work with such PDB files you'd need to adapt the codebase a bit. Happy to provide some guidance if you plan to do so. Briefly, this would require a change to the FeatureFactory (exactly the error you are getting). If you only plan to add the UNK token, should be somewhat straightforward. And of course, once the code is changed you should fine-tune / train from scratch some models with the updated code.