Original proteingym MSAs contain letter codes not supported by the current MSAProteinSequence class

[florisvdf]

Summary

There are a significant number of non-canonical or ambiguity amino acid letter codes in the MSAs provided by the original version of the proteingym. There are:

• B (Asn/Asp): 657 occurrences in 75 files (52% of files)
• U (Selenocysteine): 6 occurrences in 3 files (2% of files)
• Z (Gln/Glu): 210 occurrences in 55 files (38% of files)

Though the percentage of sequences with at least one of those letters in all sequences across MSAs is extremely small, something like less than 0.0003%, so they are definitely edge cases.

We should decide whether to support these, as well as other possible letters like:

Motivation

We can't create pgdata archives for a large number of datasets from the original proteingym if we want to use the original MSAs without modifying them. If we use the same MSA generation protocol as was done for the original proteingym, more of such MSAs could be generated. The generation protocol should also be able to inform us which potential other letter codes we could encounter. Users may benefit from an expanded alphabet if they want to create datasets with their custom MSAs.

Do we:

• Expand the alphabet, and with which letters?
• Keep the alphabet as is?

Proposed solution

Create an ADR and make a decision.

[florisvdf]

Some pros and cons for expanding the alphabet off the top of my head

Pros

• More flexibility for our data repository. We can reuse unmodified existing MSAs and keep using the same generation protocol.
• More flexibility for the user that wants to create their own dataset with custom MSAs.

Cons

• Some existing MSA based models may not support non-standard amino acid letter codes (I didn't verify this).
• Blurring of the scope: If MSAs will support non-standard aas, what about other sequences, like assay sequences or references? Why? Why not? If yes, should all models then in theory accept inputs with non-standard amino acids?

[florisvdf]

From our discussion it seems that indeed some models assume the standard 20 letter amino acid alphabet, and having pgdata datasets with MSAs that contain non-standard letter codes will cause those models to fail. From the stats posted in the issue, despite the low overall occurrence of non-standard sequences, the odds of them appearing at least once in a file are pretty high.

Henning proposed adding a validation step during dataset creation that flags and drops non-standard sequences. One option is to do this at retrieval time, i.e. when referencing msa.value.