Token for unaligned amino acids

[VGPReys]

Hi there,

I would like to generate embedding for a protein sequence that comes from a multiple sequence alignment (e.g: "AAA--AAA--A")

What kind of token should be used to describe the unaligned amino acids ?

My initial guess is that is could be one of these:

• .
• -
• <pad>
• <mask>

But which one ?

Thanks for your help

[thomas-a-neil]

Why not embed the unaligned sequence?

[VGPReys]

to be able to use the information contained in the msa

[ebetica]

ESMC was not trained for aligned sequences. I would embed the unaligned sequences and do a post-hoc alignment of the embeddings.