ESM, max length issues using ESM-gearnet-serial model

[gkwlgns323]

Hi I'm using ESM-gearnet model, to get protein representation

but in ESM module there is max_input_length which slice the graph when the protein size is bigger than this threshold

because of this, the "graph node number" and "residue number of representation from ESM" are different.

Is it impossible to get representation from big proteins? (when I set max_input_length bigger, I got out of memory error...)

In this case should I use sliced residue representation and also sliced graph? I'm worried that it will also reduce the performance.

[Oxer11]

Hi, I think the constraint on the protein size is added by ESM actually. I think the best way to deal with big proteins is to get the representations through sliced windows.

[gkwlgns323]

Thanks for kind answer, using sliced windows will ESM take only front part of the protein? Or is there any way to slice the protein with sliced windows, get representations of each windows, then combine it to achieve the representation of the whole protein?