Do protein language models store different structural elements in factorizable subnetworks?

To find out, we masked out PLM weights to suppress performance on CATH subcategories or secondary structure elements while maintaining performance on other sequences or residues.
Post image

Comments