PairSAE: Mechanistic Interpretability in Protein Co-Folding

Introduces PairSAE, a method for mechanistic interpretability of protein co-folding models by extracting pair representations from backbone features. Aims to understand internal features driving outputs of structural biology foundation models.