Translating patched counterfactual structures to viable sequences

Develop a method to translate counterfactual protein structures produced by activation patching within the ESMFold folding trunk into viable amino acid sequences that fold to those structures, overcoming current failures of inverse folding on these outputs.

Background

The study uses activation patching to transplant latent representations between proteins, generating counterfactual predicted structures that can exhibit targeted structural motifs. However, these patched structures often have low confidence and are not readily supported by sequences found via inverse folding.

The authors explicitly state that mapping such patched structures back to viable sequences remains an open problem, highlighting a gap between controllable structural manipulation in latent space and practical, physically plausible sequence design.