De Novo Design of Metamorphic α-Helical Hairpins

Anfinsen's dogma posits that a single amino acid sequence dictates one, energetically stable protein structure. Metamorphic proteins challenge the "one-fold-one-structure" hypothesis by folding into and converting between two or more native states with divergent structures and functions. They comprise up to 4% of structurally resolved proteins, however experimental identification is highly challenging due to dynamic state interconversion. Rapid advances in new machine learning (ML) tools are well suited to designing metamorphic structures in silico. This project produced a de novo, two-state helical hairpin module by applying an existing computational protein design pipeline. We generated 192 backbones per state with conditional RFDiffusion, 100 sequences and structures with multistate ProteinMPNN (PMPNN) and template-based AlphaFold2 respectively, which were validated on 8 natural metamorphic proteins. This pipeline ran twice with different degrees of partial diffusion to improve backbone stability and returned 7 and 45 designs which passed RMSD, plDDT and sequence diversity filters out of 19,200 structures per run. Alphafold3 predicted only one state for the passing designs, likely due to relative stability bias. To control the state switching, we designed helical peptide binders from the alpha helix facet which Alphafold3 predicted successfully induced switching to opposing states.