#36 · Structural bio & protein design

Dynamical data and diffusion models for glycan structure and protein switch design

The dominant thread across these preprints is the incorporation of conformational dynamics or ensemble behavior into machine learning models of biomolecules. Diff-Switch samples switch-like protein backbone ensembles from pretrained diffusion models, using reward-tilting and a metadynamics-inspired bias to target switching behavior. pHaseMD4AI supplies the dynamical substrate: an MD dataset of peptide and constant-pH protein simulations with synchronized coordinates, velocities, forces, and MSM kinetics, intended for physically and kinetically consistent AI. SweetFold is less directly dynamical; it adapts the Boltz-1x all-atom diffusion model to glycan, glycoprotein, and protein-glycan complex structure prediction using pseudo-polymer glycan modeling. What is new is the use of explicit dynamical quantities, such as velocities, forces, and kinetic models, as training data or sampling control rather than static structure alone. The evidence base is early and consists of three preprints.