Training elastic network models to change conformation

Elastic network models have been used as a coarse-grained model to study proteins’ vibrational spectrum. [1] Despite long-standing interest, foundational understanding of what’s behind protein allosteric behavior still is lacking. Diverse material properties can be achieved in physical networks whose interactions can be tuned. By iteratively tuning degrees of freedom regulating these interactions (e.g. individual spring constants in a mechanical network of nodes connected by springs), one can minimize deviations from the desired material property while also minimizing a physical Lyapunov function (e.g. mechanical energy). This dual minimization process affects the physical response to perturbations, leaving imprints of the learning procedure in the curvatures around the energy minimum [2,3]. We train elastic network models (ENMs) of unbound allosteric proteins to fold into their bound conformation and relate the physical imprints of learning to residue conservation and the effect of point mutations.