Topology representing neural networks reconcile biomolecular shape, structure, and dynamics

Topology-representing networks (TRNs) generate reduced models of biomolecules and thereby facilitate the fitting of molecular fragments into large macromolecular complexes. The components of such complexes undergo a wide range of motions, and shapes observed at low resolution often deviate from the known atomic structures. What is required for the modeling of such motions is a combination of global shape constraints based on the low-resolution data with a local modeling of atomic interactions. We present a novel Motion Capture Network that freezes inessential degrees of freedom to maintain the stereochemistry of an atomic model. TRN-based deformable models retain much of the mechanical properties of biological macromolecules. The elastic models yield a decomposition of the predicted motion into vibrational normal modes and are amenable to interactive manipulation with haptic rendering software.