Abstract

We have derived an all-scale graph partitioning approach that preserves atomistic physico-chemical detail and by using diffusive processes on the graph (both on the node and the edge space), we have shown that we can obtain the behaviour of biomolecules and bimolecular assemblies at different timescales without the need of any reparametrisation or a priori selection of relevant timescales. The approach is computationally efficient and general and can be applied to molecules, molecular assemblies as well as data. I will discuss the theory that brings together graph theory, dynamics and diffusive processes and showcase it with examples from predictions and experimental verification of mutations that control protein dynamics at different scales (AdK), prediction of allosteric sites for drug design and communication and signalling in multimers and assemblies (ATCase, Rubisco). Finally, the application of this unsupervised learning approach to trajectories and data will be briefly discussed.