One approach to calculating configurational kinetics from computer simulations of macromolecules is to place interfaces or milestones across a reaction coordinate manifold. The interfaces serve to subdivide the feature space describing the dynamics of the macromolecule (e.g. phase space) into a set of subspaces, whereby the global kinetics can be calculated from the set of fluxes between subspaces. While defining such manifolds and hyperplanes in subspaces comprised of just a few degrees of freedom, deemed important by the user, may be fairly straightforward, richer, less user-biased kinetic representations involving a larger subset of the staggering number of degrees of freedom encountered in macromolecules, like proteins and DNA, are well outside the capacity of human intuition. We have developed an algorithm that leverages the machine learning methods of Clustering and Support Vector Machines (SVM) to fully automate the placement of the hyperplanes in chemically relevant feature spaces. Additionally, due to SVM being a kernel-based method, our approach provides a framework for mapping the native input spaces, where local minima may not be linearly separable, into higher dimensional feature spaces that allow for linear separation of otherwise non-linearly separable subspaces. Although some methods of this type are exact for an arbitrary set of subspaces, in practice, the calculations will converge fastest, and are most amenable to interpretation, when the interfaces are placed in regions where they can capture transitions between local energetic minima. The algorithm requires no intuition about the system from the user, and places no theoretical limit on the number or type of degrees of freedom.

Using machine learning to automate calculations of conformational kinetics in molecular simulations

255th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.