Machine learning methods have become increasingly central in the development of a large variety of versatile tools for molecular simulations, many of which have the potential to advance significantly the fields of computational chemistry and physics. In this chapter, we present a framework for combining machine learning for local structure classification with the definition of a global classifier space as a basis for enhanced sampling of structural transformations in condensed phase systems. The transformation is represented by a path in classifier space, and the associated path collective variable is used to drive the process derived from changes in local structural motifs. Enhanced sampling along this type of path collective variable yields insight into the physical mechanism as well as corresponding free energy barriers of the transition. The idea is generally applicable, and the approach, as outlined here, can be adapted to a wide range of systems.