Descriptor-Free Collective Variables from Geometric Graph Neural Networks

Enhanced\nsampling simulations make the computational study of rare\nevents feasible. A large family of such methods crucially depends\non the definition of some collective variables (CVs) that could provide\na low-dimensional representation of the relevant physics of the process.\nRecently, many methods have been proposed to semiautomatize the CV\ndesign by using machine learning tools to learn the variables directly\nfrom the simulation data. However, most methods are based on feedforward\nneural networks and require some user-defined physical descriptors.\nHere, we propose bypassing this step using a graph neural network\nto directly use the atomic coordinates as input for the CV model.\nThis way, we achieve a fully automatic approach to CV determination\nthat provides variables invariant under the relevant symmetries, especially\nthe permutational one. Furthermore, we provide different analysis\ntools to favor the physical interpretation of the final CV. We prove\nthe robustness of our approach using different methods from the literature\nfor the optimization of the CV, and we prove its efficacy on several\nsystems, including a small peptide, an ion dissociation in explicit\nsolvent, and a simple chemical reaction.