Discovering Collective Variables of Molecular Transitions via Genetic Algorithms and Neural Networks

• DOI: 10.1021/acs.jctc.0c00981
• Publication Date: 2021-03-04T22:32:10Z
• Authors: Ferry Hooft; Alberto Pérez de ...; Bernd Ensing
• Keywords: Autoencoder

With the continual improvement of computing hardware and algorithms, simulations have become a powerful tool for understanding all sorts (bio)molecular processes. To handle large simulation data sets to accelerate slow, activated transitions, condensed set descriptors, or collective variables (CVs), is needed discern relevant dynamics that describes molecular process interest. However, proposing an adequate CVs can capture intrinsic reaction coordinate transition often extremely difficult. Here, we present framework find optimal from pool candidates using combination artificial neural networks genetic algorithms. The approach effectively replaces encoder autoencoder network with genes represent latent space, i.e., CVs. Given selection as input, trained recover atom coordinates underlying CV values at points along transition. performance used estimator fitness input Two algorithms optimize architecture. successful retrieval by this illustrated hand two case studies: well-known conformational change in alanine dipeptide molecule more intricate base pair B-DNA classic Watson-Crick pairing alternative Hoogsteen pairing. Key advantages our include following: interpretable CVs, avoiding costly calculation committor time-correlation functions, automatic hyperparameter optimization. In addition, show applying time-delay between output allows enhanced slow variables. Moreover, also be generate configurations unexplored microstates, example, augmentation data.