AbstractMolecular dynamics (MD) simulations have emerged to become the back-bone of today’s computational biophysics. Simulation tools such as, NAMD, AMBER and GROMACS have accumulated more than 100,000 users. Despite this remarkable success, now also bolstered by compatibility with graphics processor units (GPUs) and exascale computers, even the most scalable simulations cannot access biologically relevant timescales - the number of numerical integration steps necessary for solving differential equations in a million-to-billion-dimensional space is computationally in-tractable. Recent advancements in Deep Learning has made it such that patterns can be found in high dimensional data. In addition, Deep Learning have also been used for simulating physical dynamics. Here, we utilize LSTMs in order to predict future molecular dynamics from current and previous timesteps, and examine how this physics-guided learning can benefit researchers in computational biophysics. In particular, we test fully connected Feed-forward Neural Networks, Recurrent Neural Networks with LSTM / GRU memory cells with TensorFlow and PyTorch frame-works trained on data from NAMD simulations to predict conformational transitions on two different biological systems. We find that non-equilibrium MD is easier to train and performance improves under the assumption that each atom is independent of all other atoms in the system. Our study represents a case study for high-dimensional data that switches stochastically between fast and slow regimes. Applications of resolving these sets will allow real-world applications in the interpretation of data from Atomic Force Microscopy experiments.
Detecting outlying subspaces for high-dimensional data: the new task, algorithms, and performance