Fig. 5: Stable, long-timescale molecular dynamics (MD) simulations and extrapolation to larger bio-molecules.
From: A Euclidean transformer for fast and stable machine learned force fields
a Stability and speed of SO3krates enable nanosecond-long MD simulations for supra-molecular structures within a few hours. For the buckyball catcher, the ball stays in the catcher over the full simulation time of 20 ns, illustrating that the model successfully picks up on weak, non-covalent bonding. b Distribution of the radius of gyration (ROG) for Ac-Ala3-NHMe as a function of MD simulation time in 20 ns steps. The distribution converges after 60–80 ns simulation time, underlining the importance of stable, but at the same time computationally efficient, simulations. In Supplementary Fig. 8 the ROG as a function of simulation time is shown. c Dynamics of Ala15 obtained from a SO3krates model, trained on only 1k data points of a much smaller peptide (Ac-Ala3-NHMe). Analysis of the end-to-end difference shows rapid folding into helical structure illustrating the generalization capabilities of the learned local representations towards conformational changes on length-scales greatly exceeding the training data (dashed gray line).
