In by admin

Name Mr. Christian Devereux
Organization or Institution University of Florida
Presentation Type Poster
Topic Computational Chemistry

Molecular Geometry Optimization and Normal Mode Calculations via ANI-1x Deep Learning Potential


Christian  J. Devereux, Justin S. Smith, Olexandr Isayev, Adrian E. Roitberg

Author Institution(s)

Christian J. Devereux, Justin S. Smith, Adrian E. Roitberg - University of Florida
Olexandr Isayev - University of North Carolina at Chapel Hill


Predicting the way that a molecule binds to a protein is one of the most important problems in designing a new drug. In order to be an effective drug, a molecule must have the proper geometry to access the binding site of the target protein. Once the drug is bound, knowledge of the proteins normal modes of vibration is important in determining the motion of the protein. Typically classical force fields are used to these, but these methods suffer from limited accuracy on systems far from the parametrization set. Quantum mechanical approaches can provide accurate structures and energies, but are too computationally costly to model most medically relevant systems. Machine learning methods offer a bridge between classical and quantum mechanics by building potential energy surfaces empirically trained to quantum mechanical potentials but at speeds similar to classical force fields. A properly developed machine learned potential surface should be able to accurately predict not only energies, but also conservative molecular forces by construction. Given molecular forces, the calculation of normal mode frequencies and geometry optimization become trivial. The machine learning methodology ANAKIN-ME (ANI) is used to build a potential energy surface for several hundred molecules including tripeptides and drugbank database molecules. The potential energy surfaces are then used to calculate the normal mode frequencies and optimized geometries of these molecules and the results are compared to those of DFT. Based on these results, it can be seen that ANI out performs classical and semi empirical methods and has a small error compared to DFT, with a speed comparable to classical methods. ANI accomplishes this without any direct force training and while maintaining a universality in the systems it can be applied to.