ANAKIN-ME: Using deep learning to develop a chemically accurate and universal potential for the prediction of organic reactions

Kavindri D. Ranasinghe¹, Justin S. Smith¹, Olexandr Isayev², Adrian E. Roitberg¹

1. Department of Chemistry, University of Florida, Gainesville FL, USA
2. UNC Eshelman School of Pharmacy, University of North Carolina at Chapel Hill, Chapel Hill NC, USA

Organic chemistry serves as the basis for life, making it the backbone in the synthesis of biologically active compounds, pharmaceuticals and therapeutically important substances. A proper computational description of the physics involved in organic reactions will assist in solving many chemical problems. However, traditional force fields based on classical principles are unable to model bond dissociations, while quantum mechanical (QM) approaches are computationally too intense. Machine learned (ML) potentials have recently shown to be accurate in predicting reaction profiles. However, these ML potentials have only been parametrized for a specific reaction and therefore, lack transferability. To overcome these challenges, we develop the necessary tools and data set to train ANAKIN-ME (ANI), a new type of universal deep learned atomistic potential, to predict reaction profiles with an accuracy of DFT at
the computational cost of force field. The new ANI potential is shown to be a universal predictor for the types of reaction classes it was trained on; in other words, the low error extends to systems well outside of the training data set. With the enhancement of chemical space covered, while maintaining universality in the systems it can be applied to, ANI can be a powerful tool to assist organic chemists and a game changer in drug design research.