Practical Machine Learning Strategies. I. Correcting the MMFF Molecular Mechanics Model to More Accurately Provide Conformational Energy Differences in Flexible Organic Molecules

A correction to the MMFF molecular mechanics model, based on a neural network trained to reproduce conformer energy differences obtained from ωB97X-V/6-311+G(2df,2p)[6-311G*]//MMFF calculations is described. It is supported for molecules containing H, C, N, O, F, S, Cl, and Br. The correction adds only slightly to the cost of MMFF, and the resulting corrected model is several orders of magnitude faster than ωB97X-V/6-311+G(2df,2p)[6-311G*]. It properly identifies the lowest energy conformer for 82% of the molecules in a test set of flexible organic molecules (3553 total conformers), compared with 38% for MMFF. While the corrected MMFF model cannot be expected to provide sufficiently accurate Boltzmann weights for use in spectra and property calculations on flexible molecules, it is able to reduce the number of “reasonable” conformers that need to be passed on to more rigorous computational models, that can.