Knowledge of the conformational dynamics of molecules free in solution is essential for predicting molecular properties and guiding structure-based drug discovery. Accurate unbound structures provide key insights to optimize the development of a drug molecule because the conformational changes that its unbound state undergoes to assume the bioactive conformer have an impact on enthalpy and entropy and, consequently, on the observed binding affinity. Monte Carlo (MC) and molecular dynamics are the simulation methods usually employed to sample unbound molecular conformers. These methods are used alongside potential energy models such as machine-learning potentials. Owing to their lack of an explicit functional form, they can potentially achieve levels of accuracy close to the quantum level of theory, while still being significantly cheaper. However, because of their novelty, benchmark studies evaluating their accuracy are still limited. There is therefore a pressing need to assess the applicability of these potentials in different situations since they may provide the pathway towards high accuracy simulations at a moderate computational cost.<br/>Here we present a study that aims to determine the reliability of ANI-2x, a neural network (NN) potential with density functional theory (DFT) accuracy. Specifically, we determine the accuracy of ANI-2x in modeling a set of fluorinated compounds containing γ-fluorohydrins and γ,γ-difluorohydrins. For this purpose, we use a multi-level MC algorithm that involves coupling the hybrid MC algorithm with a switching step between two Markov chains, the latter as proposed by Gelb. The acceptance rate of the switching step is a metric of the similarity between ANI-2x and the DFT functional it aims to reproduce, <i>viz. w</i>B97X-D/6-31G*, and therefore it is here used as a measure of the QM-likeness of the NN potential. Furthermore, using the high-level configurational ensembles generated in the multi-level MC approach, we derive bespoke class I FFs and test their performance and efficiency in comparison to ANI-2x. This is done both in the gas phase and in aqueous solution by comparing configurational distributions, multi-level MC acceptance rates, and NMR conformational data. The results obtained show the accuracy of each model in simulating the conformational landscape of this set of fluorinated compounds and demonstrate the complex interplay of inter and intramolecular interactions in determining conformer stability.

Do machine-learning potentials perform better than bespoke-derived force fields?

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.