Dataset Information

A transferable active-learning strategy for reactive molecular force fields.

ABSTRACT: Predictive molecular simulations require fast, accurate and reactive interatomic potentials. Machine learning offers a promising approach to construct such potentials by fitting energies and forces to high-level quantum-mechanical data, but doing so typically requires considerable human intervention and data volume. Here we show that, by leveraging hierarchical and active learning, accurate Gaussian Approximation Potential (GAP) models can be developed for diverse chemical systems in an autonomous manner, requiring only hundreds to a few thousand energy and gradient evaluations on a reference potential-energy surface. The approach uses separate intra- and inter-molecular fits and employs a prospective error metric to assess the accuracy of the potentials. We demonstrate applications to a range of molecular systems with relevance to computational organic chemistry: ranging from bulk solvents, a solvated metal ion and a metallocage onwards to chemical reactivity, including a bifurcating Diels-Alder reaction in the gas phase and non-equilibrium dynamics (a model S_N2 reaction) in explicit solvent. The method provides a route to routinely generating machine-learned force fields for reactive molecular systems.

SUBMITTER: Young TA

PROVIDER: S-EPMC8372546 | biostudies-literature | 2021 Aug

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

A transferable active-learning strategy for reactive molecular force fields.

Young Tom A TA Johnston-Wood Tristan T Deringer Volker L VL Duarte Fernanda F

Chemical science 20210705 32

Predictive molecular simulations require fast, accurate and reactive interatomic potentials. Machine learning offers a promising approach to construct such potentials by fitting energies and forces to high-level quantum-mechanical data, but doing so typically requires considerable human intervention and data volume. Here we show that, by leveraging hierarchical and active learning, accurate Gaussian Approximation Potential (GAP) models can be developed for diverse chemical systems in an autonomo ...[more]

PMID: 34476072

Similar Datasets

Project description:Ion descriptors in molecular mechanics models are calibrated against reference data on ion-water interactions. It is then typically assumed that these descriptors will also satisfactorily describe interactions of ions with other functional groups, such as those present in biomolecules. However, several studies now demonstrate that this transferability assumption produces, in many different cases, large errors. Here we address this issue in a representative polarizable model and focus on transferability of cationic interactions from water to a series of alcohols. Both water and alcohols use hydroxyls for ion-coordination, and, therefore, this set of molecules constitutes the simplest possible case of transferability. We obtain gas phase reference data systematically from "gold-standard" quantum Monte Carlo and CCSD(T) methods, followed by benchmarked vdW-corrected DFT. We learn that the original polarizable model yields large gas phase water → alcohol transferability errors - the RMS and maximum errors are 2.3 and 5.1 kcal/mol, respectively. These errors are, nevertheless, systematic in that ion-alcohol interactions are overstabilized, and systematic errors typically imply that some essential physics is either missing or misrepresented. A comprehensive analysis shows that when both low- and high-field responses of ligand dipole polarization are described accurately, then transferability improves significantly - the RMS and maximum errors in the gas phase reduce, respectively, to 0.9 and 2.5 kcal/mol. Additionally, predictions of condensed phase transfer free energies also improve. Nevertheless, within the limits of the extrathermodynamic assumptions necessary to separate experimental estimates of salt dissolution into constituent cationic and anionic contributions, we note that the error in the condensed phase is systematic, which we attribute, at least, partially to the parametrization in long-range electrostatics. Overall, this work demonstrates a rational approach to boosting transferability of ionic interactions that will be applicable broadly to improving other polarizable and nonpolarizable models.

Project description:Accurate and fast evaluation of electrostatic interactions in molecular systems is one of the most challenging tasks in the rapidly advancing field of macromolecular chemistry and drug design. Electrostatic interactions are of crucial importance in biological systems. They are well represented by quantum mechanical methods; however, such calculations are computationally expensive. In this study, we have evaluated the University of Buffalo Pseudoatom Databank (UBDB)1,2 approach for approximation of electrostatic properties of macromolecules and their complexes. We selected the S663 and JSCH-20054 data sets (208 molecular complexes in total) for this study. These complexes represent a wide range of chemical and biological systems for which hydrogen bonding, electrostatic, and van der Waals interactions play important roles. Reference electrostatic energies were obtained directly from wave functions at the B3LYP/aug-cc-pVTZ level of theory using the SAPT (Symmetry-Adapted Perturbation Theory) scheme for calculation of electrostatic contributions to total intermolecular interaction energies. Electrostatic energies calculated on the basis of the UBDB were compared with corresponding reference results. Results were also compared with energies computed using a point charge model from popular force fields (AM1-BCC and RESP used in AMBER and CGenFF from CHARMM family). The energy trends are quite consistent (R2 ≈ 0.98) for the UBDB method as compared to the AMBER5 and CHARMM force field methods6(R2 ≈ 0.93 on average). The RSMEs do not exceed 3.2 kcal mol-1 for the UBDB and are in the range of 3.7-7.6 kcal mol-1 for the point charge models. We also investigated the discrepancies in electrostatic potentials and magnitudes of dipole moments among the tested methods. This study shows that estimation of electrostatic interaction energies using the UBDB databank is accurate and reasonably fast when compared to other known methods, which opens potential new applications to macromolecules.

Dataset Information

A transferable active-learning strategy for reactive molecular force fields.

Publications

A transferable active-learning strategy for reactive molecular force fields.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets