Active learning accelerates ab initio molecular dynamics on reactive energy surfaces

Ang, Shi Jun; Wang, Wujie; Schwalbe-Koda, Daniel; Axelrod, Simon; Gómez‐Bombarelli, Rafael

doi:10.1016/j.chempr.2020.12.009

Cited by 62 publications

(64 citation statements)

References 61 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…[7,41]). While this approach is not new, recent advances and increase interest in the field of QML potentials have motivated new developments in this field (e.g., [42][43][44]).…”

Section: Qml Potentials Can Be Trained Using Experimental Thermodynamic Data To Systematically Improve Accuracymentioning

confidence: 99%

Teaching free energy calculations to learn from experimental data

Wieder

Fass

Chodera

2021

Preprint

View full text Add to dashboard Cite

Alchemical free energy calculations are an important tool in the computational chemistry tool-box, enabling the efficient calculation of quantities critical for drug discovery such as ligand binding affinities, selectivities, and partition coefficients. However, modern alchemical free energy calculations suffer from three significant limitations: (1) modern molecular mechanics force fields are limited in their ability to model complex molecular interactions, (2) classical force fields are unable to treat phenomena that involve rearrangements of chemical bonds, and (3) these calculations are unable to easily learn to improve their performance if readily-available experimental data is available. Here, we show how all three limitations can be overcome through the use of quantum machine learning (QML) potentials capable of accurately modeling quantum chemical energetics even when chemical bonds are made and broken. Because these potentials are based on mathematically convenient deep learning architectures instead of traditional quantum chemical formulations, QML simulations can be run at a fraction of the cost of quantum chemical simulations using modern graphics processing units (GPUs) and machine learning frameworks. We demonstrate that alchemical free energy calculations in explicit solvent are especially simple to implement using QML potentials because these potentials lack singularities and other pathologies typical of molecular mechanics potentials, and that alchemical free energy calculations are highly effective even when bonds are broken or made. Finally, we show how a limited number of experimental free energy measurements can be used to significantly improve the accuracy of computed free energies for unrelated compounds with no significant generalization gap. We illustrate these concepts on the prediction of aqueous tautomer free energies (related to tautomer ratios), which are highly relevant to drug discovery in that more than a quarter of all approved drugs exist as a mixture of tautomers.

show abstract

“…[7,41]). While this approach is not new, recent advances and increase interest in the field of QML potentials have motivated new developments in this field (e.g., [42][43][44]).…”

Section: Qml Potentials Can Be Trained Using Experimental Thermodynamic Data To Systematically Improve Accuracymentioning

confidence: 99%

Teaching free energy calculations to learn from experimental data

Wieder

Fass

Chodera

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…3 Machine learning (ML) approaches have the potential to revolutionise force-field based simulations, aiming to provide the best of both worlds, [4][5][6] and have indeed begun to provide new insights into a range of challenging research problems. [7][8][9][10][11][12][13][14][15][16] The development of an ML potential applicable to the whole periodic table mapping nuclear coordinates to total energies and forces is, however, precluded by the curse of dimensionality. Within small chemical subspaces, models can be achieved using neural networks (NNs), 6,[17][18][19][20][21] kernel-based methods such as the Gaussian Approximation Potential (GAP) framework 22,23 or gradient-domain machine learning (GDML), 24 and linear fitting with properly chosen basis functions, 25,26 each with different data requirements and transferability.…”

Section: Introductionmentioning

confidence: 99%

“…Very recently, AL approaches have started to be adopted for fitting reactive potentials for organic molecules based on single point evaluations at quantum-chemical levels of theory. Notable examples include the modelling of gas-phase pericyclic reactions, 12 the exploration of reactivity during methane combustion, 50 and the decomposition of urea in water. 41 In the present work -with a view to developing potentials to simulate solution phase reactions -we consider bulk water as a test case and develop a strategy which requires just hundreds of total ground truth evaluations and no a priori knowledge of the system, apart from the molecular composition.…”

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

A transferable active-learning strategy for reactive molecular force fields

et al. 2021

View full text Add to dashboard Cite

show abstract

“…Conversely, if the training set already has similar structures of the given one, the predicted results of these models should be consistent. This algorithm was also called "active learning" and has been used by many works [42][43][44][45] . Recently, a concurrent learning algorithm was proposed by E and co-workers 46 .…”

Section: Introductionmentioning

confidence: 99%

Automated Construction of Neural Network Potential Energy Surface: The Enhanced Self-Organizing Incremental Neural Network Deep Potential Method

Xu¹,

Zhu²,

Zhang³

2021

Preprint

View full text Add to dashboard Cite

In recent years, the use of deep learning (neural network) potential energy surface (NNPES) in molecular dynamics simulation has experienced explosive growth as it can be as accurate as quantum chemistry methods while being as efficient as classical mechanic methods. However, the development of NNPES is highly non-trivial. In particular, it has been troubling to construct a dataset that is as small as possible yet can cover the target chemical space. In this work, an ESOINN-DP method is developed, which has the enhanced self-organizing incremental neural-network (ESOINN) and a newly proposed error indicator at its core. With ESOINN-DP, One can construct the NNPES with little human intervention, and this method ensures that the constructed reference dataset covers the target chemical space with minimum redundancy. The performance of the ESOINN-DP method has been well validated by developing neural network potential energy surfaces for water clusters and by de-redundancy of a sub-data set of the ANI-1 database. We believe that the ESOINN-DP method provides a novelty idea for the construction of NNPES and especially, the reference datasets, and it can be used for MD simulations of various gas-phase and condensed-phase chemical systems.

show abstract

Active learning accelerates ab initio molecular dynamics on reactive energy surfaces

Cited by 62 publications

References 61 publications

Teaching free energy calculations to learn from experimental data

Teaching free energy calculations to learn from experimental data

A transferable active-learning strategy for reactive molecular force fields

Automated Construction of Neural Network Potential Energy Surface: The Enhanced Self-Organizing Incremental Neural Network Deep Potential Method

Contact Info

Product

Resources

About