Choosing the right molecular machine learning potential

Quantum-chemistry simulations based on potential energy surfaces of molecules provide invaluable insight into the physicochemical processes at the atomistic level and yield such important observables as reaction rates and spectra. Machine learning potentials promise to significantly reduce the computational cost and hence enable otherwise unfeasible simulations. However, the surging number of such potentials begs the question of which one to choose or whether we still need to develop yet another one. Here, we address this question by evaluating the performance of popular machine learning potentials in terms of accuracy and computational cost. In addition, we deliver structured information for non-specialists in machine learning to guide them through the maze of acronyms, recognize each potential''s main features, and judge what they could expect from each one.

This article provides a lifeline for those lost in the sea of the molecular machine learning potentials by providing a balanced overview and evaluation of popular potentials.     

First Principles Neural Network Potentials for Reactive Simulations of Large Molecular and Condensed Systems

Modern simulation techniques have reached a level of maturity which allows a wide range of problems in chemistry and materials science to be addressed. Unfortunately, the application of first principles methods with predictive power is still limited to rather small systems, and despite the rapid evolution of computer hardware no fundamental change in this situation can be expected. Consequently, the development of more efficient but equally reliable atomistic potentials to reach an atomic level understanding of complex systems has received considerable attention in recent years. A promising new development has been the introduction of machine learning (ML) methods to describe the atomic interactions. Once trained with electronic structure data, ML potentials can accelerate computer simulations by several orders of magnitude, while preserving quantum mechanical accuracy. This Review considers the methodology of an important class of ML potentials that employs artificial neural networks.     

Incorporation of local structure into kriging models for the prediction of atomistic properties in the water decamer

Machine learning algorithms have been demonstrated to predict atomistic properties approaching the accuracy of quantum chemical calculations at significantly less computational cost. Difficulties arise, however, when attempting to apply these techniques to large systems, or systems possessing excessive conformational freedom. In this article, the machine learning method kriging is applied to predict both the intra‐atomic and interatomic energies, as well as the electrostatic multipole moments, of the atoms of a water molecule at the center of a 10 water molecule (decamer) cluster. Unlike previous work, where the properties of small water clusters were predicted using a molecular local frame, and where training set inputs (features) were based on atomic index, a variety of feature definitions and coordinate frames are considered here to increase prediction accuracy. It is shown that, for a water molecule at the center of a decamer, no single method of defining features or coordinate schemes is optimal for every property. However, explicitly accounting for the structure of the first solvation shell in the definition of the features of the kriging training set, and centring the coordinate frame on the atom‐of‐interest will, in general, return better predictions than models that apply the standard methods of feature definition, or a molecular coordinate frame. © 2016 The Authors. Journal of Computational Chemistry Published by Wiley Periodicals, Inc.     

Reaction-based machine learning representations for predicting the enantioselectivity of organocatalysts

Hundreds of catalytic methods are developed each year to meet the demand for high-purity chiral compounds. The computational design of enantioselective organocatalysts remains a significant challenge, as catalysts are typically discovered through experimental screening. Recent advances in combining quantum chemical computations and machine learning (ML) hold great potential to propel the next leap forward in asymmetric catalysis. Within the context of quantum chemical machine learning (QML, or atomistic ML), the ML representations used to encode the three-dimensional structure of molecules and evaluate their similarity cannot easily capture the subtle energy differences that govern enantioselectivity. Here, we present a general strategy for improving molecular representations within an atomistic machine learning model to predict the DFT-computed enantiomeric excess of asymmetric propargylation organocatalysts solely from the structure of catalytic cycle intermediates. Mean absolute errors as low as 0.25 kcal mol⁻¹ were achieved in predictions of the activation energy with respect to DFT computations. By virtue of its design, this strategy is generalisable to other ML models, to experimental data and to any catalytic asymmetric reaction, enabling the rapid screening of structurally diverse organocatalysts from available structural information.

A machine learning model for enantioselectivity prediction using reaction-based molecular representations.     

支持向量机与KStar模型预测细胞色素P450酶催化的氧脱烃反应

分别以支持向量机(SVM)和KStar方法为基础, 构建了代谢产物的分子形状判别和代谢反应位点判别的嵌套预测模型. 分子形状判别模型是以272个分子为研究对象, 计算了包括分子拓扑、二维自相关、几何结构等在内的1280个分子描述符, 考查了支持向量机、决策树、贝叶斯网络、k最近邻这四种机器学习方法建立分类预测模型的准确性. 结果表明, 支持向量机优于其他方法, 此模型可用于预测分子能否被细胞色素P450酶催化发生氧脱烃反应. 代谢反应位点判别模型以538个氧脱烃反应代谢位点为研究对象, 计算了表征原子能量、价态、电荷等26个量子化学特征, 比较了决策树、贝叶斯网络、KStar、人工神经网络建模的准确率. 结果显示, KStar模型的准确率、敏感性、专一性均在90%以上, 对分子形状判别模型筛选出的分子, 此模型能较好地判断出哪个C―O键发生断裂. 本文以15个代谢反应明确的中药分子为验证集, 验证模型准确性, 研究结果表明基于SVM和KStar的嵌套预测模型具有一定的准确性, 有助于开展中药分子氧脱烃代谢产物的预测研究.     

Modeling of aqueous poly(oxyethylene) solutions. 2. Mesoscale simulations

We extend our work on aqueous solutions of poly(oxyethylene) oligomers H-(CH2-O-CH2)n -H (POEn). On the basis of atomistic simulations of trimer and decamer solutions (first part of this series of papers), different sets of coarse-grained implicit-solvent potentials have been constructed using the iterative Boltzmann inversion technique. The comparison of structures obtained from coarse-grained simulations (gyration radii, end-to-end distances, radial distribution functions) with atomistic reference simulations and experiments shows that the state-specific potentials are transferable both to a wide concentration range, if the same molecule size is considered, and to at least 2 orders of magnitude larger molecules (in terms of molecular mass). Comparing the performance of different mesoscale potentials, we find different applicability ranges in terms of molecule sizes. The experimental gyration radii for chains comprising up to 1500 monomers are reproduced almost quantitatively by the decamer-fitted potentials with dihedral interactions included. The trimer-fitted potentials reproduce experimental chain dimensions of up to some hundred monomers but seem to become metastable beyond a certain chain length, as we evidenced some chain collapses. Relaxation of large-scale features is 1-2 orders of magnitude faster in the mesoscale simulations than in the atomistic simulations. The diffusion behavior in dependence of concentration is captured correctly when the decamer potential is applied to the decamer itself. For all other chain lengths, we find that time mapping from coarse-grained to atomistic trajectories has to be determined separately for each concentration. Overall, diffusion is 1-2 orders of magnitude faster on the mesoscale, depending considerably on the Lowe-Andersen thermostat parameters. The CG simulations provide an overall speed-up of about 3 orders of magnitude.     

Machine learning potential era of zeolite simulation

Zeolites, owing to their great variety and complexity in structure and wide applications in chemistry, have long been the hot topic in chemical research. This perspective first presents a short retrospect of theoretical investigations on zeolites using the tools from classical force fields to quantum mechanics calculations and to the latest machine learning (ML) potential simulations. ML potentials as the next-generation technique for atomic simulation open new avenues to simulate and interpret zeolite systems and thus hold great promise for finally predicting the structure–functionality relation of zeolites. Recent advances using ML potentials are then summarized from two main aspects: the origin of zeolite stability and the mechanism of zeolite-related catalytic reactions. We also discussed the possible scenarios of ML potential application aiming to provide instantaneous and easy access of zeolite properties. These advanced applications could now be accomplished by combining cloud-computing-based techniques with ML potential-based atomic simulations. The future development of ML potentials for zeolites in the respects of improving the calculation accuracy, expanding the application scope and constructing the zeolite-related datasets is finally outlooked.

The machine learning atomic simulation will usher the research of zeolite, as other complex materials, into a new era featuring the easy access to zeolite functionalities predicted from theory.     

基于机器学习方法的H1N1神经氨酸苷酶抑制剂的分类预测

流感是一种主要的呼吸道传染病, 在普通人群中有着较高的发病率, 而对于一些年老和高危病人还有较高的死亡率. 研究显示抑制神经氨酸苷酶(NA)可以阻断病毒RNA复制, 因此NA是有效治疗H1N1型流感病毒的重要药物靶标. 通过计算机方法进行虚拟筛选和预测NA抑制剂已经变得越来越重要. 针对酶活性位点进行基于结构的合理药物设计, 开发H1N1 病毒神经氨酸苷酶抑制剂, 已成为药物研究的热点之一. 本文通过多种机器学习方法(支持向量机(SVM)、k-最近相邻法(k-NN)和C4.5决策树(C4.5DT))对已知的神经氨酸苷酶抑制剂(NAIs)与非神经氨酸苷酶抑制剂(non-NAIs)建立分类预测模型. 其中227个结构多样性化合物(72个NAIs与155个non-NAIs)被用于测试分类预测系统, 并用递归变量消除法选择与神经氨酸苷酶抑制剂分类相关的性质描述符以提高预测精度. 本研究对独立验证集的总预测精度为75.9%-92.6%, NA 抑制剂的预测精度为64.3%-78.6%, 非H1N1抑制剂的预测精度为77.5%-97.5%. SVM法给出最好的总预测精度(92.6%). 本研究表明支持向量机等机器学习方法可以有效预测未知数据集中潜在的NA抑制剂, 并有助于发现与其相关的分子描述符.     

FFLUX: Transferability of polarizable machine‐learned electrostatics in peptide chains

The fully polarizable, multipolar, and atomistic force field protein FFLUX is being built from machine learning (i.e., kriging) models, each of which predicts an atomic property. Each atom of a given protein geometry needs to be assigned such a kriging model. Such a knowledgeable atom needs to be informed about a sufficiently large environment around it. The resulting complexity can be tackled by collecting the 20 natural amino acids into a few groups. Using substituted deca‐alanines, we present the proof‐of‐concept that a given atom's charge can be modeled by a few kriging models only. © 2017 Wiley Periodicals, Inc.