/mldl-md-dynamics

A repository of update in molecular dynamics field by recent progress in machine learning and deep learning.

Awesome machine learning/deep learning in molecular dynamics

A repository of update in molecular dynamics field by recent progress in machine learning and deep learning. Those efforts are cast into the following categories:

  1. Learn force field or molecular interactions
  2. Enhanced sampling methods
  3. Learn collective variable
  4. Learn kinetic model
  5. Capture dynamics of molecular system
  6. Map between all atoms and coarse grain
  7. Design proteins

 

Machine learning molecular dynamics for the simulation of infrared spectra

(Picture from *Machine learning molecular dynamics for the simulation of infrared spectra*. )  

1. Learn force field or molecular interactions

Molecular Graph Convolutions: Moving Beyond Fingerprints
Steven Kearnes, Kevin McCloskey, Marc Berndl, Vijay Pande, Patrick Riley. (2016)
This paper from Standford Univ and Google proposed graph representation of molecules and graph convolution to capture the interactions in the molecule. The authors used a weave module, where the atom feature and edge feature are weaved to preserve invariance of atom and pair permutation. They used Gaussian membership functions to preserve overall order invariance.

An implementation of artificial neural-network potentials for atomistic materials simulations: Performance for TiO2
NongnuchArtrith, Alexander Urban. (2016)
The authors from UC Berkeley developed open-source atomic energy network package, based on Behler-Parrinello machine learning potential, which uses multilayer perceptron to learn the potential of molecules. The atomic coordinates are transformed into invariant representation of the local atomic environments and potential is trained on such representation. The authors applied the model to TiO2, ZrO2, and alpha-PbO2.

Machine learning prediction errors better than DFT accuracy
Felix A. Faber, Luke Hutchison, Bing Huang, Justin Gilmer, Samuel S. Schoenholz, George E. Dahl, Oriol Vinyals, Steven Kearnes, Patrick F. Riley, O. Anatole von Lilienfeld. (2017)
The authors from Univ of Basel and Google used elastic network, bayesian regression, random forest, kernel ridge regression, gated graph NN, graph convolutions to predict QM9 data set. The representations are Coulomb matrix, BAML (bonds, angles, machine learning), ECFP4 (extended connectivity fingerprints), MARAD (molecular atomic radial angular distribution), HD, HDA, HDAD (histogram methods). They demonstrated the machine learning methods have smaller error than DFT error.

Quantum-Chemical Insights from Deep Tensor Neural Networks
Kristof T. Schütt, Farhad Arbabzadah, Stefan Chmiela, Klaus R. Müller, Alexandre Tkatchenko. (2017)
The authors from Technische Universitat Berlin, Korea Univ, Fritz-Haber-Institut der Max-Planck-Gesellschaft and Univ of Luxembourg developed DTNN. The network used atom features and edge features for input. Edges are processed by Gaussian expansion. The edges and atoms interact through an interaction module through tensor multiplications. The authors applied this to predict chemical potentials, ring stability of molecules etc.

Machine learning molecular dynamics for the simulation of infrared spectra
Michael Gastegger, Jörg Behler, Philipp Marquet. (2017)
The authors from Univ. of Vienna and Universität Göttingen developed a molecular dipole moment model based on environment-dependent NN and combined with NN potential approach of Behler and Parrinello for ab inito MD. As an application, they obtained accurate models for predicting infrared spectra.

ANI-1: an extensible neural network potential with DFT accuracy at force field computational cost
J. S. Smith, Isayev, A. E. Roitberg. (2017)
This paper from Univ. of Florida and Univ. of North Carolina presented ANI-1, which used Behler and Parrinello symmetry functions to build single-atom atomic environment vectors (AEV) as molecular representation. This is similar to the context representation of work in NLP.

Applying machine learning techniques to predict the propertiesof energetic materials
Daniel C. Elton, Zois Boukouvalas, Mark S. Butrico, Mark D. Fuge, Peter W. Chung. (2018)
The authors from Univ of Maryland applied several machine learning methods (KRR, ridge, SVR, RF, k-nearest neighbor) based on features (sum over bonds, custom descriptors, Coulomb matrices, Bag of Bonds, and fingerprints). They concluded the best featurization is sum over bonds and best model is kernel ridge regression.

Deep Potential Molecular Dynamics: A Scalable Model with the Accuracy of Quantum Mechanics
Linfeng Zhang, Jiequn Han, Han Wang, Roberto Car, Weinan E. (2018)
The authors from Peking Univ., Princeton Univ., and Institute of Applied Physics and Computational Mathematics, China developed DeepMD method based on a many-body potential and interatomic forces generated by NN, which is trained with ab initio data.

Neural Message Passing with Edge Updates for Predicting Properties of Molecules and Materials
Peter Bjørn Jørgensen, Karsten Wedel Jacobsen, Mikkel N. Schmidt. (2018)
This paper from Univ. of Denmark extended neural message passing model with an edge update NN, so that information exchanges between atoms depend on hidden state of the receiving atom. They also explored ways to construct the graph.

SchNet – A deep learning architecture for molecules and materials
K. T. Schütt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko, K.-R. Müller. (2018)
This paper from Technische Universita ̈t Berlin, Univ. of Luxembourg, Max Planck Institute, and Korea University presented SchNet, a variant of DTNN to learn the molecular properties and studied local chemical potential and the dynamics of C20-fullerene.

Pixel Chem: A Representation for Predicting Material Properties with Neural Network Shuqian Ye, Yanheng Xu, Jiechun Liang, Hao Xu, Shuhong Cai, Shixin Liu, Xi Zhu.(2019)
The authors designed a Pixel Chemistry network to learn a representation for predicting molecular properties. The authors proposed three new matrices, which reflect charge transfer ability, bond binding strength, and Euclidean distances between atoms. They also designed an angular interaction matrix A, describes the interaction between two atomic orbitals.

Message-passing neural networks for high-throughput polymer screening
Peter C. St. John1, Caleb Phillips, Travis W. Kemper, A. Nolan Wilson, Yanfei Guan, Michael F. Crowley, Mark R. Nimlos, Ross E. Larsen. (2019)
This paper from National Renewable Energy Lab, USA, used message-passing NN to predict polymer properties for screening purpose. They focused on larger molecules and tested the model with/without 3D conformation information, since accurate 3D structure calculation is also expensive.

Accurate and transferable multitask prediction of chemical properties with an atoms-in-molecules neural network
Roman Zubatyuk, Justin S. Smith, Jerzy Leszczynski and Olexandr Isayev. (2019)
This paper from Univ. of North Carolina, Los Alamos National Lab, and Jackson State Univ presented AIMNet to leearn implicit solvation energy in MNSol database. Atoms in molecules are embedded and interact with each in several layers.

LanczosNet: Multi-Scale Deep Graph Convolutional Networks
Renjie Liao, Zhizhen Zhao, Raquel Urtasun, Richard S. Zemel. (2019)
The authors from Univ. of Toronto, Uber ATG, Vector Institute, UIUC and Canadian Institute of Advanced Research developed this spectral-based graph NN, which uses Lanczos algorithms to construct low rank approximations of the graph Laplacian. They benchmarked the model on citation networks and QM8 dataset.

Molecule-Augmented Attention Transformer
Łukasz Maziarka, Tomasz Danel, Sławomir Mucha, Krzysztof Rataj, Jacek Tabor, Stanisław Jastrzebski. (2019)
The authors from Jagiellonian Univ, Ardigen and New York Univ designed this MAT graph NN model with self-attention mimicking the Transformer, consisting of multiple blocks of layer norm, multi-head self-attention, and residual net. The model achieved comparable or better results on BBBP and FreeSolv datasets comparing with MPNN.

Machine Learning for Scent: Learning Generalizable Perceptual Representations of Small Molecules
Benjamin Sanchez-Lengeling, Jennifer N. Wei, Brian K. Lee, Richard C. Gerkin, Alán Aspuru-Guzik, Alexander B. Wiltschko. (2019)
This paper from Google, Arizona State Univ, Univ of Toronto, Vector Institute, Canadian Institute for Advanced Research used MPNN (message passing NN) based on graph representation, to predict quantitative structure-odor relationship (QSOR), very similar to QSAR. The model out-performed molecular fingerprint-based methods. The authors showed their learned embeddings from GNN capture a meaningful odor space representation.

ProDyn0: Inferring calponin homology domain stretching behavior using graph neural networks
Ali Madani, Cyna Shirazinejad, Jia Rui Ong, Hengameh Shams, Mohammad Mofrad. (2019)
This paper from UC Berkeley used MPNN and residual gated graph convnets to predict the pattern and mode of SMD (steered MD) simulation results. The authors created this data set of 2020 mutants of calponin homology domain (CH, an actin-binding domain) with SMD simulation results. Capturing the force between CH domains is capturing molecular interactions between amino acid residues.

2. Enhanced sampling methods with ML/DL

Reinforced dynamics for enhanced sampling in large atomic and molecular systems
Linfeng Zhang, Han Wang, Weinan E. (2018)
This paper from Peking Univ., Princeton Univ, and IAPCM, China used reinforcement learning to calculate the biasing potential on the fly, with data collected judiciously from exploration and an uncertainty indicator from NN serving as the reward function.

Reinforcement Learning Based Adaptive Sampling: REAPing Rewards by Exploring Protein Conformational Landscapes
Zahra Shamsi, Kevin J. Cheng, Diwakar Shukla. (2018)
This paper from UIUC used reinforcement learning to adaptively biase the sampling potential. The action in this RL problem is to pick new structures to start a swarm of simulations, and the reward function is how far order parameters sample the landscape.

Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning
Frank Noé, Simon Olsson, Jonas Köhler, Hao Wu. (2019)
This paper from Freie Universität Berlin, Rice Univ and Tongji Univ used a generative model, Boltzmann generator machine, to generate unbiased equilibrium samples from different metastable states in one shot. This model is said to overcome rare event-sampling problems in many-body systems.

Targeted Adversarial Learning Optimized Sampling
Justin Zhang, Yi Isaac Yang, Frank Noé (2019)
The authors from Freie Universität Berlin use adversarial training to steer a molecular dynamics ensemble towards a desired target distribution, overcoming rare-event sampling problems.

Neural networks-based variationally enhanced sampling
Luigi Bonati, Yue-Yu Zhang, Michele Parrinello. (2019)
The authors from ETH Zurich, Universita della Svizzera italiana, MARVEL (Switzerland), and Italian Institute of Technology presented a NN-based bias potential for enhanced sampling, building on their previous work of variationally enhanced sampling. Deep learning provides an expressive tool for mapping from CV to actual bias potential.

3. Learn collective variables

Machine Learning Based Dimensionality Reduction Facilitates Ligand Diffusion Paths Assessment: A Case of Cytochrome P450cam
Jakub Rydzewski, and Wieslaw Nowak. (2016)
The authors from Nicolaus Copernicus University showed how t-distributed stochastic neighbor embedding (t-SNE) can be applied to analyze the process of camphor unbinding from cytochrome P450cam via multiple reaction pathways.

Transferable Neural Networks for Enhanced Sampling of Protein Dynamics
Mohammad M. Sultan, Hannah K. Wayment-Steele, Vijay S. Pande. (2018)
The authors from Stanford Univ used variational autoencoder with time-lagged information to learn the collective variable in latent space. They then used the latent space representation in well-tempered ensemble metadynamics. The authors showed such learned latend space is transferrable for proteins with certain mutations or between force fields.

Time-lagged autoencoders: Deep learning of slow collective variables for molecular kinetics
Christoph Wehmeyer, Frank Noé. (2018)
The authors from Freie Universität Berlin built time-lagged autoencoders to learn the slow collective variables. They show that time-lagged autoencoders are a nonlinear generalization of the time-lagged independent component analysis (TICA) method.

Reweighted autoencoded variational Bayes for enhanced sampling (RAVE)
João Marcelo Lamim Ribeiro, Pablo Bravo, Yihang Wang, and Pratyush Tiwary. (2018)
This paper from Univ of Maryland and Pontificia Universidad Catolica de Chile used variational autoencoder and Bayes theorem to find the reaction coordinates and approapriate weights. Kullback-Leibler divergence is calculated between this latent space distribution and the distribution of various trial reaction coordinates sampled from the simulation.

Learning protein conformational space by enforcing physics with convolutions and latent interpolations
Venkata K. Ramaswamy, Chris G. Willcocks, Matteo T. Degiacomi. (2019)
This paper from Durhan Univ designed a CNN-based autoencoder to learn a continuous latent space for protein conformations. Based on the latent space, they derived a transition path between two states. The authors also augmented the network with MD simulation data, incorporating physics-based constraints, achieving high accuracy.

Nonlinear discovery of slow molecular modes using state-free reversible VAMPnets
Wei Chen, Hythem Sidky, Andrew L. Ferguson. (2019)
The authors from UIUC and Univ of Chicago introduced SRV, state-free reversible VAMPnets to learn nonlinear CV approximants. The work built on VAMPNet (variational approach for Markov processes networks). SRV learns the first few slow eigenfunctions of the spectral decomposition of the transfer operator, which evolves probability distribution at equilibrium through time.

Past–future information bottleneck for sampling molecular reaction coordinate simultaneously with thermodynamics and kinetics
Yihang Wang, João Marcelo Lamim Ribeiro, Pratyush Tiwary. (2019)
The authors from Univ of Maryland used variational inference implemented in deep neural networks to infer reaction coordinates/CV. To sample the rare event, the authors took analogue of predictive information bottleneck, trying to maximize the prediction of future by utilizing the informaiton from the past.

Artificial Intelligence Assists Discovery of Reaction Coordinates and Mechanisms from Molecular Dynamics Simulations
Hendrik Jung, Roberto Covino, Gerhard Hummer. (2019) The authors from Max Planck Institute of Biophysics and Goethe Univ introduced an NN-based model to find the reaction coordinates. Based on the transition path sampling (TPS) theory, the authors did MD simulations, built transition path ensemble, find reaction coordinates and do more MD simulations.

4. Learn kinetic model

VAMPnets for deep learning of molecular kinetics
Andreas Mardt, Luca Pasquali, Hao Wu, Frank Noé (2018)
The authors from Freie Universität Berlin employ the variational approach for Markov processes (VAMP) to develop a deep learning framework for molecular kinetics using neural networks, dubbed VAMPnets. A VAMPnet encodes the entire mapping from molecular coordinates to a Markov state model (MSM), thus combining the MSM whole data processing pipeline in a single end-to-end framework.

5. Capture the dynamics of the molecular system

Equivariant Hamiltonian Flows
Danilo Jimenez Rezende, Sébastien Racanière, Irina Higgins, Peter Toth. (2019)
This paper from Google uses Lie algebra to prove what hamiltonian flow learns and how addition of symmetry invariance constraints can improve data efficiency.

Equivariant Flows: sampling configurations formulti-body systems with symmetric energies
Jonas Köhler, Leon Klein, Frank Noé. (2019)
This paper from Freie Universität Berlin model flows that have symmetries in the energy built in, such as roto-translational and permutational invariances, as a system of interacting particles. Can be used both for learning particle dynamics and sampling equilibrium states.

Symplectic ODE-NET: learning Hamiltonian dynamics with control
Yaofeng Desmond Zhong, Biswadip Dey, Amit Chakraborty. (2019)
This paper from Princeton University and Siemens Corp infers the dynamics of a physical system from observed state trajectories. They embedded high dimensional coordinates into low dimensions and velocity into general momentum.

Hamiltonian Neural Networks
Sam Greydanus, Misko Dzamba, Jason Yosinski. (2019)
This paper from Google, PetCube and Uber trains models to learn conservation law of Hamiltonian in unsupervised way.

Symplectic Recurrent Neural Networks
Zhengdao Chen, Jianyu Zhang, Martin Arjovsky, Léon Bottou. (2019)
The authors from NYU, Tianjin University, and Facebook proposes SRNN to capture the dynamics of physical systems from observed trajectories.

Physical Symmetries Embedded in Neural Networks
M. Mattheakis, P. Protopapas, D. Sondak, M. Di Giovanni, E. Kaxiras. (2019)
The authors from Harvard and Polytechnic Milan used symplectic neural network to embed physics symmetry in the neural network to characterize the dynamics.

Neural Canonical Transformation with Symplectic Flows
Shuo-Hui Li, Chen-Xiao Dong, Linfeng Zhang, Lei Wang. (2019)
The authors from CAS, Princeton Univ., and Songshan Lake Materials Lab constructed canonical transformation with symplectic neural networks. Such formulations help understand the physical meaning of latend space in the model. The authors applied this to learn slow CV of analine dipeptide and conceptual compression of MNIST dataset.

6. Coarse grain models

Machine Learning of coarse-grained Molecular Dynamics Force Fields
Jiang Wang, Simon Olsson, Christoph Wehmeyer, Adrià Pérez, Nicholas E. Charron, Gianni de Fabritiis, Frank Noé, Cecilia Clementi. (2018)
The authors from Rice University, Freie Universität Berlin, and Universitat Pompeu Fabra presented CGnet which learns coarse grain force field by using variational force matching. They also recast force-matching as a machine learning problem, allowing to decompose the force matching error into bias, variance and noise. They demonstrated the model performance on dialanine peptide simulation and Chignolin folding/unfolding in water.

DeePCG: Constructing coarse-grained models via deep neural networks
Linfeng Zhang, Jiequn Han, Han Wang, Roberto Car, and Weinan E. (2018)
The authors from Peking Univ, Princeton Univ, and IAPCM, China presented DeepCG to construct a many-body CG potential. The authors applied this to liquid water and did CG simulation starting from an atomistic simulation at ab inito level.

Adversarial-Residual-Coarse-Graining: Applying machine learning theory to systematic molecular coarse-graining
Aleksander E. P. Durumeric, Gregory A. Voth. (2019)
The authors from Univ. of Chicago employed generative adversial network (GAN) for systematic molecular coarse-graining. They showed that the resulting framework can rigorously parameterize CG models containing CG sites with no prescribed connection to the reference atomistic system.

7. Design proteins

(Though this part is less connected to MD simulation, some of the ML-based protein design algorithms are actually inditectly learning the potential energy of proteins. So we keep a small portion here.)

Generative models for graph-based protein design
John Ingraham, Vikas K. Garg, Regina Barzilay, Tommi Jaakkola. (2019)
This paper from MIT used generative graph model to design proteins. View this as a reverse problem of protein folding/structure prediction, the authors showed their approach efficiently captures the long-range interactions that are distant in sequence but local in 3D structure.