Collective variable discovery and enhanced sampling using autoencoders: Innovations in network architecture and error function design

Chen, Wei; Tan, Aik Rui; Ferguson, Andrew L.

doi:10.1063/1.5023804

Cited by 130 publications

(163 citation statements)

References 91 publications

Supporting

Mentioning

161

Contrasting

Order By: Relevance

“…Proteins represent a particularly interesting application case in this area, since they do not feature a difficulty typical of raw 3D point clouds: as the position of their constituent atoms is constrained by covalent interactions, protein conformations can be interpreted as ordered sets of points. Generative neural networks have been recently proposed as a tool for the discovery of collective variables, useful to extract kinetic information from molecular simulations or to guide the sampling of poorly explored regions (Chen et al, 2018;Chiavazzo et al, 2017;Herná ndez et al, 2018;Mardt et al, 2018;Ribeiro et al, 2018).…”

Section: Introductionmentioning

confidence: 99%

Coupling Molecular Dynamics and Deep Learning to Mine Protein Conformational Space

Degiacomi

2019

Structure

100

View full text Add to dashboard Cite

Graphical Abstract Highlights d A neural network is trained with protein atomic structures d The trained network can generate new protein-like structures d The neural network is used to generate candidate subunits for protein docking Authors Matteo T. Degiacomi Correspondence matteo.t.degiacomi@durham.ac.uk In Brief Degiacomi presents a usage of generative neural networks for the characterization of the conformational space of proteins featuring domain-level dynamics. The network can generate protein-like structures and can be combined to a protein-protein docking algorithm to identify conformations close to the bound state. SUMMARYFlexibility is often a key determinant of protein function. To elucidate the link between their molecular structure and role in an organism, computational techniques such as molecular dynamics can be leveraged to characterize their conformational space. Extensive sampling is, however, required to obtain reliable results, useful to rationalize experimental data or predict outcomes before experiments are carried out. We demonstrate that a generative neural network trained on protein structures produced by molecular simulation can be used to obtain new, plausible conformations complementing preexisting ones. To demonstrate this, we show that a trained neural network can be exploited in a protein-protein docking scenario to account for broad hinge motions taking place upon binding. Overall, this work shows that neural networks can be used as an exploratory tool for the study of molecular conformational space.

show abstract

Section: Introductionmentioning

confidence: 99%

Coupling Molecular Dynamics and Deep Learning to Mine Protein Conformational Space

Degiacomi

2019

Structure

100

View full text Add to dashboard Cite

show abstract

“…[11][12][13][14] Popular methods include principal component analysis (PCA) 15,16 which represents a linear transformation to collective variables that maximize the variance of the first components, and time-lagged independent component analysis (TICA) [17][18][19] which aims to maximize the timescales of the first components. Moreover, various kinds of nonlinear techniques [20][21][22][23] as well as a variety of machine learning approaches [24][25][26][27][28][29][30] have been proposed.…”

Section: Introductionmentioning

confidence: 99%

Dynamical coring of Markov state models

Nagel

Weber

Lickert

et al. 2019

The Journal of Chemical Physics

View full text Add to dashboard Cite

The accurate definition of suitable metastable conformational states is fundamental for the construction of a Markov state model describing biomolecular dynamics. Following the dimensionality reduction of a molecular dynamics trajectory, these microstates can be generated by a recently proposed density-based geometrical clustering algorithm [J. Chem. Theory Comput. 12, 2426 (2016)], which by design cuts the resulting clusters at the energy barriers and allows for a data-based identification of all parameters. Nevertheless, projection artifacts due to the inevitable restriction to a low-dimensional space combined with insufficient sampling often leads to a misclassification of sampled points in the transition regions. This typically causes intrastate fluctuations to be mistaken as interstate transitions, which leads to artificially short life time of the metastable states. As a simple but effective remedy, dynamical coring requires that the trajectory spends a minimum time in the new state for the transition to be counted. Adopting molecular dynamics simulations of two well-established biomolecular systems (alanine dipeptide and villin headpiece), dynamical coring is shown to considerably improve the Markovianity of the resulting metastable states, which is demonstrated by Chapman-Kolmogorov tests and increased implied timescales of the Markov model. Providing high structural and temporal resolution, the combination of density-based clustering and dynamical coring is particularly suited to describe the complex structural dynamics of unfolded biomolecules.

show abstract

“…We thus retain as much of the physical foundations and robustness of traditional modeling as possible, while being pragmatic about the parts of a problem, where that is not possible (see, e.g., Refs. [29][30][31]).…”

Section: B Creation Of New Modeling Techniquesmentioning

confidence: 99%

Advances of machine learning in molecular modeling and simulation

Haghighatlari

Hachmann

2019

Current Opinion in Chemical Engineering

108

View full text Add to dashboard Cite

In this review, we highlight recent developments in the application of machine learning for molecular modeling and simulation. After giving a brief overview of the foundations, components, and workflow of a typical supervised learning approach for chemical problems, we showcase areas and state-of-the-art examples of their deployment. In this context, we discuss how machine learning relates to, supports, and augments more traditional physics-based approaches in computational research. We conclude by outlining challenges and future research directions that need to be addressed in order to make machine learning a mainstream chemical engineering tool.

show abstract

Collective variable discovery and enhanced sampling using autoencoders: Innovations in network architecture and error function design

Cited by 130 publications

References 91 publications

Coupling Molecular Dynamics and Deep Learning to Mine Protein Conformational Space

Coupling Molecular Dynamics and Deep Learning to Mine Protein Conformational Space

Dynamical coring of Markov state models

Advances of machine learning in molecular modeling and simulation

Contact Info

Product

Resources

About