Hierarchical Dynamics Encoder

Hierarchical dynamics encoder (HDE) is a deep learning-based framework to learn multiple hierarchical nonlinear kinetic slow modes. They are built on top of transfer operator theory, the variational approach to conformational dynamics (VAC), and use a neural network as a featurizer to provide an optimal nonlinear basis set for VAC, which finds an optimal linear combination of the bases.

HDEs are similar to (and inspired by) VAMPNets and VDEs with some key differences. VAMPNets seek to replace the entire MSM construction pipeline of featurization, dimensionality reduction, and state assignment. In the one-dimensional limit, HDEs are formally equivalent to VDEs with an exclusive autocorrelation loss, subject to Gaussian noise. VDEs however, cannot currently generalize to multiple dimensions due to the lack of an orthogonality constraint on the learned eigenfunctions.

Requirements

HDE depends on the following libraries:

numpy
scipy
keras
tensorflow
scikit-learn

Installation

With the necessary requirements all you need to do is clone the repository and pip install.

$ git clone https://github.com/hsidky/hde.git
$ pip install ./hde

Examples

Below is an example that demonstrates basic usage of HDE. Here we are using PyEMMA to extract features from a trajectory.

For the examples presented in the original paper see the paper_notebooks folder. For other detailed examples see the examples and notebooks folders.

import pyemma as py
from hde import HDE 

features = py.coordinates.featurizer('system.pdb')
features.add_backbone_torsions(cossin=True)
data = py.coordinates.load('trajectory.pdb', features=features)

n_components = 3 # Number of eigenfunctions to learn.
model = HDE(
    features.dimension(), 
    n_components=n_components, 
    n_epochs=20, 
    lag_time=100,
    batch_normalization=True
)

slow_modes = model.fit_transform(data)

Acknowledgements

The HDE method and code is inspired by and built upon VAMPNets and VDEs.

License

HDE is provided under an MIT license that can be found in the LICENSE file. By using, distributing, or contributing to this project, you agree to the terms and conditions of this license.

Cite

If you use this code in your work, please cite:

hsidky / srv