Unrolr

Conformational analysis of MD trajectories based on (pivot-based) Stochastic Proximity Embedding using dihedral distance as a metric (https://github.com/jeeberhardt/unrolr).

Prerequisites

You need, at a minimum (requirements.txt):

• Python
• NumPy
• H5py
• Pandas
• Matplotlib
• PyOpenCL
• MDAnalysis

Installation on UNIX (Debian/Ubuntu)

1 . First, you have to install OpenCL:

• MacOS: Good news, you don't have to install OpenCL, it works out-of-the-box. (Update: bad news, OpenCL is now depreciated in macOS 10.14. Thanks Apple.)
• AMD: You have to install the AMDGPU graphics stack.
• Nvidia: You have to install the CUDA toolkit.
• Intel: And of course it's working also on CPU just by installing this runtime software package. Alternatively, the CPU-based OpenCL driver can be also installed through the package pocl (http://portablecl.org/) using Anaconda.

For any other informations, the official installation guide of PyOpenCL is available here.

2 . I highly recommand you to install the Anaconda distribution (https://www.continuum.io/downloads) if you want a clean python environnment with nearly all the prerequisites already installed. To install everything properly, you just have to do this:

$ conda create -n unrolr python=3
$ conda activate unrolr
$ conda install -c conda-forge mkl numpy scipy pandas matplotlib h5py MDAnalysis pyopencl ocl-icd-system

3 . Install unrolr

$ pip install unrolr

... or from the source directly

$ git clone https://github.com/jeeberhardt/unrolr
$ cd unrolr
$ python setup.py build install

OpenCL context

Before running Unrolr, you need to define the OpenCL context. And it is a good way to see if everything is working correctly.

$ python -c 'import pyopencl as cl; cl.create_some_context()'

Here in my example, I have the choice between 3 differents computing device (2 graphic cards and one CPU).

Choose platform:
[0] <pyopencl.Platform 'AMD Accelerated Parallel Processing' at 0x7f97e96a8430>
Choice [0]:0
Choose device(s):
[0] <pyopencl.Device 'Tahiti' on 'AMD Accelerated Parallel Processing' at 0x1e18a30>
[1] <pyopencl.Device 'Tahiti' on 'AMD Accelerated Parallel Processing' at 0x254a110>
[2] <pyopencl.Device 'Intel(R) Core(TM) i7-3820 CPU @ 3.60GHz' on 'AMD Accelerated Parallel Processing' at 0x21d0300>
Choice, comma-separated [0]:1
Set the environment variable PYOPENCL_CTX='0:1' to avoid being asked again.

Now you can set the environment variable.

$ export PYOPENCL_CTX='0:1'

Example

from unrolr import Unrolr
from unrolr.feature_extraction import Dihedral
from unrolr.utils import save_dataset


top_file = 'examples/inputs/villin.psf'
trj_file = 'examples/inputs/villin.dcd'

# Extract all calpha dihedral angles from trajectory and store them into a HDF5 file
d = Dihedral(top_file, trj_file, selection='all', dihedral_type='calpha').run()
X = d.result
save_dataset('dihedral_angles.h5', "dihedral_angles", X)

# Fit X using Unrolr (pSPE + dihedral distance) and save the embedding into a csv file
# The initial embedding is obtained using PCA (init = 'pca') with the OpenCL implementation
# to run SPE, a CPU implementation can be used as an alternative (platform='CPU')
U = Unrolr(r_neighbor=0.27, n_iter=50000, init='pca', platform='OpenCL', verbose=1)
U.fit_transform(X)
U.save(fname='embedding.csv')

print('%4.2f %4.2f' % (U.stress, U.correlation))

Todo list

• Compare SPE performance with UMAP
• Compatibility with python 3
• Compatibility with the latest version of MDAnalysis (==0.17)
• Unit tests
• Accessible directly from pip
• Improve OpenCL performance (global/local memory)

Citation

Eberhardt, J., Stote, R. H., & Dejaegere, A. (2018). Unrolr: Structural analysis of protein conformations using stochastic proximity embedding. Journal of Computational Chemistry, 39(30), 2551-2557. https://doi.org/10.1002/jcc.25599

License

MIT