mosaic-clustering 0.3.1

Correlation based feature selection for MD data

Docs • Features • Installation • Usage • FAQ

Molecular Systems Automated Identification of Cooperativity

MoSAIC is an unsupervised method for correlation analysis which automatically detects the collective motion in MD simulation data, while simultaneously identifying uncorrelated coordinates as noise. Hence, it can be used as a feature selection scheme for Markov state modeling or simply to obtain a detailed picture of the key coordinates driving a biomolecular process. It is based on the Leiden community detection algorithm which is used to bring a correlation matrix in a block-diagonal form.

The method was published in:

G. Diez, D. Nagel, and G. Stock, Correlation-Based Feature Selection to Identify Functional Dynamcis in Proteins, J. Chem. Theory Comput., 2022, XXXX, XXX, XXX-XXX, 10.1021/acs.jctc.2c00337

If you use this software package, pleaes cite the above mentioned paper.

Features

• Intuitive usage via module and via CI
• Sklearn-style API for fast integration into your Python workflow
• No magic, only a single parameter which can be optimized via cross-validation
• Extensive documentation and detailed discussion in publication

Installation

The package is called mosaic-clustering and is available via PyPI or conda. To install it, simply call:

python3 -m pip install --upgrade mosaic-clustering

or

conda install -c conda-forge mosaic-clustering

or for the latest dev version

# via ssh key
python3 -m pip install git+ssh://[e-mail]/moldyn/MoSAIC.git

# or via password-based login
python3 -m pip install git+https://github.com/moldyn/MoSAIC.git

In case one wants to use the deprecated UMAPSimilarity or the module mosaic umap one needs to specify the extras_require='umap', so

python3 -m pip install --upgrade moldyn-mosaic[umap]

Shell Completion

Using the bash, zsh or fish shell click provides an easy way to provide shell completion, checkout the docs. In the case of bash you need to add following line to your ~/.bashrc

eval "$(_MOSAIC_COMPLETE=bash_source mosaic)"

Usage

In general one can call the module directly by its entry point $ MoSAIC or by calling the module $ python -m mosaic. The latter method is preferred to ensure using the desired python environment. For enabling the shell completion, the entry point needs to be used.

CI - Usage Directly from the Command Line

The module brings a rich CI using click. Each module and submodule contains a detailed help, which can be accessed by

$ python -m mosaic
Usage: python -m mosaic [OPTIONS] COMMAND [ARGS]...

  MoSAIC motion v0.3.1

  Molecular systems automated identification of collective motion, is
  a correlation based feature selection framework for MD data.
  Copyright (c) 2021-2022, Georg Diez and Daniel Nagel

Options:
  --help  Show this message and exit.

Commands:
  clustering  Clustering similarity matrix of coordinates.
  similarity  Creating similarity matrix of coordinates.
  umap        Embedd similarity matrix with UMAP.

For more details on the submodule one needs to specify one of the three commands.

A simple workflow example for clustering the input file input_file using correlation and Leiden with CPM and the default resolution parameter:

# creating correlation matrix
$ python -m mosaic similarity -i input_file -o output_similarity -metric correlation -v

MoSAIC SIMILARITY
~~~ Initialize similarity class
~~~ Load file input_file
~~~ Fit input
~~~ Store similarity matrix in output_similarity

# clustering with CPM and default resolution parameter
# the latter needs to be fine-tuned to each matrix
$ python -m mosaic clustering -i output_similarity -o output_clustering --plot -v

MoSAIC CLUSTERING
~~~ Initialize clustering class
~~~ Load file output_similarity
~~~ Fit input
~~~ Store output
~~~ Plot matrix

This will generate the similarity matrix stored in output_similarity, the plotted result in output_clustering.matrix.pdf, the raw data of the matrix in output_clustering.matrix and a file containing in each row the indices of a cluster.

Module - Inside a Python Script

import mosaic

# Load file
# X is np.ndarray of shape (n_samples, n_features)

sim = mosaic.Similarity(
    metric='correlation',  # or 'NMI', 'GY', 'JSD'
)
sim.fit(X)


# Cluster matrix
clust = mosaic.Clustering(
    mode='CPM',  # or 'modularity
)
clust.fit(sim.matrix_)

clusters = clust.clusters_
clusterd_X = clust.matrix_
...

Cross-Validation of Parameters

Selecting the optimal parameters, e.g., resolution_parameter, or n_clusters, can be quite difficult. Here we show a short example how one can use cross-validation for optimizing the parameters. Nevertheless, one should keep in mind that the here used silhouette score is not optimal for our task. Hence, the here obtained optimal parameters should be considered as a good first guess.

Here a figure visualizing the optimal cluster value n_clusters=12 and the code to produce it.

import mosaic
import numpy as np
import prettypyplot as pplt
from matplotlib import pyplot as plt

pplt.use_style(colors='tab20c', figsize=2.4)

# traj = np.loadtxt(filename)

# specify parameters grid
n_clusters = np.arange(2, traj.shape[1])
params = {'n_clusters': n_clusters}
search = mosaic.GridSearchCV(
    similarity=mosaic.Similarity(),
    clustering=mosaic.Clustering(
        mode='kmedoids',
        n_clusters=2,  # any dummy value is good here
    ),
    param_grid=params,
).fit(traj)

# plotting result
fig, ax = plt.subplots()
mean_score = search.cv_results_['mean_test_score']
std_score = search.cv_results_['std_test_score']

ax.fill_between(
    n_clusters,
    mean_score + std_score,
    mean_score - std_score,
    color='C2',
)
ax.plot(n_clusters, mean_score + std_score, c='C1')
ax.plot(n_clusters, mean_score - std_score, c='C1')
ax.plot(n_clusters, mean_score, c='C0')

ax.set_xlim([0, traj.shape[1]])
ax.set_xlabel(r'$k$ no. of clusters')
ax.set_ylabel(r'silhouette score')

pplt.savefig('cv_silhouette.pdf')

FAQ

How to load the clusters file back to Python?

Simply use the function provided in tools:

import mosaic

clusterfile = '...'
clusters = mosaic.tools.load_clusters(clusterfile)

Is it possible to use cross validation together with silhouette score?

The new release v0.3.0 refactored the classes, so that the sklearn.model_selection.GridSearchCV can be used. Check out the cv example.

I get an error.

Please open an issue.

Should I upgrade the package?

You can check out the CHANGELOG.md to see what changed.

How can I interpretate the results?

Check out our publication for two detailed examples.

Is it possible to install the CLI only?

Partially, yes. If you do not want to screw up your current Python environment there are multiples possibilities. Either create a virtual environment on your own via conda or venv, or you can simply use pipx

Is the silhouette method implemented?

Yes, simply use the score method implemented in the Clustering class.