frobnitzem / classifier

Monte Carlo Sampler for Bernoulli Mixture Models

Multivariate Bernoulli Mixture Model Sampler

This project implements a Monte Carlo sampler for the above model. When fed data on pair-contacts, it will automatically identify prototypical shapes.

All the source files in this distribution were written by David M. Rogers. If you make use of this code, we ask you to read and cite the following reference:

David M. Rogers, Protein Conformational States: A First Principles Bayesian Method. submitted, 2020. arXiv:2008.02353.

Lots of helper programs are included to analyze protein conformations.

They are being made available under the terms of the GNU GPL version 3. A copy of that license must accompany this work and all derivatives (see LICENSE.txt).

Incuded Programs

This source distribution contains several top-level programs.

• classify.py features.npy runs the classification algorithm based on an NxM boolean matrix. The matrix should be stored in compressed format using numpy's packbits.

  At present, the number of MCMC chains and samples to collect is selected by editing the source code. Running this program generates indices.npy, listing the indices of important features used for classification. It also generates several sub-directories named sz1, sz2, etc. Each contins features.npy which is a Kxlen(indices) matrix of feature probabilities (0 <= mu <= 1).

• reclassify.py indices.npy features.npy assigns categories to each input sample, and saves a file, rep.txt, listing the population in each cluster and the 1-based index of its "most representative" sample from the input.

• pull.py is a helper script that parses the rep.txt files from above and pulls out the corresponding frames from a DCD trajectory. This generates a multi-model PDB file showing the examplary structures from each conformational state.

• plot.py is a helper script that reads the category labels and permutes an all-all RMSD matrix - so that categories appear in blocks.

• test_bbm.py is the top-level testing routine that checks key properties of the implementation. When run with the right command-line arguments, it reproduces the test results shown in the publication.

  Some helpful code bits for working with these test results are in the test/ subdirectory.

• compare2.py calculates the Bhattacharyya Similarity between two different classifications.

• diff.py compares structures within each

• plotP.py creates a scatter-plot of cluster assignment probabilities. The plot is projected down to 2D by using a weighted sum of vertices -- where the probability of assignment to each cluster is the weight. Each vertex is the corner of a K-sided polyhedron.

The programs above are mostly self-documenting. A few of them depend on ucgrad, which is available in one of the subdirectories of forcesolve. Email the author if you're interested in turning that into a formal python package.

Protein Helpers

The prot_helpers subdirectory contains useful source code for generating pair distance maps. The main program, contacts.py, passes each frame to a user-specified python function which generates the frame's binary fingerprint.

An example fingerprint generator is provided as mpro.py. This script generates pair contacts that focus on a catalytic site involving residues 41, 145, and 188 of chain A.

How to Contribute

Submit an issue describing your use case, and we'll discuss. Some things on my wishlist include:

• maximization of the likelihood function from a given starting point

• A more user-friendly interface for classify.py

• Additional tests created from simple classification problems.