This module implements DISCERN, a method of identifying features whose conditional dependancy structure differs between two networks. An example is in the field of biology, where features may be genes, and the two networks may be those with cancer and those without cancer. DISCERN would identify those genes who are conditionally dependent on a different group of genes in the two networks.

DISCERN is available on PyPi, making installation as easy as running

pip install discern

Alternatively, you can clone this repo to get the cutting edge.

DISCERN makes use of rpy2, which can be a pain to install. If you do not have it installed, you have to set environment variables R_HOME and R_USER, where R_HOME is where your R installation is, and R_USER is where your python executable is (C:\Python27 or C:\Anaconda) usually. Once those are set properly, rpy2 is also pip installable with http://www.lfd.uci.edu/~gohlke/pythonlibs/

pip install rpy2

Many people have difficulty installing rpy2, but that means there is help online if you're having issues installing it. If you are on a windows, check out the windows binaries available at

Using DISCERN is simple. Here is a sample script, assuming that A and B are two datasets with the same features, but allowing different numbers of samples. DISCERN takes in pandas dataframes exclusively.

from discern import DISCERN

A, B = ** some pandas dataframes **
discern = DISCERN()

# Split the data into training and testing
A_train, A_test = A[::2], A[1::2]
B_train, B_test = B[::2], B[1::2]

scores = discern.fit_score( A_train, A_test, B_train, B_test, names=A.columns,
	l=0.3 )

print scores

This will return the scores. DISCERN makes use of Elastic Net regressions in order to build the underlying network, and so parameters lambda (l) and alpha (alpha) should be specified. alpha=1.0 means pure LASSO, and alpha=0.0 means pure ridge regression.

If you want to automatically determine lambda, you can pass in l='auto', or do it yourself with discern.lambda_opt( data, names... ), which will return both the optimal lambda and SSE across all regression problems which the optimal lambda minimizes.

DISCERN has built-in parallelization from R. For any function, simply pass in n_cores=n where n is the number of cores you want to use. Since these methods are natively scalable, the more cores the better. Not passing anything in assumes only using 1 core.