A genetic algorithm to designing distributed cell classifier circuits.

RAccoon has the following dependencies:

• Python 3+

Check requirements.txt for required packages or use:

$ conda create --name <env_name> --file requirements.txt

To download the RAccoon from Github, do:

$ git clone https://github.com/MelaniaNowicka/RAccoon

Use the following format of the .csv file: the first column includes unique IDs of samples, the second column includes annotation (0 - negative samples, 1 - positive samples), the following columns include miRNA profiles. Use semicolon as a separator. An example may be found below:

Continuous data:

If you use continuous data, keep the discretization on.

Discretized data:

To train a classifier on train data run:

python raccoon.py --train train_data.csv 

Use exemplary data to try it: train_data.csv, test_data.csv.

Description of parameters used from command line:

--train - training data set in the .csv format (obligatory)

--test - testing data set in the .csv format (default: None)

Training and test data should be formatted according to the description in the Data format section.

--config - config file name

Here, you can use most of the following parameters using config file instead of the command line. Check the exemplary config file here.

--rules - path to a file of pre-optimized rules (default: None)

--filter - filtering non-relevant features (default: t, f to turn off)

Features that are non-relevant (columns filled with only 0s or 1s) are filtered out as such features are not informative.

--discretize - discretize the data (default: t, f to turn off)

--mbin - discretization parameter: m segments (default: 50)

--abin - discretization parameter: alpha (default: 0.5)

--lbin - discretization parameter: lambda (default: 0.1)

Data discretization according to Wang et al. (2014). To know more please look into the mentioned publication.

-c - maximal size of a classifier (maximal number of single rules in the classifier, default: 5)

-a - classification threshold (default: None)

If classification threshold is set to a certain value (e.g., 0.5) the threshold is fixed for all classifiers in the GA run, if None - different values of thresholds (0.25, 0.45, 0.5, 0.75 and 1.0) are randomly assigned to the classifiers.

-w - multi-objective function weight (default: 0.5)

-u - uniqueness option (related to calculation of CDD classifier score, default: True)

-i - number of iterations without improvement after which the algorithm terminates (default: 30)

-f - number of fixed iterations after which the algorithm terminates (default: None)

-p - population size (default: 300)

--elitism - if True the best found solutions are added to the population in each selection operation (default: True)

--poptfrac - pre-optimized fraction of population, the rest of solutions is generated randomly (default: 0.5)

-x - crossover probability (default: 0.8)

-m - mutation probability (default: 0.1)

-t - tournament size (default: 0.2)

Run the analysis using:

python run_tests.py --train train_data.csv --config config_tuning.ini [--val validation_data.csv --test test_data.csv --rules rule_file.csv --run_id name_of_the_run]

Use exemplary data to try it: train_data.csv, test_data.csv.

python run_tests.py --train train_data.csv --test test_data.csv --config config_tuning.ini

You may change all the parameters in config.ini. Description may be found here.

Output log description:

READING CONFIG - config parameter values

READING DATA - data processing information

PARAMETER TUNING - parameter tuning section including data division and pre-processing information as well as parameter tuning results

FINAL TEST - results of the final tests (the classifiers are trained with tuned parameters and tested on test data)

simDataGenerator allows to generate a simulated GED data set with compcodeR package, preprocess it by splitting into train/test data sets and normalize with TMM normalization method (edgeR).

LIBRARIES REQUIRED: compcodeR, edgeR, matrixStats

Run prepareSimulatedDataset() with parameters:

n.genes - number of genes, e.g., 500

samples.per.cond - number of samples per class, e.g., 100

n.diffexp - number of differentially expressed genes, e.g., 50 (10%)

fraction.upregulated - fraction of differentially expresed genes that are upregulated, e.g., 0.5

random.outlier.high.prob - number of random outliers (higher values), e.g., 0.5

random.outlier.low.prob - number of random outliers (lower values), e.g., 0.5

train.fraction - fraction of data that becomes training data set, e.g, 0.8 (1-train.fraction = test.fraction)

is.seed - set to TRUE to be able to reproduce the results for certain conditions, set to FALSE if you want to generate different data sets

generateSummary - set to TRUE if you want to generate a compCodeR data report

imbalanced - set to TRUE if the data set should be imbalanced