supervised_learning.py – Run supervised classification using OTUs as predictors and a mapping file category as class labels.

Description:

This script trains a supervised classifier using OTUs (or other continuous input sample x observation data) as predictors, and a mapping file column containing discrete values as the class labels.

Outputs:

• cv_probabilities.txt: the label probabilities for each of the given samples. (if available)
• mislabeling.txt: A convenient presentation of cv_probabilities for mislabeling detection.
• confusion_matrix.txt: confusion matrix for hold-out predictions.
• summary.txt: a summary of the results, including the expected generalization error of the classifier
• feature_importance_scores.txt: a list of discriminative OTUs with their associated importance scores (if available)

It is recommended that you remove low-depth samples and rare OTUs before running this script. This can drastically reduce the run-time, and in many circumstances will not hurt performance. It is also recommended to perform rarefaction to control for sampling effort before running this script. For example, to rarefy at depth 200, then remove OTUs present in < 10 samples run:

single_rarefaction.py -i otu_table_filtered.txt -d 200 -o otu_table_rarefied200.txt filter_otu_table.py -i otu_table_rarefied200.txt -s 10

For an overview of the application of supervised classification to microbiota, see PubMed ID 21039646.

This script requires that R be installed and in the search path. To install R visit: http://www.r-project.org/. Once R is installed, run R and excecute the command “install.packages(“randomForest”)”, then type q() to exit.

Usage: supervised_learning.py [options]

Input Arguments:

Note

[REQUIRED]

-i, --input_data

Input data file containing predictors (e.g. otu table)

-m, --mapping_file

File containing meta data (response variables)

-c, --category

Name of meta data category to predict

[OPTIONAL]

-o, --output_dir

The output directory [deafult: .]

-f, --force

Force overwrite of existing output directory (note: existing files in output_dir will not be removed) [default: None]

--ntree

Number of trees in forest (more is better but slower) [default: 500]

-e, --errortype

Type of error estimation. Valid choices are: oob, loo, cv5, cv10. oob: out-of-bag, fastest, only builds one classifier, use for quick estimates; cv5: 5-fold cross validation, provides mean and standard deviation of error, use for good estimates on very large data sets; cv10: 10-fold cross validation, provides mean and standard deviation of error, use for best estimates; loo: leave-one-out cross validation, use for small data sets (less than ~30-50 samples) [default oob]

Output:

Outputs a ranking of features (e.g. OTUs) by importance, an estimation of the generalization error of the classifier, and the predicted class labels and posterior class probabilities according to the classifier.

Simple example of random forests classifier:

supervised_learning.py -i otutable.txt -m map.txt -c Individual -o ml

Simple example, filter OTU table first:

single_rarefaction.py -i otu_table_filtered.txt -d 200 -o otu_table_rarefied200.txt
filter_otu_table.py -i otu_table_rarefied200.txt -s 10
supervised_learning.py -i otutable_filtered_rarefied200.txt -m map.txt -c 'Individual' -o ml

Running with 10-fold cross-validation for improved estimates of generalization error and feature importances:

supervised_learning.py -i otutable.txt -m map.txt -c Individual -o ml -e cv10

Running with 1,000 trees for improved generalization error:

supervised_learning.py -i otutable.txt -m map.txt -c Individual -o ml --ntree 1000