slime2

This is a wrapper around SciKit Learn's Random Forest Classifier (rfc) routines. Given an OTU table and known classes for some of the samples, train a rfc on all the data and get the output: how well could the samples be classified, what's the out-of-bag error, and what are the important features?

Input

You'll need an OTU table in QIIME format (i.e., OTUs on the rows, samples on the columns). You'll also need a classes file that specifies what samples belong to what class.

The classes file can have two formats: two-column or one-column. In one-column format, each line specifies a sample-class pair, like

healthy_guy1	healthy
healthy_guy2	healthy
sick_guy1	sick
sick_guy2	sick

In two-column format, the first line of a block is the class name and the rest of the lines are sample names, like

# you can put in a comment
healthy
healthy_guy1
healthy_guy2

sick
sick_guy1
sick_guy2

Options

The command-line arguments will show you what kinds of options you can pass to the random forest classifier. Of special note in -n, the number of trees in the forest.

Make sure you put the --help after you've specified the classifier (viz., rf or ab) so that you can see the options specific to that classifier.

Output

All the output files are given a tag to help organize the output. By default, a tag is produced that is specific to that OTU table and class file. (It's a hex digest of the concatenation of the two files.)

File name	Contents
tag_classes.txt	the true classes that you specified
tag_cmd.txt	the command you input to run slime2
tag_featimp.txt	feature importances. column1 is the OTUs; col2 is the important; col3 is the cumulative importance
tag_params.txt	the parameters used when making the RFC
tag_results.txt	an explicit confusion matrix. if a sample was correctly classified, the line is "--". if a sample of true class X was misclassified as class Y, you get ">> X misclassified as Y".
tag_rfc.pkl	a pickled python object. it's the RFC object, which has a bunch of extra data attached to it.
tag_scores.txt	the "mean" score (i.e., the fraction of samples correctly classified by the RFC when trained on all data) and the out-of-bag score.

Common gotcha's

• If you get a ValueError, check the format of your classes file. Output from R can be problematic, since by default it writes the row numbers, which means you end up with too many columns.
• Beware of outputting information from Excel, which will use Windows-style newlines even when running on a Mac. You'll have to .

Upcoming features

• Cross-validation
• Universal newline support (cf. the gotcha's)