Algorithm::TrunkClassifier - Implementation of the Decision Trunk Classifier algorithm
use Algorithm::TrunkClassifier qw(runClassifier);
This module contains the implementation of the Decision Trunk Classifier. The algorithm can be used to perform binary classification on numeric data, e.g. the result of a gene expression profiling experiment. Classification is based on so-called decision trunks, which consist of a sequence of decision levels, represented as nodes in the trunk. For each decision level, a probe is selected from the input data, and two decision threshold are calculated. These threshold are associated to two outgoing edges from the decision level. One edge represents the first class and the other edge represents the second class.
During classification, the decision levels of a trunk are considered one at a time. To classify a sample, its expression of the probe at the decision level is compared to the thresholds of outgoing edges. If the expression is less than the first threshold, class1 is assigned to the sample. If, on the other hand, the expression is greater than the second threshold, class2 is assigned to the sample. In the case expression is in-between the thresholds, the algorithm proceeds to the next decision level of the trunk.
By default, classification is done by leave-one-out cross validation (LOOCV) meaning that a single sample is used as test set, while the remaining samples are used to build the classifier. This is done for every sample in the input dataset. See the the algorithm publication for more details. A PubMed link can be found in "SEE ALSO".
Following installation, the algorithm can be run from the terminal using the run_classifier.pl script supplied in the t/ folder. The command should be in this form
perl run_classifier.pl [Options] [Input data file]
The last argument must be the name of the input data file containing the expression data in table format, where columns are tab-separated. The first row must contains column names and the first column must contain row names. Samples need to be given in columns and probes/attributes in rows. Before the name of the input data file, a number of optional arguments may be given, see "OPTIONS" below. A data file containing random data is provided in the t/ folder.
At the top of the input data file, before the expression data table, an number of meta data rows starting with # can be given. The purpose of these rows is to tell the algorithm what classification variables that are defined for the data and what classes the samples belong to. The classification variable is the name of the property by which the samples are divided into two groups. For example, if the samples should be classified as either early or late stage cancer, the name of the classification variable would be STAGE.
Classification variables are defined on rows starting with #CLASSVAR, followed by the name of the variable and the two class labels.
#CLASSVAR name classLabel1 classLabel2
Class labels (e.g. EARLY and LATE) are assigned to samples on rows staring with #CLASSMEM, followed by the name of the classification variable and the class labels for all samples.
#CLASSMEM name sampleOneClass sampleTwoClass sampleThreeClass ...
Since this would be very tedious to fill in manually for large datasets, the algorithm accepts a supplementary file with class information for all samples. See the
-s value option below. An example of a supplementary file is given in the t/ folder.
An example of meta rows for the classification variable STAGE in a dataset with five samples is
#CLASSVAR STAGE EARLY LATE #CLASSMEM STAGE LATE LATE EARLY LATE EARLY
By default, the algorithm trains and classifies one dataset using leave-one-out cross validation. Two other classification procedures are supported: split-sample and dual datasets. The split-sample procedure causes the input dataset to be split into two sets, a training set and a test set. Trunks are built using the training set and classification is done only on the test set. It is also possible to supply two datasets. The input data file (final command line argument) is then used as training set and the second dataset as test set. Thus the value for the -c option should be
loocv, for leave-one-out cross validation,
split for split-sample procedure, or
dual when using two datasets.
The percentage of samples in the input data file that should be used as test set when using
-p split. Must be from 1 to 99. Default is 20.
The name of the testset data file when using the
-p dual option.
The value should be the name of the classification variable to use. Default is TISSUE.
The value should be the name of the output folder. Created if it does not exist in the current directory. Default is current directory.
By default, the algorithm selects the number of decision levels to use for classification. To override this, supply the -l option and an integer from 1 to 5. This will force the algorithm to use that number of decision levels.
This option can be used to inspect the dataset without running the classifier. The option takes one of three possible values:
samples: prints the number of samples in each class for the classification variable probes: prints the number of probes in the dataset classes: prints all classification variables in the meta data
Name of a supplementary file containing class information for the samples in the dataset. The contents should be in table format with columns being tab-separated. The first row needs to contain column names and the first column should contain sample names. The second and subsequent columns can contain class information, with the name of the classification variable given as the column name, followed by class labels on the rows starting with sample names. Examples of classification variables are STAGE, GRADE and HISTOLOGY. Class labels could be EARLY and LATE for STAGE, or LOW and HIGH for GRADE. The format of the file is illustrated here.
Samples ClassVar1 ClassVar2 sample1 classLabel1 classLabel3 sample2 classLabel1 classLabel4 sample3 classLabel2 classLabel3 sample4 classLabel1 classLabel4 sample5 classLabel2 classLabel4
When this option is given, the algorithm first processes the supplementary file and writes a new data file containing meta data. This data file is then used as input.
Note: If the
-p dual option is used, two datasets must be supplied. In this case the supplementary file needs to contain the class information of all samples in both datasets.
This option makes the algorithm report its progress to the terminal during a run.
This option circumvents selection of decision levels and makes the algorithm use trunks with 1, 2, 3, 4 and 5 decision levels during classification.
This option causes argument documentation to be printed to the terminal.
The algorithm produces five files as output: performance.txt, loo_trunks.txt, cts_trunks, class_report.txt and log.txt. The classification accuracy can be found in performance.txt. In case of leave-one-out cross validation, the accuracy for each fold is reported along with the average accuracy across all folds. Since the test set consists of one sample, the accuracy of one LOOCV fold is either 0 % (wrong) or 100 % (correct). For split-sample and dual datasets classification, only the average accuracy is reported since there is only one test set.
The loo_trunks.txt file contains the decision trunks resulting from leave-one-out training on the training set. Since the training set is different in each fold, different probes may be selected in the trunks. The decision levels of a trunk are shown in order starting with the first level at the top. Each level consists of two rows: the first row shows the name of the probe and the second row contains the decision thresholds and the associated class labels. An illustration of a decision trunk with three levels is shown here
Probe X <= A (class1) > B (class2) Probe Y <= C (class1) > D (class2) Probe Z <= E (class1) > F (class2)
Classification of a sample S using this decision trunk would proceed as follows.
Compare the expression of probe X in sample S with thresholds A and B. If the expression is less than A, the sample is classified as class1. If the expression is greater than B, the sample is classified as class2. If the expression is in-between A and B, the algorithm proceeds to the next decision level. This continues until the last level, where the thresholds E and F are equal, meaning that sample S is guaranteed to be classified as either class1 or class2.
The cts_trunks.txt file contains decision trunks built using the complete training set.
The classification of each sample can be found in the class_report.txt file. The rows in this file start with a sample name, followed by "in X-class". X is the level in the decision trunk where the sample was classified, and class is the class label assigned to the sample.
The log.txt file gives a summary of the classifier run. The information given includes the name of the input data file, the name of the test data file (if any), the name of the classification procedure, the split-sample percentage (if any), number of decision levels used for classification, the name of the classification variable, the sizes of class1 and class2 in the training and test set respectively, and the version of the algorithm.
In case the
-u option is used, the output files will contain the results from using decision trunks with 1, 2, 3, 4 and 5 levels.
To provide an easy way of testing the algorithm, the t/ folder contains two test files. The test_data.txt contains a random dataset with 200 samples and 1000 probes. This set has been generated such that the first 100 samples (healthy) have a mean gene expression of 0 and standard deviation of 0.5 (normal distribution) for all genes, while the remaining 100 samples (malignant) have a mean of 1 and standard deviation of 0.5. The test_supp.txt is a supplementary file containing the class information associated to the random dataset. To run the algorithm with this dataset, use the following command.
perl run_classifier.pl -v -o test_set_tissue -s test_supp.txt test_data.txt
Since a supplementary file is given, a new data file with class information will be written. Following this, the algorithm will build decision trunks and determine how many decision levels to use for classification. Finally, LOOCV will be performed using the selected trunks and output written. If no classification variable is explicitly given, the algorithm will default to TISSUE. For the random dataset, this variable states if the sample comes from healthy tissue or from a tumor. The supplementary file labels healthy samples as T_HEALTHY and tumor samples as T_MALIGN. By looking in the supplementary file it can also be seen that the random dataset comes with a second classification variable: GRADE. This variable states if the tumor samples comes from an low- or high-state tumor. This is indicated by G_LOW and G_HIGH. Since the healthy samples do not come from tumors, they do not have GRADE classes. To indicate this, #NA is used. The #NA symbol is interpreted by the algorithm as a null class, causing the sample to be excluded if GRADE is given as the classification variable. To test this, use the following command.
perl run_classifier.pl -v -c GRADE -o test_set_stage -s test_supp.txt test_data.txt
By comparing the output files, differences can be seen in how many folds of LOOCV has been carried out, and in what probes where selected for the decision trunks. The log file will also reflect that a different classification variable was used. Accuracy will be good when classifying TISSUE, because the healthy and tumor samples have sufficiently different gene expression values. For GRADE, however, all tumor samples have the same mean and standard deviation, so the algorithm is not able to separate them.
If an invalid argument is given, or if there is something wrong with the input data file or supplementary file, the algorithm will output a warning or error message. Warnings will not prevent the algorithm from running, but errors will. Here is a list of all warnings/errors and how to interpret them
Indicates that the supplementary file has less than two columns. The algorithm expects the first column to contain sample names and the following columns to contain class labels of samples. The first row in the file must contain the names of the columns. For column 2, 3 and so on, the column name should be the classification variable name.
Indicates that a class label is missing on line index. The missing label is replaced with #NA, the symbol for the null class.
Indicates that no class labels (rows) were found in the supplementary file.
Indicates that sample sampleName in the input/testset data file is missing a class label for classVar in the supplementary file. The sample's class label becomes #NA.
Indicates that a meta data row (starting with #CLASSVAR) is missing the name value. The expected format is
#CLASSVAR name classLabel1 classLabel2
Indicates that a meta data row (starting with #CLASSVAR) is missing one/both the classLabel values.
Indicates that a meta data row (starting with #CLASSMEM) is missing the name value. The expected format is
#CLASSVAR name sampleOneClass sampleTwoClass sampleThreeClass ...
Indicates that two samples names in the input or testset data file are identical. This does not affect classification.
Indicates that an expression value for probe index is missing.
Indicates that the number of levels given to the
-l argument was to high. This can happen when there are not enough samples to create five levels in a trunk. If 4 is given to the
-l argument but only two levels could be created, the algorithm will use two levels instead.
Indicates that the command line option com was not recognized as a valid command.
Indicates that command line option com did not accept arg as an argument.
Indicates that command line option com requires an argument but none was supplied.
Indicates that an input data file (final command line argument) was not supplied.
Indicates that the -t command line option for supplying a test dataset is missing. This must be given when -p dual is used.
Indicates that the file filename given as argument to the
-s option was not readable.
Indicates that the number of columns at line index differs from the number of column names on the first row. The first row defines the correct number of columns, and every subsequent row must have the same number of columns.
Indicates that the sample name (first column) is missing on line index.
Indicates that the class variable column with name classVar contains fewer/more than two different class labels. Since the algorithm has been designed to only handle binary classification, a classification variable is only allowed to have two class labels.
Indicates that the input/testset data file filename was not readable. The input data file must be given as the last command line argument. A testset data file must be supplied using the -t option when the
-p dual option is used.
Indicates that the new input data file with meta data could not be written.
Indicates that the input/testset data file, which is supposed to contain the expression data table, does not contain any sample columns.
Indicates that one of the classLabel values for a #CLASSVAR meta data row is equal to the null class symbol #NA. The classLabel values are the class labels for the classification variable and cannot be equal to #NA, because this symbol is reserved as a null class symbol.
Indicates that the classification variable name given to the -c option is missing in the meta data of the input or testset data file.
Indicates that the number of class labels on the #CLASSMEM row for classVar is different from the number of samples in the input or testset data file.
Indicates that one or more class labels on the #CLASSMEM row for classVar are invalid. Valid class labels are those given on the #CLASSVAR row for classVar and the null class symbol #NA.
Indicates that no samples have the class label classLabel for classification variable classVar. This means that classification cannot be carried out, since all samples belong to the same class.
Indicates that a probe row in the input or testset data file has wrong number of columns with respect to the number of samples.
probenamein input data file not found in testset data file
Indicates that a probe in the input data file is missing in the testset. For classification to be carried out using two datasets, all probes in the input data file must be present in the testset.
Indicates that output files were not writable in the output folder.
The publication describing the algorithm can be found in PubMed by this link: http://www.ncbi.nlm.nih.gov/pubmed?Db=pubmed&Cmd=DetailsSearch&Term=23467331%5Buid%5D
None by default. The runClassifier subroutine is exported on request.
Benjamin Ulfenborg, <email@example.com>
Copyright (C) 2012 by Benjamin Ulfenborg
This module is free to use, modify and redistribute for academic purposes.