Identification of True Somatic Variants Using Machine Learning

Usage

Requirements

Environment

• Python 2.7
• virtualenv

Per Sample

• two tab-delimited files containing variants in format Chr Pos Allele, one file containing positive variants and another containing artifacts
• a sorted sample BAM and index file
• two sorted negative-control BAM files and their respective index files
• at least one other sorted sample BAM file sequenced in the same batch

Setup

Create a Python environment via virtualenv, activate it, and install the provided requirements.

$ virtualenv --python=python2.7 som-var-env
$ source som-var-env/bin/activate
$ pip install -r requirements.txt

Generating Features

In order to train/apply a Random Forest model, features need to be generated for coverage, bias, VAF, similarity score vs two control samples, and similarity vs a intra-run batch sample. This process can be automated using auto_featuregeneration.sh, a wrapper for the featuregeneration.py script.

Once a datafile (e.g. inputs.txt) is in the proper tab separated format SAMPLE_ID CLASS_TYPE VARIANT_FILE SAMPLE_BAM NORMAL_BAMS BATCH_BAMS, run:

$ ./auto_featuregeneration.sh inputs.txt

This will create a features/ directory and populate it with an updated "VARIANT_FILE" that now includes features.

Example

For two samples sequenced in the same batch, "TEST01" and "TEST02", inputs.txt would contain 4 lines:

TEST01  NEG test01_neg.txt  test01.bam  ctrl1.bam,ctrl2.bam batchX-1.bam,batchX-2.bam
TEST01  POS test01_pos.txt  test01.bam  ctrl1.bam,ctrl2.bam batchX-1.bam,batchX-2.bam
TEST02  NEG test02_neg.txt  test02.bam  ctrl1.bam,ctrl2.bam batchX-1.bam,batchX-2.bam
TEST02  POS test02_pos.txt  test02.bam  ctrl1.bam,ctrl2.bam batchX-1.bam,batchX-2.bam

If "TEST01" and "TEST02" were sequenced in different batches, inputs.txt would change for the BATCH_BAMS field:

TEST01  NEG test01_neg.txt  test01.bam  ctrl1.bam,ctrl2.bam batchX-1.bam,batchX-2.bam
TEST01  POS test01_pos.txt  test01.bam  ctrl1.bam,ctrl2.bam batchX-1.bam,batchX-2.bam
TEST02  NEG test02_neg.txt  test02.bam  ctrl1.bam,ctrl2.bam batchY-1.bam,batchY-2.bam
TEST02  POS test02_pos.txt  test02.bam  ctrl1.bam,ctrl2.bam batchY-1.bam,batchY-2.bam

Training a Model

While a single "VARIANT_FILE" can be used to a train a model, this is not very useful. All the files in the new features/ directory can be combined into one text file using merge_features.py:

$ python merge_features.py --directory features/ --output_filename allfeatures.txt

train_model.py makes use of sklearn's RandomForestClassifier utility and Pickle to save the trained model for later use. A single model can be created with default arguments:

$ python train_model.py --training_data allfeatures.txt

A brief report will be output to the console displaying the parameters provided to the classifier, the Pickle filename, as well as any variants that were mislabeled by this newly trained model.

This tool also provides arguments for creating many models iteratively. This is done by using a combination of a) the number of models to include in the Random forest, b) the depth of each tree in the forest, and c) weights for the positive and negative classes.

For example, over 1,000 models can be generated with one command:

$ python train_model.py \
    --training_data allfeatures.txt \
    --iterate \
    --trees 50 201 50 \
    --leaves 8 22 2 \
    --weights 1 101 10 1 101 10

Applying a Model

A model that was saved can be loaded and applied to another dataset made with merge_features.py by using apply_model.py:

$ python apply_model.py \
    --pickle_model mymodel.sav \
    --test_set mytestset.tsv \
    --lower_limit 0.05 \
    --upper_limit 0.90

where --lower_limit and --upper_limit are the boundaries of the "uncertain" region. Ideally, no misclassifications occur outside the region of uncertainty. This can be a major criteria in model selection.

This tool will output a histogram of confidence interval scores for all variants in the --test_set as well as any misclassifications, if applicable.