# Tutorial for using and training vica¶

## Installation¶

Instructions for Genepool.nersc.gov (more to come):

module unload python
module load python/3.6-anaconda_4.3.0
module load bbtools
module load pigz
module load prodigal
conda create -n vicaenv python=3.6
source activate vicaenv
conda config --add channels bioconda
conda install pyfaidx
conda install khmer
git clone https://github.com/USDA-ARS-GBRU/vica.git
cd vica
python setup.py build
pip install -e .
# pip install vica


## Classifying sequences with vica¶

To classify contigs using vica no additional data is needed except for the pre-trained model. minhash sketched are sent to minhash server to avoid having to download refseq minhash files:

vica classify [options]


## Training models with vica¶

To train the model without evaluation we use all the available data.

• Download the NCBI taxonomy data and process the data for use by bbtools using the fetchTaxonomy script:

vica/pipelines/fetchTaxonomy.sh

• Download the Refseq data and process for use by bbtools using using the script:

vica/pipelines/fetchRefSeq.sh

• If you want to add additional fastas please format them with a header in the format:

tid|[taxid number]|[any identifier]| optional info...

• The taxid does not need to be at the species level but it must be equal to or lower than the classification group to be included. For example, If you have sequences you want to include in the group bacteria (taxid 2) you can use a family or genus level taxid.

• Shred the data without splitting using the command:

vica shred [options]

• Extract the features for each fasta using:

vica get_features [options]

• Feature selection is embarrassingly parallel. If you are working on an HPC system with a job scheduler you can split your files up and submit them in parallel. At the end, all of the TFrecord feature files can be combined into a single file with the command

cat *.tfrecord > all.tfrecord

• train the model with:

vica train [options]

• During training you can monitor the process using tensorflow’s tensorboard:

tensorboard --logdir [directory]


## Training models and and evaluating performance with vica¶

The process of training and evaluating a model is the most complex because data for evaluation (the test data) must be split from the data for training the model to get an accurate assessment of performance. In many Machine learning applications this is a trivial matter of random data selection, but because our data are taxonomically structured and some taxa are overrepresented, a simple random split is insufficient. This means that we need to split out test and training data at the level of taxonomic novelty we are hoping to discover.

For the most accurate assessment of the classifier performance we also must exclude test data from and databases used to create features. We currently use Three feature sets: Codon usage and 5mers (which do not contain external databases) and the phylum level taxonomic assignment from minhash sketches, which uses an external database. Codon usage and 5mer feature sets help the model to generalize well and are used in the deep neural network portion of the classifier. The purpose of the minhash feature set is quite different. It serves the purpose of improving classifier precision by making sure that known non-viral taxa do not enter the viral bin. This feature set is incorporated using a logistic regression model.

All of this information from the logistic regression and the DNN are combined by jointly training both models. Then, “the wide component and deep component are combined using a weighted sum of their output log odds as the prediction, which is then fed to one common logistic loss function for joint training.” (Cheng et al., 2015). This means that the predictive ability of the minhash features must be realistic or else too much weight could be given to these features. For that reason we leave the test taxa out of the minhash database used to generate minhash features.

The overall process for training from scratch using the Refseq dataset is this:

• Download the NCBI taxonomy data and process the data for use by bbtools using the script:

vica/pipelines/fetchTaxonomy.sh

• Download the Refseq data and process for use by bbtools using using the script:

vica/pipelines/fetchRefSeq.sh

• Split the test and train data, and fragment it into fragments of the selected length using the command:

vica split_and_shred [options]


Currently we use 4 classes (Viruses, bacteria, archaea, eukaryotes). This will create a directory with test and train folders. Each folder will contain 1 folder per class, and a file contain the taxids to be excluded from RefSeq.

• Extract the features for each fasta using:

vica get_features [options]


The feature selection is embarrassingly parallel if you are working on an HPC system with a job scheduler you can split your files up and submit them in parallel. at the end all of the TFrecords files can be combined into a single file with the command

cat *.tfrecord > all.tfrecord

• train the model with:

vica train [options]


During training you can monitor the process using tensorflow’s tensorboard:

tensorboard --logdir [directory]

• evaluate the model with:

vica evaluate [options]