Machine learning models for VirFinder trained using the Tara Ocean viromes and metagenomes
Logistic classifier for the classification of viral and non viral contig sequences from marine metagenomes. The model was trained on Tara Ocean metagenomes and viromes (the list of samples used for the training is available in Model/list.txt and raw sequence files available in Model/Training_set).
The Tara-trained model was trained using VirFinder training function, on 10 000 viral sequences and 10 000 non viral sequences of 5000pb with a kmer size of 8pb.
The installation of VirFinder is described here https://github.com/jessieren/VirFinder
To use the custom Tara-trained model, download the model from Models/VF.trainModUser.mod1_Tara_10k.rda
In order to use the model on contigs, use the following code :
library(VirFinder)
### load the model
modFile <- "<path_to_the_model_file>/VF.trainModUser.model_Tara_10k.rda"
load(modFile)
### predict the contigs using the customized model
inFaFile <- system.file("data", "<path_to_the_contig_file>/contigs.fa")
predResultUser <- VF.pred.user(inFaFile, VF.trainModUser)
predResultUser
Assembled sequences from the Tara Ocean expedition were downloaded from EBI metagenomics (data available at https://www.ebi.ac.uk/ena/about/tara-oceans-assemblies). The list of metagenomes assemblies used is available in Document/list.txt. Contigs with a size less than 5kb were filtered out, and Centrifuge v1.0.4-beta (https://github.com/infphilo/centrifuge) was run on the remaining sequences using the Bacteria, Archaea, Human and Virus compressed index (updated 12/06/2016), using the default settings. Sequences in the microbiomes with viral matches were removed from the negative training set. Sequences in viromes with a Bacterial, Archeal or Human match were removed from the positive training set. The viromes were further cleaned using BLAST against prokaryotic genomes (reference prokaryotic genomes released on ftp://ftp.ncbi.nlm.nih.gov/blast/db/ last modified on 10/29/18). A cutoff for the e-value was set at 0.01. Sequences with a significant hit against this prokaryotic database was removed from the training set. The remaining sequences were broken down to 5000bp to be used as training set. The training set used for the model is available in Model/Training_set
VirFinder version 1.1 available at https://github.com/jessieren/VirFinder was used, and the Tara-trained models were trained using the built-in training function, using a kmer-size of 8bp and a training set of 10 000 viral and 10 000 non-viral sequences randomly selected from the cleaned Tara metagenomic sequences broken down to 5000bp. The model is available in: Model/VF.trainModUser.model_Tara_10k.rda