Discovering Deeply Divergent RNA Viruses in Existing Metatranscriptome Data with Machine Learning

Most sampling of RNA viruses and phages has been directed toward a narrow range of hosts and environments. Several marine metagenomic studies have examined the RNA viral fraction in aquatic samples and found a number of picornaviruses and uncharacterized sequences. The lack of homology to known protein families has limited the discovery of new RNA viruses. We developed a computational method for identifying RNA viruses that relies on information in the codon transition probabilities of viral sequences to train a classifier. This approach does not rely on homology, but it has higher information content than other reference-free methods such as tetranucleotide frequency. Training and validation with RefSeq data gave true positive and true negative rates of 99.6% and 99.5% on the highly imbalanced validation sets (0.2% viruses) that, like the metatranscriptomes themselves, contain mostly non-viral sequences. To further test the method, a validation dataset of putative RNA virus genomes were identified in metatransciptomes by the presence of RNA dependent RNA polymerase, an essential gene for RNA viruses. The classifier successfully identified 99.4% of those contigs as viral. This approach is currently being extended to screen all metatranscriptome data sequenced at the DOE Joint Genome Institute, presently 4.5 Gb of assembled data from 504 public projects representing a wide range of marine, aquatic and terrestrial environments.