Short k-mer Abundance Profiles Yield Robust Machine Learning Features and Accurate Classifiers for RNA Viruses
High throughout sequencing technologies have greatly enabled the study of genomics, transcriptomics and metagenomics. Automated annotation and classification of the vast amounts of generated sequence data has become paramount for facilitating biological sciences. Genomes of viruses can be radically different from all life, both in terms of molecular structure and primary sequence. Alignment-based and profile-based searches are commonly employed for characterization of assembled viral contigs from high-throughput sequencing data. Recent attempts have highlighted the use of machine learning models for the task but these models rely entirely on DNA genomes and owing to the intrinsic genomic complexity of viruses, RNA viruses have gone completely overlooked. Here, we present a novel short k-mer based sequence scoring method that generates robust sequence information for training machine learning classifiers. We trained 18 classifiers for the task of distinguishing viral RNA from human transcripts. We challenged our models with very stringent testing protocols across different species and evaluated performance against BLASTn, BLASTx and HMMER3 searches. For clean sequence data retrieved from curated databases, our models display near perfect accuracy, outperforming all similar attempts previously reported. On de-novo assemblies of raw RNA-Seq data from cells subjected to Ebola virus, the area under the ROC curve varied from 0.6 to 0.86 depending on the software used for assembly. Our classifier was able to properly classify the majority of the false hits generated by BLAST and HMMER3 searches on the same data. The outstanding performance metrics of our model lays the groundwork for robust machine learning methods for the automated annotation of sequence data. Author SummaryIn this age of high-throughput sequencing, proper classification of copious amounts of sequence data remains to be a daunting challenge. Presently, sequence alignment methods are immediately assigned to the task. Owing to the selection forces of nature, there is considerable homology even between the sequences of different species which draws ambiguity to the results of alignment-based searches. Machine Learning methods are becoming more reliable for characterizing sequence data, but virus genomes are more variable than all forms of life and viruses with RNA-based genomes have gone overlooked in previous machine learning attempts. We designed a novel short k-mer based scoring criteria whereby a large number of highly robust numerical feature sets can be derived from sequence data. These features were able to accurately distinguish virus RNA from human transcripts with performance scores better than all previous reports. Our models were able to generalize well to distant species of viruses and mouse transcripts. The model correctly classifies the majority of false hits generated by current standard alignment tools. These findings strongly imply that this k-mer score based computational pipeline forges a highly informative, rich set of numerical machine learning features and similar pipelines can greatly advance the field of computational biology.