Machine Learning for detection of viral sequences in human metagenomic datasets

Abstract

BackgroundDetection of highly divergent or yet unknown viruses from metagenomics sequencing datasets is a major bioinformatics challenge. When human samples are sequenced, a large proportion of assembled contigs are classified as “unknown”, as conventional methods find no similarity to known sequences. We wished to explore whether machine learning algorithms using Relative Synonymous Codon Usage frequency (RSCU) could improve the detection of viral sequences in metagenomic sequencing data.ResultsWe trained Random Forest and Artificial Neural Network using metagenomic sequences taxonomically classified into virus and non-virus classes. The algorithms achieved accuracies well beyond chance level, with area under ROC curve 0.79. Two codons (TCG and CGC) were found to have a particularly strong discriminative capacity.ConclusionRSCU-based machine learning techniques applied to metagenomic sequencing data can help identify a large number of putative viral sequences and provide an addition to conventional methods for taxonomic classification.Electronic supplementary materialThe online version of this article (10.1186/s12859-018-2340-x) contains supplementary material, which is available to authorized users.

Short k-mer abundance profiles yield robust machine learning features and accurate classifiers for RNA viruses

“…The performance of the classifiers at this point far exceeded the outcomes of other classifiers built for the same task in previous studies [11,16,17,21]. Given that the focus of these previous classifier were genomic DNA data, in order to test whether this performance amplification was entirely attributable to the greater homogeneity RNA sequences and their shorter lengths compared to genomic DNA sequences, we studied the performance of the classifier on…”

ViraMiner: Deep Learning on Raw DNA Sequences for Identifying Viral Genomes in Human Samples

“…In such a case, while losing information 289 about the maximal activation (best match), we gain information about frequency -the 290 average cannot be high if only a few good matches were found. In previous work the 291 authors of DVF and the authors of the current article have shown that methods based 292 on pattern frequency (k-mer counts, relative synonymous codon usage) are effective in 293 separating viral samples from non-viral ones [24,28]. Using convolution + average is a 294 natural extension to these pattern counting-based models.…”

“…To compare against the performance of such 112 PLOS 5/16 methods, we also extracted k-mers from the investigated dataset and trained Random 113 Forest (RF) classifiers on the extracted values, while keeping the same data partitioning 114 as above. RF is a competitive machine learning algorithm for non-linearly separable 115 classes and it has already been used on this type of datasets [23,28]. The best test 116 performance with RF models was achieved with 6-mers and it produced test AUROC 117 0.875 (Fig 4).…”

“…As a consequence, many sequences generated from NGS technologies are 24 classified as "unknown" by BLAST [5,18]. Another commonly used algorithm for viral 25 discovery in metagenomic samples is HMMER3 [20], which applies profile Hidden 26 Markov Models (pHMM) in relation to the vFams [21] database of viral family of 27 proteins, that have been built by multiple sequence alignments from all viral proteins in 28 RefSeq. With this method, sequences are compared to entire viral families which 29 enables it to be more effective in detecting distant homologs [22].…”

PACIFIC: A lightweight deep-learning classifier of SARS-CoV-2 and co-infecting RNA viruses

“…In contrast, machine learning based tools have the advantage of extracting required features and encapsulating the necessary information for sequence classification in a computationally efficient model. This approach has been successfully used in the past for sequence classification problems (20). For example, DeepMicrobes (21)…”