Rapid identification of human‐infecting viruses

Abstract

Viruses have caused much mortality and morbidity to humans and pose a serious threat to global public health. The virome with the potential of human infection is still far from complete. Novel viruses have been discovered at an unprecedented pace as the rapid development of viral metagenomics. However, there is still a lack of methodology for rapidly identifying novel viruses with the potential of human infection. This study built several machine learning models to discriminate human‐infecting viruses from other viruses based on the frequency of k‐mers in the viral genomic sequences. The k‐nearest neighbor (KNN) model can predict the human‐infecting viruses with an accuracy of over 90%. The performance of this KNN model built on the short contigs (≥1 kb) is comparable to those built on the viral genomes. We used a reported human blood virome to further validate this KNN model with an accuracy of over 80% based on very short raw reads (150 bp). Our work demonstrates a conceptual and generic protocol for the discovery of novel human‐infecting viruses in viral metagenomics studies.

Interpretable detection of novel human viruses from genome sequencing data

“…Table 1 presents the results of a benchmark using the “All” test set. Low performance of the k -NN classifier (Zhang et al, 2019) is caused by frequent conflicting predictions for each read in a read pair. In a single-read setting it achieves 75.5% accuracy, while our best model achieves 87.8% (Table S2).…”

“…We benchmarked our models against the human blood virome dataset used by Zhang et al (2019). Our models outperform their k -NN classifier.…”

“…Compared to the previous state-of-the-art in viral host prediction directly from next-generation sequencing reads (Zhang et al, 2019), our models drastically reduce the error rates. This holds also for novel viruses not present in the training set.…”

“…Zhang et al (2019) designed several classifiers explicitly predicting whether a new virus can potentially infect humans. Their best model, a k -NN classifier, uses k -mer frequencies as features representing the query sequence and can yield predictions for sequences as short as 500 base pairs (bp).…”

“…In this paper, we first improve the performance of read-based predictions of the viral host (human or non-human) from next-generation sequencing reads. We show that reverse-complement (RC) neural networks (Bartoszewicz et al, 2019) significantly outperform both the previous state-of-the-art (Zhang et al, 2019) and the traditional, alignment-based algorithm – BLAST (Altschul et al, 1990; Camacho et al, 2009), which constitutes a gold standard in homology-based bioinformatics analyses. We show that defining the negative (non-human) class is non-trivial and compare different ways of constructing the training set.…”