The Promises and Pitfalls of Machine Learning for Detecting Viruses in Aquatic Metagenomes.

Alise J Ponsero,Bonnie L Hurwitz

doi:10.3389/fmicb.2019.00806

Abstract

Tools allowing for the identification of viral sequences in host-associated and environmental metagenomes allows for a better understanding of the genetics and ecology of viruses and their hosts. Recently, new approaches using machine learning methods to distinguish viral from bacterial signal using k-mer sequence signatures were published for identifying viral contigs in metagenomes. The promise of these content-based approaches is the ability to discover new viruses, with no or few known relatives. In this perspective paper, we examine the use of the content-based machine learning tool VirFinder for the identification of viral sequences in aquatic metagenomes and explore the possibility of using ecosystem-focused models targeted to marine metagenomes. We discuss the impact of the training set composition on the tool performance and the current limitation for the retrieval of low abundance viral sequences in metagenomes. We identify potential biases that could arise from machine learning approaches for viral hunting in real-world datasets and suggest possible avenues to overcome them.

Highlights

Viruses infect host cells from all domains of life and are highly adapted to their host genetics and their environmental niches (Hurwitz et al, 2014; Brum et al, 2015)
Using a k-Disagreeing Neighbors algorithm, Soueidan et al (2015) show that, for a k-mer size of 3 bp, the high-level classification of viral sequences mixed with non-viral sequences is a hard task, whereas low-level classification is easier. These results suggest that machine learning models trained to classify viral sequences against cellular sequences may have a hard time generalizing to unknown viral families
While we do not claim that all ecosystem-focused models would perform better in the detection of rare events, this experiment shows how valuable high precision models can be in the case of very imbalanced datasets, with a significant improvement in the precision (Figure 2A) and Area under the precision-recall curve (AUPRC) (Figure 2B) of the Tara-trained model compared to the VirFinder “phage-prok.”

Summary

Introduction

Viruses infect host cells from all domains of life and are highly adapted to their host genetics and their environmental niches (Hurwitz et al, 2014; Brum et al, 2015). We review the potential bias and pitfalls of composition-based machine learning approaches such as VirFinder for the detection of viral sequences in aquatic ecosystems.

Results

Conclusion