Answers are grounded in this paper and cite their sources.
Introduction Emerging and neglected viral diseases often circulate silently, remaining undetected by conventional surveillance systems, particularly when associated with unknown or genetically divergent agents. Metagenomic sequencing has become a powerful tool for the untargeted detection of viruses in clinical samples. However, a substantial fraction of metagenomic sequences remains taxonomically unclassified, commonly referred to as viral “dark matter”, owing to the extensive genetic diversity of viruses and limitations of current reference databases and reference-based taxonomic classification methods. Methods Here, we investigated the viral dark matter recovered from plasma metagenomes of pediatric patients with febrile neutropenia. Random Forest and eXtreme Gradient Boosting models trained on sequence-derived features from 34 human-associated viral families were applied to prioritize candidate viral contigs among the Kraken2-unclassified fraction . Results Of the 8,640 assembled contigs considered, 149 (1.72%) had at least one protein profile compatible with the training set, whereas the remaining 8,491 (98.28%) were excluded from the machine-learning analysis because they did not meet this requirement. Although both models achieved excellent performance on reference-genome datasets, application to clinical metagenomes highlighted the importance of independent biological validation. Among the 149 eligible contigs, most candidate contigs were assigned to the family Anelloviridae (n = 91, 61.1%), and subsequent BLASTx, protein-domain annotation, and phylogenetic analyses provided biological support for the majority of these predictions, including a lineage consistent with the recently proposed genus Samektorquevirus. Discussion Together, these findings demonstrate that supervised tree-based machine-learning approaches can help prioritize candidate viral sequences within the subset of metagenomic dark matter that contains recognizable protein domains represented in the training data, while highlighting the limitations of this approach for contigs lacking such features.
No takes yet. Share an insight, caveat, or question.
Campos et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: