ViroNEXT: A multilayer hybrid system for high-precision detection of known and divergent viruses in metagenomics

Technical Summary: ViroNEXT – A Multilayer Hybrid System for High-Precision Viral Detection

Problem Statement
Metagenomic next-generation sequencing (mNGS) offers a hypothesis-free approach for viral detection, essential for outbreak response, surveillance, and pathogen discovery. However, current implementation in routine diagnostics faces two primary challenges: (1) the inability of homology-based methods to detect highly divergent or previously uncharacterized viruses as evolutionary distance increases, and (2) the generation of substantial false-positive assignments due to spurious alignments and taxonomic ambiguities. These false positives complicate the interpretation of metagenomic findings, particularly in complex clinical and environmental samples dominated by host-derived nucleic acids. Existing pipelines often optimize for either known virus classification or exploratory discovery, lacking an integrated approach that balances sensitivity with stringent precision.

Methodology
The authors present ViroNEXT, a fully automated, end-to-end bioinformatics pipeline designed to integrate reference-based identification with machine-learning (ML)-assisted detection. The system operates on a multilayer hybrid architecture comprising two complementary branches:

1. Reference-Based Branch: This branch targets viruses represented in current reference databases (specifically the Clustered Reference Viral Database, C-RVDB). It employs a sequential, multilayer classification strategy to maximize precision:

   • Primary Classification: Nucleotide alignment using bwa-mem2.
   • Secondary Classification: Amino acid alignment using Kaiju against the C-RVDB-prot database.
   • Coverage Filtering: Stringent filtering based on non-adjacent regions, minimum region length, and total covered bases to eliminate localized spurious alignments.
   • Tertiary Classification: Consensus sequence generation followed by BLASTn classification, requiring species-level concordance with primary and secondary steps.
   • Final Checkpoint: Reporting only "high-quality read assignments" (mapping quality ≥ 60).

2. Machine-Learning Branch: Designed to capture highly divergent sequences lacking close reference homologs, this branch processes reads unclassified by the reference branch.

   • Assembly: de novo assembly of unclassified reads using SPAdes in rnaviral mode.
   • Prediction: Contigs are evaluated using VirHunter, a deep learning model (convolutional neural network) trained on generalized viral sequence patterns.
   • Validation: Candidate contigs undergo read-support thresholds (minimum 100 mapped reads) and protein-level homology screening via DIAMOND (blastx) against the clustered NCBI protein database.

Key Contributions

• Integrated Hybrid Workflow: ViroNEXT uniquely combines homology-based and ML-based approaches within a single, automated pipeline, allowing for the detection of both known and divergent viruses without manual intervention.
• Multilayer Filtering Architecture: By requiring candidates to pass sequential, orthogonal classification steps (nucleotide, amino acid, coverage, consensus, and ML prediction), the system significantly reduces false-positive rates while maintaining high recall.
• Accessibility: The system is available as open-source software for local execution (Snakemake-based) and via a publicly accessible web interface, removing barriers related to computational infrastructure and bioinformatics expertise.
• Interpretable Reporting: The pipeline generates interactive HTML reports with visualizations of abundance, genome coverage, and filtering metrics, facilitating clinical interpretation.

Results
ViroNEXT was evaluated across four distinct scenarios using synthetic, spike-in, and clinical datasets, comparing its performance against established pipelines (Genome Detective, One Codex, and CZID).

• Synthetic Datasets (Sequence Divergence): In viral coinfection scenarios with substitution rates up to 25%, the reference-based branch of ViroNEXT maintained high precision (zero false positives) but showed declining recall at high divergence. The ML branch compensated for this, detecting highly divergent sequences (up to 25% divergence) that the reference branch missed, achieving a superior F1 score compared to benchmarks.
• Environmental Virome: In a complex dataset containing 27 viral species, ViroNEXT detected all species with perfect precision (no false positives), outperforming CZID and One Codex, which generated multiple false-positive assignments.
• Diagnostic Escape: In scenarios simulating viruses with sequence similarities as low as 45% to known references, ViroNEXT maintained a recall of 1.0 by integrating the ML branch, whereas benchmark pipelines showed significant drops in recall (e.g., One Codex dropped to 0.4).
• Spike-in Experiments: Across WHO International Reference Panel and multi-virus spike-in datasets, ViroNEXT demonstrated the lowest number of false-positive assignments. Statistical analysis (negative-binomial GLMM) revealed that Genome Detective and One Codex produced ~7-fold more false positives, while CZID produced ~288-fold more than ViroNEXT.
• Novel Virus Detection: In a "leave-one-out" experiment where entire viral families were removed from both the reference database and ML training data, the ML branch successfully recovered genomes for 5 of 7 families (e.g., Circoviridae, Spinareoviridae), demonstrating the ability to detect viruses from previously unrepresented families.
• Clinical Samples: In datasets from patients with encephalitis, respiratory infections, and suspected arboviral infections, ViroNEXT achieved high recall (1.0) and precision (0.96). It reported fewer unsupported viral assignments than other pipelines, with 86% of its assignments supported by at least three independent pipelines, compared to only 14% for One Codex and CZID.

Significance and Claims
The authors claim that ViroNEXT addresses a critical barrier in metagenomic analysis: the trade-off between sensitivity and precision. By integrating ML-assisted detection with rigorous, multilayer homology-based filtering, ViroNEXT extends the effective detection range to highly divergent viruses while substantially reducing false-positive classifications. The system is presented as a robust, practical tool for pathogen discovery, epidemiological surveillance, biosafety testing, and clinical diagnostics. The authors emphasize that minimizing false positives is crucial not only for clinical accuracy but also for preventing unnecessary follow-up investigations in surveillance and biosafety contexts. While acknowledging that larger clinical studies are needed for full diagnostic validation, the paper positions ViroNEXT as a significant step toward reliable, automated, and interpretable viral metagenomic analysis.