Sensitivity is cheap; specificity is not — hybrid viral classifiers on short, low-abundance contigs

Question: How do hybrid machine learning and alignment-based viral contig classifiers perform on short, low-abundance contigs from environmental metagenomes, and what strategies reduce false-positive host contamination?

Answer

Taken together, these papers support a deflationary answer: hybrid detectors buy recall on short, low-abundance contigs, but they do not buy specificity, and nothing in the retrieved literature demonstrates that classifier architecture is the lever that controls host contamination. The best-performing configurations reported are two-tool combinations of a learning-based detector (DeepVirFinder) with a second, differently-biased detector, followed by an explicit taxonomic rejection step [6]; feature-fusion hybrids of k-mer statistics and foundation-model embeddings report accuracy gains on short contigs and low false-positive rates in zero-shot environmental tests [11], but the false positives that matter — host-derived sequence, host genes shared with viruses, and provirus flanks — are segmentation and provenance problems rather than scoring problems [3], and neither standard genome-quality metrics nor classifier confidence detect them [17]. The strategies that demonstrably reduce host contamination are external to the classifier: contig-level taxonomic pre-filtering, contamination-aware quality control with base-resolution boundary calls, orthogonal evidence such as DNA-read recruitment, and upstream enrichment or depletion that changes the ratio of viral to host signal before assembly.

Recall on short, low-abundance contigs is the binding constraint, and no single detector holds it

Benchmarking eleven prediction tools on artificial low-abundance metagenomes — built because prior benchmarks used virus-enriched, high-abundance data — found that DeepVirFinder and DIAMOND gave the highest recall, at median 67% and 58% respectively, across increasing in silico mutation loads [6]. The reason is informational: earlier methods struggle on short contigs specifically because the fragments carry too little signal and training sets were small and close-set [11]. The same limitation holds for read-level classification, where exact k-mer matching and alignment lose discriminative power on 20–100 bp fragments, and where compression-based models gain most precisely on ultra-short input [7]. Even end-to-end pipelines that report 21% and 13% accuracy gains over four competing pipelines on mock and experimental data are reporting a pipeline-level aggregate, not performance in the short-contig, low-abundance corner where the failure occurs [2].

Fusion of sequence-composition and learned representations is the mechanism that rescues short fragments

The clearest direct evidence for hybridisation is VIRALpre, which fuses genomic-foundation-model embeddings with k-mer features and reports a 4% accuracy gain over prior methods, roughly a 10% F1 improvement under a BLASTn similarity cut-off designed to mimic contigs with no close training relative, and near-complete recovery with very low false-positive rate in zero-shot cross-dataset tests sampled from natural environments [11]. Fusing heterogeneous representations appears independently: DeepVirus combines protein-level foundation-model embeddings with genome-aware representations for hierarchical and open-set viral classification [10], and fine-tuning protein language models on viral sequences measurably improves representation quality for the under-represented viral proteome [9]. At the tool-combination level, the best-performing formulations on mutated RefSeq contigs were all two-tool hybrids — DeepVirFinder combined with DeepMicroclass, ViraLM, geNomad, or DIAMOND [6].

Combining differently-biased tools outperforms tuning any single model

The complementary-combination approach increased the number of viral contigs with functionally annotated CDSs and revealed diversity missed by gold-standard methods, while limiting false positives [6]. The same pattern is reported for individual tools: Cenote-Taker 3 performs comparably to geNomad on virus discovery and the two produce complementary results [5]; tyPPing surpasses alternatives on sensitivity for phage-plasmids yet is explicitly positioned as a component to be combined with other tools [18]; and DeepVir pairs classical homology alignment with profile-HMM mining of RdRp rather than choosing between them [12].

Host contamination has three separable mechanisms, so no score threshold can suppress it

Learning-based detectors such as VirSorter2 and geNomad show varying false-positive rates attributable to sequencing and assembly noise, genes shared between viruses and their hosts, and provirus formation inside host genomes [3]. These are not one failure mode. A contamination-aware tool that identifies contamination regions with foundation models trained on viral versus cellular genomes and estimates completeness by protein-organisation alignment identifies 38% more contamination than CheckV at a median absolute error of 3% [3]. Critically, host-proximal errors survive standard quality control: incorporating substantial amounts of tetranucleotide-matched non-host phage sequence into high-quality MAGs does not substantially alter completeness and contamination estimates, and CRISPR-associated regions diverge compositionally from their host genomes, raising misbinning risk — motivating scaffold-level validation that integrates structural linkage and taxonomic consistency rather than MAG-level metrics [17]. Large-scale protein-family resources handle the same problem by iteratively removing genomic sequences likely of cellular origin before model construction [8].

Pre-filtering and orthogonal evidence reduce false positives more reliably than classifier confidence does

The combination that achieved the best sensitivity on real human airway and air metagenomes paired two-tool detection with Kraken2X filtering to remove contigs classified as Eukaryota, Bacteria, or Archaea [6]. A simple score pre-filter at threshold 0.7 improved precision by ten percentage points with negligible recall loss in a compression-based classifier [7]. Orthogonal evidence functions similarly: structure consistency and low DNA-read mapping to the corresponding contig supported genuine RNA viral origin and reduced false positives in a two-stage structure-guided framework applied to over 10,000 metatranscriptomes [4], and dedicated prophage-extraction modules exist to separate integrated provirus from host backbone [5]. Host assignment is a parallel contamination channel: the choice of spacer-matching tool drastically changes CRISPR spacer–protospacer detection and hence virus–host inference [20].

Upstream sample processing sets the false-positive budget before any classifier runs

In paired DNA-metagenome and VLP-enriched libraries from the same nasopharyngeal samples, contig overlap between workflows was under 1%, protocol choice reshaped viral recovery (PERMANOVA R² = 0.448), DNA metagenomes favoured assembly contiguity while VLP enrichment maximised viral detection [15]. Systematic protocol comparison for faecal viromes likewise exposes explicit trade-offs between viral genome recovery and contamination removal, with an optimised protocol required to minimise non-viral sequence [13], and viral richness, taxonomy, and integration status shift with environment and with VLP enrichment versus bulk DNA extraction [1]. Targeted depletion is the most quantitative lever: probe-based removal of rRNAs and abundant tobamoviruses cut rRNA reads by 90%, reduced tobamoviruses from up to 25% of viral reads to under 2%, produced an order-of-magnitude gain in viral read signal, and increased human and animal pathogen reads 42-fold, rendering previously undetectable pathogens detectable [16]. Reproducible host and rRNA decontamination of reads and assemblies is available as a defined pipeline step [14].

Conflicts and gaps

The central gap is that no retrieved paper measures host contamination as an endpoint for short, low-abundance environmental contigs. The most relevant benchmark scores recall and F1 on RefSeq genomes fragmented and mutated in silico, with false-positive control introduced as a post hoc filtering step rather than quantified per length bin [6]. The field therefore disagrees implicitly about what a false positive is: VIRALpre claims very low false-positive rates in zero-shot natural-environment tests [11], while ViralQC documents variable false-positive rates in the same class of learning-based tools and attributes them to host-shared genes and proviruses [3] — different definitions, not necessarily different observations. There is also disagreement about where the short-contig gain comes from: fusion of embeddings with k-mers [11], compression-based finite-context models [7], or simply tool combination plus taxonomic filtering [6]; no ablation isolates these. A further tension: binners systematically underperform for prophages and fail on contigs shared between genomes, and CheckM2 systematically overestimates completeness while underestimating contamination (partially ameliorated by GUNC) [19] — so the ground truth used to validate provirus handling is itself biased in exactly the region that generates host-contaminated contigs.

What would change the answer

A length-binned, abundance-tiered benchmark would settle the primary question: shear viral genomes to 300, 500, 1000, 2000, and 5000 bp, spike them into real environmental DNA at relative abundances spanning roughly 10⁻⁶ to 10⁻⁴, and score precision at fixed recall in each bin. The negative set must be adversarial rather than random — prophage-containing host contigs, host genomic islands, gene-transfer agents, eukaryotic repeat-rich sequence — because random host contigs are too easy [17]. Base-resolution contamination labels, of the kind ViralQC produces, would allow boundary error to be reported in base pairs and would let contamination-aware segmentation be evaluated as a detector in its own right rather than as a post-processing filter [3]. A three-way feature ablation — k-mer only, embedding only, fused — as a continuous function of contig length would identify where fusion actually pays and where it merely correlates with contig length. Finally, prospective precision estimation is needed: apply the best pipeline to a real environmental sample, stratify novel calls by score, and verify a random sample of each stratum by read recruitment or targeted PCR, since every retrieved performance figure derives either from simulated data or from mock communities that lack the diversity of the target environment [2].

References

[1] Birkeland, E., Istvan, P., Avershina, E., Rognes, T., Rounge, T. B.. "VirMake: a flexible and user-friendly pipeline for viral taxonomic and functional characterisation from shotgun metagenomic sequencing data". bioRxiv. 2025-02-08T00:00:00.000Z. URL: https://doi.org/10.1101/2025.02.07.637044

[2] SHEKARRIZ, E., VIJENDRAN, E., Ho, J. W. K.. "vOMIX-MEGA: An ultra-fast end-to-end pipeline for terabyte-scale viral metagenomics analysis.". bioRxiv. 2026-07-31T00:00:00.000Z. URL: https://doi.org/10.64898/2026.07.28.741255

[3] Cheng Peng, Jiayu Shang, Jiaojiao Guan, Yanni Sun. "ViralQC: A Tool for Assessing Completeness and Contamination of Predicted Viral Contigs". arxiv. 2025-04-09T00:00:00.000Z. URL: https://arxiv.org/abs/2504.0579

[4] Luo, G., Zang, Z., Yuan, L., Zhou, J., Dong, A., Huang, Y., Li, S. Z., Ju, F.. "Expanding the RNA Virus Universe by Scalable Structure-Guided Discovery". bioRxiv. 2025-11-27T00:00:00.000Z. URL: https://doi.org/10.1101/2025.11.24.690314

[5] Tisza, M. J., Varsani, A., Petrosino, J. F., Cregeen, S. J. J.. "Cenote-Taker 3 for Fast and Accurate Virus Discovery and Annotation of the Virome". bioRxiv. 2026-03-25T00:00:00.000Z. URL: https://doi.org/10.1101/2025.08.20.671380

[6] Alami, S., DESTRAS, G., Rimeur, M., Navratil, V., Josset, L., Anani, H.. "Benchmarking and integration of viral prediction tools to uncover viral dark matter across divergent low abundance metagenomes". bioRxiv. 2026-10-19T00:00:00.000Z. URL: https://doi.org/10.64898/2026.10.14.759168

[7] Marques, L. L., Pinho, A. J., Pratas, D.. "Metagenomic Classification of Ancient Viruses". bioRxiv. 2025-11-07T00:00:00.000Z. URL: https://doi.org/10.1101/2025.11.07.687203

[8] Bucio, R. E. P., Enault, F., Galiez, C.. "EnVhogDB: an extended view of the viral protein families on Earth through a vast collection of HMM profiles". bioRxiv. 2025-09-01T00:00:00.000Z. URL: https://doi.org/10.1101/2024.06.25.600602

[9] Sawhney, R., Ferrell, B. D., Dejean, T., Schreiber, Z., Harrigan, W., Polson, S. W., Wommack, K. E., Belcaid, M.. "Fine-Tuning Protein Language Models Unlocks the Potential of Underrepresented Viral Proteomes". bioRxiv. 2025-06-11T00:00:00.000Z. URL: https://doi.org/10.1101/2025.04.17.649224

[10] Cao, C., He, L., Li, C., Jiang, Y., Tang, C., Li, Y., He, Y., Min, Y., Liu, H., Qin, T., Liu, T.-Y.. "Illuminating the Virosphere's Dark Matter using Hierarchical Deep Learning". bioRxiv. 2025-09-25T00:00:00.000Z. URL: https://doi.org/10.1101/2025.09.22.677955

[11] Wang, Z., Yu, Q., Li, Y.. "VIRALpre: Genomic Foundation Model Embedding Fused withK-mer Feature for Virus Identification". bioRxiv. 2024-11-15T00:00:00.000Z. URL: https://doi.org/10.1101/2024.11.12.623150

[12] Cosentino, M., FERNANDEZ NUNEZ, N., Soares, M. A., Ayouba, A., Santos, A., D'arc, M.. "DeepVir: A reproducible workflow for large-scale viral dark matter discovery". bioRxiv. 2026-09-18T00:00:00.000Z. URL: https://doi.org/10.64898/2026.09.15.751837

[13] De Chiara, L., Doughty, R., Estevez-Gomez, N., Gallego-Garcia, P., Alvarino, P., Diez-Martin, A., Davila Pinon, P., Treangen, T. J., Cubiella, J., Posada, D.. "A Comparison of Methods for the Optimal Recovery of the Human Fecal Virome". medRxiv. 2025-05-13T00:00:00.000Z. URL: https://doi.org/10.1101/2025.05.12.25327428

[14] Lataretu, M., Krautwurst, S., Huska, M. R., Marquet, M., Viehweger, A., Braun, S. D., Brandt, C., Hölzer, M.. "Targeted decontamination of sequencing data with CLEAN". bioRxiv. 2025-01-24T00:00:00.000Z. URL: https://doi.org/10.1101/2023.08.05.552089

[15] Mencia-Ares, O., Deneke, C., Martinez-Martinez, S., Malorny, B., Gutierrez-Martin, C. B., Gruetzke, J.. "Paired viromics resolves the modular ecological architecture of the swine nasopharyngeal phageome". bioRxiv. 2026-05-29T00:00:00.000Z. URL: https://doi.org/10.64898/2026.05.28.728360

[16] Bautista, M. A., Bhardwaj, T., Barbera, D. J., Rulli, S. J., Zorz, J., McCalder, J., Papparis, C., Pang, X., Lee, B. E., Qiu, J., Clark, R. G., Acosta, N., O' Grady, C., Parkins, M. D., Hubert, C. R.. "Targeted Depletion of rRNAs and Tobamoviruses Enhances Viral Discovery in Wastewater Metatranscriptomes". bioRxiv. 2026-10-10T00:00:00.000Z. URL: https://doi.org/10.64898/2026.10.08.757481

[17] Yuan, L., Anantharaman, K., Ma, B., Camargo, A. P., Chen, L.. "Revisiting phage-host inference from metagenome-assembled genomes reveals the need for scaffold-level validation". bioRxiv. 2026-04-23T00:00:00.000Z. URL: https://doi.org/10.64898/2026.04.22.720271

[18] Ilchenko, K., Bonnin, R., Rocha, E. P. C., Pfeifer, E.. "Efficient detection and typing of phage-plasmids". bioRxiv. 2025-08-31T00:00:00.000Z. URL: https://doi.org/10.1101/2025.08.29.673033

[19] Coleman, I., Ma, J., Qian, G., Jiang, Y., Brown Kav, A., Korem, T.. "End-to-end evaluation of pipelines for metagenome-assembled genomes reveals hidden performance gaps". bioRxiv. 2026-04-09T00:00:00.000Z. URL: https://doi.org/10.64898/2026.04.06.712906

[20] Neri, U., Camargo, A. P., Bushnell, B., Roux, S.. "Tool Choice drastically Impacts CRISPR Spacer-Protospacer Detection". bioRxiv. 2025-05-10T00:00:00.000Z. URL: https://doi.org/10.1101/2025.05.06.652306

How this survey was made

This survey answers one question: “How do hybrid machine learning and alignment-based viral contig classifiers perform on short, low-abundance contigs from environmental metagenomes, and what strategies reduce false-positive host contamination?”

We keep our own collection of recent preprints and refresh it every day from bioRxiv, medRxiv, arXiv, PsyArXiv, SocArXiv and ChemRxiv. For each survey, we search that collection for recent work and keep the strongest matches: papers that use the same terms as the question, and papers on the same topic in different words.

The text was written from the titles and abstracts of those papers; the full texts were not read. Statements that draw on a particular paper are marked with a number in brackets. The reference list just above is the 20 papers whose work this text draws on, chosen from the 164 papers those searches returned; the rest of the shortlist is not printed. It is not a complete review of everything published on the question.

Treat this as a starting point and a reading list, not a peer-reviewed review. The searches behind this survey ran on October 22, 2026 and covered work from the previous two years.