DeepVirus

Most sequences from metagenomic surveys have no close relative in any reference database. This unclassified fraction — the virosphere's "dark matter" — is invisible to homology search and marker-gene profiling, which recognize only what already resembles something known. DeepVirus, developed at Beijing Zhongguancun Academy with the Zhongguancun Institute of Artificial Intelligence and academic collaborators, treats a genome not as a nucleotide string but as an ordered sequence of protein-coding genes, modeled with a transformer.

The framework answers three questions in a single pass: whether an assembly is viral, which known lineage it belongs to across a seven-level taxonomy, and whether no known lineage fits at all. That third capability — open-set recognition, implemented as a statistical rejection test on the classifier's outputs — is what separates DeepVirus from closed-set classifiers forced to assign every input to a training class.

It joins a growing set of learned alternatives to alignment-based virus identification, including k-mer-based DeepVirFinder, the DNA-level transformer BERTax, the hybrid tool geNomad, and the multimodal viral foundation model LucaVirus. Its bet is that genome-wide protein repertoires — functional motifs, more conserved than primary sequence, read in genomic order — carry enough signal to recognize viruses with almost no sequence identity to references.

#Key Features

• Genome as a sequence of proteins: Each protein is embedded by a pretrained protein language model, and the embeddings are read in genome order by a second transformer, so both function and gene organization inform the prediction.
• Taxonomy-consistent hierarchical head: A recurrent classification head predicts every taxonomic level jointly, with a violation loss penalizing cases where a child class outweighs its parent.
• Open-set recognition by hypothesis testing: Per-class distributions of logit-transformed probabilities are fitted on training data; a genome rejected by a one-sided z-test against every known class is flagged as a candidate novel lineage.
• Iterative lineage expansion: Rejected genomes can be given a new label and folded back into training, enabling a second discovery round that recovers lineages missed on the first pass.
• Imbalance-aware training: Loss re-weighting by class frequency counters a RefSeq taxonomy ranging from thousands of genomes in one kingdom to six in another.

#Technical Details

The protein encoder is a transformer pretrained on UniRef50 (roughly 30 million sequences) with masked language modeling plus a pairwise objective that predicts pairs of masked residues to capture co-evolutionary signal. The genome-level model is BERT-style with 8 transformer layers and a hidden size of 1,280, pooling a genome CLS token that feeds the hierarchical head. Training ran 300 epochs on 8 GPUs with Adam at a learning rate of 1e-5 and batch size 64. Data came from NCBI RefSeq split chronologically at 2022-12-31 to prevent leakage: 11,480 viral genomes for training and validation, 3,340 held out, plus archaeal, bacterial, fungal, protozoan, plant, invertebrate, and vertebrate genomes as negatives. Genomes average about 45 proteins, taken from RefSeq annotation or predicted with Prodigal.

On a held-out set of 28,702 assemblies (14,820 viral, 13,882 non-viral) comprising 100 million proteins, DeepVirus reached 0.999 accuracy and 0.997 MCC on binary virus identification. Restricted to viral genomes, it exceeded 0.97 accuracy and 0.93 MCC at realm, kingdom, and phylum, and outperformed MMseqs2, BERTax, DeepVirFinder, and geNomad across six ranks from realm to family. On geNomad's own length-stratified benchmark, it improved MCC on contigs under 3 kb from 0.944 to 0.960. Leave-one-phylum-out experiments, withholding an entire phylum from training, recovered roughly half of the tested phyla.

#Applications

On an ocean metatranscriptome of 1,943,085 assemblies and 10 million proteins, DeepVirus called 14.2% of genomes viral at high confidence, 92% matching the original study's annotations, with nearly 12% falling outside all known phyla. On a larger compendium of 8,256,141 assemblies and over 900 million proteins, it called 10.2% viral with 72.9% overlap and 7.6% candidate novel lineages. Follow-up recovered RNA-dependent RNA polymerases that retain the hallmark GDD catalytic motif, form clades separate from reference families, and adopt the canonical right-hand polymerase fold in AlphaFold 3 models while diverging in the fingers and thumb domains. It targets virologists, metagenomics groups mining environmental and host-associated sequence archives, and biosurveillance efforts that must triage assemblies with no database match.

#Impact

DeepVirus is a September 2025 preprint that has not been peer-reviewed; no code, weights, or hosted inference have been released, so results cannot be reproduced or the model reused. Its evidence is entirely in silico, and the authors are explicit that isolation, host identification, and functional characterization are needed before any candidate lineage enters formal viral taxonomy. Within those bounds the methodological argument is concrete: combining protein foundation model embeddings with genome-order context and an explicit rejection mechanism improves on marker- and homology-based tools on short, fragmentary contigs — the regime metagenomic assembly produces — while naming what it cannot classify instead of forcing a wrong label.