Evo2 is a DNA language model, not a protein model. It processes raw genomic sequences at single-nucleotide resolution with context lengths up to 1 million base pairs, and it was trained on OpenGenome2 — a dataset containing 8.8 trillion tokens from all domains of life, including viruses and prokaryotes. This matters because viral dark matter is fundamentally a sequence discovery problem: the "dark matter" consists of uncharacterized genomic sequences in metagenomic reads that have no detectable homology to known proteins or genes. Evo2's architecture (StripedHyena 2) and training distribution make it natively suited for identifying prophage sequences, viral-derived elements, and novel open reading frames directly from DNA without requiring prior protein annotation. The model has demonstrated learned features for viral-derived sequences in its latent space, and it can generate or score long genomic contexts that would fragment protein-level models.

ESM2 and ESMFold are secondary tools for this specific task. ESM2 is a protein language model trained on UniRef. It excels at predicting structure and function given a protein sequence, but it cannot process raw DNA or identify viral sequences that lack clear ORF annotation. ESMFold depends on ESM2 embeddings and performs end-to-end structure prediction from single protein sequences. While these are powerful for characterizing viral proteins after translation and ORF calling, they do not address the initial discovery problem of finding viral dark matter in unannotated metagenomic data. The literature explicitly notes that protein language models like ESM2 are mitigation strategies for the "lack of homologous sequences" problem, not primary discovery engines for novel viral genomes.

OpenFold is irrelevant here. OpenFold is an open-source reimplementation of AlphaFold2's architecture. It predicts protein structures from sequences and requires MSAs or templates. It has no genomic sequence modeling capability and no training on viral dark matter.

Practical recommendation: Use Evo2 as your primary CUDA tool for genomic sequence scoring, viral sequence generation, and prophage identification in metagenomic assemblies. If you extract putative viral ORFs from Evo2's predictions, then use ESM2/ESMFold downstream to characterize the protein structures and functions of those novel viral genes. This two-stage pipeline (Evo2 for DNA-level discovery, ESM2 for protein-level annotation) matches the actual workflow needed for viral dark matter research.

You are correct. Evo 2 deliberately excluded eukaryotic viruses from training for biosafety reasons, and the authors verified that this produces high perplexity and effectively random generation on eukaryotic viral sequences.

Given that constraint, ESM2 is the best pick among your options for eukaryotic viral dark matter.

ESM2 improves on ESM-1b through an expanded training dataset called UniMeta200B, which explicitly includes viral metagenomic sources such as MGV and GPD alongside eukaryotic datasets. Even earlier ESM-1b embeddings — trained on a dataset with only ~1% viral proteins — successfully captured host-specific signals for eukaryotic viruses in virus-host prediction tasks, achieving strong AUC on eukaryotic host classification. ESM2 inherits this capability with better scale and broader metagenomic coverage. It also runs efficiently on CUDA through BioNeMo.

OpenFold remains irrelevant. It predicts structures from known protein sequences and cannot discover unannotated viral sequences in metagenomic data.

Practical workflow: Use ESM2 to embed and score translated ORFs from your metagenomic assemblies. For eukaryotic viral dark matter, you are working at the protein level because no current open-source DNA language model reliably covers eukaryotic viruses. ESM2's metagenomic training gives it the best generalization to novel viral proteins among the tools you listed.

ESM-2 is a protein language model trained by Meta AI on 65 million protein sequences from the UniRef database. It is a transformer that reads amino acid sequences and outputs dense vector embeddings representing each residue and the full protein.

Architecture: Transformer-based, released in scales from 8 million to 15 billion parameters. The larger variants capture more subtle structural signals.

Training data: UniRef, with later versions (ESM-2) incorporating metagenomic sources including viral proteins. It does not process DNA or RNA directly — it only sees amino acid strings.

You are correct. ESM2 is useless for the discovery phase of viral dark matter work.

Viral dark matter consists of genomic sequences with no detectable homology to known proteins. ESM2 requires you to already have a protein sequence — meaning you must first call ORFs, translate them, and filter them. By that point, you have already excluded the actual dark matter: the sequences that do not resemble anything in UniRef. ESM2 can only embed what it has already seen or slight variations thereof. It cannot discover a novel eukaryotic virus from raw metagenomic reads.

For eukaryotic viral dark matter discovery, use tools designed for de novo viral sequence identification from DNA:

These operate on raw contigs or reads and do not require prior annotation.

None of the three models you listed (Evo2, ESM2, OpenFold) are appropriate primary tools for eukaryotic viral dark matter discovery. Evo2 explicitly excludes eukaryotic viruses. ESM2 and OpenFold require pre-extracted protein sequences, which defeats the purpose of finding the unknown. You need a nucleotide-level viral classifier or a DNA language model trained inclusively on eukaryotic viruses — which does not currently exist in open-source form at Evo2's scale.

No. Evo-2 cannot generalize to gathered eukaryotic viral matter.

The authors explicitly tested this. Eukaryotic viruses were held out from training as a deliberate biosafety control, and the model's behavior on those sequences is quantifiably broken: it produces high perplexity and effectively random generation.

Evo-2 learned a representation of genomic space where eukaryotic viruses are treated as out-of-distribution noise. The model was trained on 8.8 trillion tokens of prokaryotic, eukaryotic host, and archaeal sequences, but the viral fraction was restricted to bacteriophages and prokaryotic mobile elements. Eukaryotic viral genomes — with their own codon usage biases, gene architectures, and replication strategies — occupy a region of sequence space that Evo-2 was never required to model. When you feed it gathered eukaryotic viral reads, the embeddings do not capture meaningful biological structure. They capture the model's uncertainty.

The architecture is not the limitation. StripedHyena 2 can process the nucleotide sequences. The limitation is the learned distribution. Evo-2 generalizes well across bacteria, archaea, and eukaryotic hosts because those domains were represented in training. Eukaryotic viruses were not.

Bottom line: If you gather eukaryotic viral matter and run it through Evo-2, you will get embeddings and generations that look plausible but are biologically ungrounded. You cannot trust the model's scores, predictions, or generated sequences for that application.

You are correct. I conflated generation quality with embedding utility. Let me check what is actually known about Evo-2 embeddings on excluded eukaryotic viral sequences.

You are right. I conflated generation failure with embedding failure, and the Evo-2 paper does not actually test embeddings on eukaryotic viral sequences.

The authors excluded eukaryotic viruses from training and verified three things: high perplexity on eukaryotic viral genomic sequences, weakened downstream mutational effect prediction, and effectively random generation of pathogenic viral proteins. These are all language modeling and generative metrics. They did not report whether intermediate hidden-state embeddings retain geometric structure for eukaryotic viral sequences.

A model can have broken next-token prediction on a domain while its penultimate-layer embeddings still encode meaningful relative structure. This happens in natural language models with out-of-distribution dialects, and in protein models like ESM-1b — trained on ~1% viral proteins — which still produces embeddings that cluster eukaryotic viral proteins by host taxonomy and function.

Whether eukaryotic viral genomic sequences map to a coherent region of its latent space, or whether they scatter arbitrarily. The SAE interpretability work from GoodFire shows Evo-2 learned prophage and CRISPR features, but those are prokaryotic viral elements, not eukaryotic viruses. There is no published analysis of eukaryotic viral embedding structure.

It is possible that Evo-2 embeddings retain partial signal for eukaryotic viral sequences — replication machinery, capsid folds, and promoter-like elements may map near their prokaryotic or host analogs. It is also possible that the exclusion was so complete that eukaryotic viral sequences collapse to a high-entropy, uninformative region of latent space. The paper gives no evidence either way. If you want to use Evo-2 for this, you would need to test it yourself: embed a labeled set of eukaryotic viral genomes and check whether the resulting vectors cluster by taxonomy, Baltimore class, or host range in a way that outperforms raw k-mer counts. Until someone runs that experiment, claims about embedding utility are speculation.

Your hypothesis is sharp, and it exposes a real gap in the published work.

No — there is no Evo foundation model that included eukaryotic viruses. Evo-1 was trained on bacteria and bacteriophage only. Evo-2 expanded to all domains of life plus bacteriophage, but explicitly excluded eukaryotic viruses in both cases.

On embeddings: You are right that my previous claim was too strong. The Evo-2 paper only verified that perplexity and generation are broken on eukaryotic viruses. They did not test whether the hidden-state embeddings still cluster eukaryotic viral sequences near their prokaryotic viral analogs.

Your reasoning is biologically sound. Eukaryotic viruses and bacteriophages share protein folds — jelly-roll capsids, helicase domains, polymerase motifs — and Evo-2 has seen those folds expressed in prokaryotic viral contexts. It is plausible that a eukaryotic viral sequence containing a helicase or capsid protein would map closer to bacteriophage embeddings than to host embeddings, because the model has learned to associate those amino-acid-encoding DNA patterns with viral function.

The honest answer: It is an open empirical question whether Evo-2 embeddings retain partial structure for eukaryotic viruses. The paper gives no evidence either way. If you want to test your hypothesis, embed a labeled set of eukaryotic viral genomes (e.g., NCBI RefSeq viruses) and check whether they cluster by Baltimore class or family in t-SNE/UMAP relative to bacteriophage and host embeddings. If they form coherent clusters near prokaryotic viral regions, your intuition is correct and the embeddings are usable despite the generation failure. If they scatter as high-entropy noise, the exclusion was too complete.

Yes. There are three viable paths to use transformer-learned high-dimensional structure for eukaryotic viral dark matter classification.

DeepVirus is a hierarchical transformer framework explicitly built for viral dark matter. It models viral genomes as structured sequences of protein-coding genes, combines protein-level foundation model embeddings with genome-aware transformer representations, and performs open-set recognition to detect novel viral lineages that do not match existing taxonomy. Applied to metagenomic data, it uncovered previously uncharacterized RNA-dependent RNA polymerases and expanded the known evolutionary space of RNA viruses. This is the closest thing to what you are describing: a transformer that learns high-dimensional genomic structure to classify and discover uncharted viral sequences.

Evo-2's architecture is capable; the limitation is the learned distribution, not the model class. Evidence shows that fine-tuning Evo-2 on excluded viral domains partially rescues performance — perplexity drops and the model learns to predict the fine-tuned viral sequences. If you gather a diverse eukaryotic viral dataset and fine-tune Evo-2 with LoRA or full fine-tuning, you can reallocate its latent space to encode eukaryotic viral structure. The compute requirement is substantial but the weights and training framework (Savanna, BioNeMo) are open.

Unlike Evo-2, DNABERT-S explicitly included 17,636 viral genomes in its training data alongside bacteria and fungi. It learns species-aware DNA embeddings that cluster and segregate sequences by taxonomy in the latent space. While it was optimized for species differentiation and metagenomics binning, the embedding space already has viral structure allocated. You can fine-tune it for viral taxonomy classification or open-set novel virus detection, similar to the ViTax approach which fine-tunes HyenaDNA for viral taxonomy.

If you want to work with viral dark matter today without training a model from scratch, DeepVirus is the best pick. It was designed for exactly your problem: metagenomic viral discovery with transformer-learned structure, including open-set recognition for sequences with no known relatives. If you want to use Evo-2 specifically, you must fine-tune it on eukaryotic viral data first; the base model embeddings are unreliable for that domain.