Genome-scale context, not just proteins. ESM2 and OpenFold operate on protein sequences or single protein structures. Viral dark matter is dominated by ORFans—genes with no detectable homologs in any database. Evo2 processes long genomic contexts (up to 1M base pairs), allowing it to capture the regulatory architecture, gene neighborhood, and non-coding regions that viral proteins depend on. ESM2 gives you an embedding for a single protein; Evo2 gives you the genomic ecosystem that protein lives in.

Trained on the full tree of life, including viruses. Evo2's training data spans bacteria, archaea, eukaryotes, and viruses. This matters because viral proteins often evolve via horizontal gene transfer, gene shuffling, and rapid divergence—patterns that require cross-domain genomic context to detect. ESM2 was trained on UniRef and MGnify protein clusters, which under-represents the extreme divergence of viral sequences. OpenFold is a structure prediction tool; it tells you how a protein folds, not whether that protein is biologically meaningful in a viral context.

Functional impact prediction without homology. The viral dark matter problem is fundamentally a "no homologs" problem. Up to 99% of viral sequences in metagenomes share no homology with known viruses. Evo2's architecture is built to predict functional impacts of genetic variation and identify functional elements in sequences that have no BLAST hits. ESM2 embeddings can help with remote homology detection, but they still rely on the assumption that the protein fold or sequence pattern exists somewhere in the training data. Evo2 can infer function from genomic context even when the protein itself is entirely novel.

Eukaryotic viral proteins are in the training data. ESM2 was trained on UniRef and MGnify protein clusters, which include eukaryotic viral proteins from databases like RefSeq Viral and UniProt viral proteomes. The embeddings carry signal for eukaryotic viral sequence features, even for remote homologs.

The ORFan problem is still hard, but tractable. ESM2 embeddings can detect structural and functional similarity beyond primary sequence homology. For eukaryotic viral dark matter—where you have protein sequences but no genomic context—ESM2 gives you a dense representation that captures evolutionary and functional signal better than raw sequence or BLAST.

OpenFold is downstream, not discovery. OpenFold predicts structure from sequence. It does not help you find or prioritize novel viral proteins in a metagenomic pile. Use it after ESM2 identifies candidates worth folding.

Input: A protein sequence as a string of amino acids (e.g., "MKTAYIAKQRQISFVK").

Processing: The model reads the sequence through a transformer architecture (33 layers, 15 billion parameters in the largest version) and produces a fixed-length vector for each amino acid position plus a single vector for the whole sequence.

Output: Embeddings that encode biophysical properties—secondary structure, solvent accessibility, contact maps, functional sites—without ever seeing the protein's 3D structure during training.

Viral dark matter is 90% non-coding and genomic context. Viral genomes are compact, with overlapping reading frames, regulatory RNA elements, and host-interaction domains scattered in non-coding regions. ESM-2 only sees amino acid strings. It misses the genomic architecture that defines viral replication strategy and host tropism.

ORFans are ORFans for a reason. The "dark matter" proteins that ESM-2 might embed are, by definition, those with no detectable homology to anything known. ESM-2 embeddings capture evolutionary signal from training data. If a viral protein has no evolutionary relatives in UniRef, ESM-2 has no meaningful signal to embed—it just generates a vector of noise with no interpretable structure.

No host-virus interaction signal. Eukaryotic viral dark matter is interesting because of how the virus manipulates host machinery. ESM-2 has no concept of host proteins, cellular compartments, or immune evasion. It embeds a viral protein in isolation, stripped of the biological context that makes

ESM2 for protein-level work — limited, but at least the training data includes eukaryotic viral proteins.

Train a custom model — collect eukaryotic viral genomes and train a genomic language model from scratch or fine-tune a general DNA model.

Hybrid approach — use general DNA models (Nucleotide Transformer, HyenaDNA) that were not explicitly safety-filtered for eukaryotic viruses, though they may still underperform due to underrepresentation in training data.

The exclusion was training-data-level. Eukaryotic viral sequences were removed from the pretraining corpus. If the model never saw these sequences, the embedding space has no reason to allocate meaningful structure to them. They likely project to the same "unknown" region as random DNA.

Embedding quality correlates with generation quality. In language models, representations for out-of-distribution inputs typically collapse to high-entropy, uninformative states. The high perplexity suggests the model has no internal model of eukaryotic viral sequence structure, which implies the embeddings encode noise rather than biology.

No evidence of emergent generalization. The paper notes that Evo2 fails on eukaryotic viral fitness prediction. If embeddings captured latent biological signal, you would expect some correlation with experimental fitness even if generation is poor. The absence of this correlation is weak evidence that embeddings are also uninformative.