Evo 2 Generates Candidate Phage Genomes From Examples Without Fine-Tuning

Genome language models learn to predict the next base in DNA, and the usual way to aim one at a particular design task is fine-tuning, which means further training on a collection of the target sequences. The preprint reviewed here compares against exactly that ^[1]. Large language models can instead pick up a task from a few examples placed in the prompt, a behaviour called in-context learning, and the authors ask whether Evo 2, a genome language model trained on natural genomic sequence, does the same ^[1]. The work was posted to bioRxiv on 29 October 2026 and has not been peer reviewed. It reports designs for proteins, antibodies and whole phage genomes, all judged by computational filters.

The finding

Examples in the prompt steer generation

The authors place several natural sequences in the prompt, separated by short marker sequences, and let Evo 2 continue the text without updating any weights, then screen what it writes with computational filters ^[1], Methods. The number of examples varies by task, from 1 up to 128 for anti-CRISPR proteins, 4 to 32 natural CRISPR-Cas9 loci for Cas9 systems, and 4 or 8 natural genomes for the ΦX174-like phages ^[1].

Feeding generations back into the prompt

The second idea is what the authors call genomic chain-of-thought. As generation proceeds, the oldest natural examples in the prompt are replaced one at a time by newly generated sequences, so later prompts contain synthetic genomes as well as natural ones. The aim is more diverse designs at the same compute cost when prompts are refreshed at the same frequency ^[1]. They also describe a faster inference engine, up to 35-fold quicker than the original depending on context length, that fits a 1-megabase context on a single H100 GPU ^[1].

The evidence

Every result below is computational. The authors write that in silico metrics do not guarantee biological function and that all evaluations in the paper are purely computational ^[1].

Phage genomes against a fine-tuned model

The cleanest comparison is for ΦX174-like phages, using a fine-tuned Evo 2 as the baseline. A design counts as correct if it passes six checks: valid DNA letters only, a length of 4,000 to 6,000 bases, GC content of 30 to 65%, no homopolymer longer than 10 nucleotides, at least 7 genes with significant matches in a phage protein database, and 60 to 100% amino acid identity to the ΦX174 gene G. Novelty adds three more filters, on mean amino acid identity (at most 95%), a genome architecture score, and a gene order (synteny) criterion under which both the total and the syntenic gene counts must fall between 10 and 12, with the combination of 11 and 11 excluded ^[1], Methods.

In-context learning passed the ΦX174 correctness filters for 51.7% of designs, compared with 13.0% for a fine-tuned model, and 0.25% of in-context designs passed correctness and all three novelty filters, compared with 0.015%. Producing a library of 64 designs that pass the correctness filters and protein novelty took about 200 GPU hours with fine-tuning (162 with the faster engine) and 2.1 GPU hours with in-context learning on natural genomes ^[1]. The 0.25% figure is small in absolute terms.

Chain-of-thought and two other phages

With chain-of-thought, the share passing all three novelty filters rose from 0.25% to 2.9%, while the share passing the quality filters fell from 51.7% to 46.7% ^[1]. For the filamentous phage M13, prompted with 4 genomes, 92.8% of designs passed the correctness filters, and protein novelty rose from 36.6% to 49.6% with chain-of-thought. For the RNA phage MS2, prompted with 2 genomes, 57.5% passed correctness but none passed novelty. Only two natural MS2 genomes passed the authors' filters, and those same two supplied every prompt, with their order shuffled. The authors say the lack of novel designs likely reflects the high amino acid identity (99%) of MS2 genes across genomes, which they do not establish as the cause. Prompting Evo 2 directly, without examples, failed to complete either genome ^[1].

Proteins, CRISPR systems and antibodies

For anti-CRISPR proteins, with 128 examples, 47% of generations passed computational filters when the examples were type II and 23% when they spanned types I to VI, against 6% for a semantic design baseline. Greater variety among the examples produced more divergent designs but a lower pass rate ^[1].

Cas9 is the mixed case. Pooled across the prompt settings (4 to 32 loci) and sampling temperatures tested, in-context learning had a lower per-sample pass rate than a fine-tuned Evo 1 (0.14% against 0.39%). Fine-tuning consumed 881 H100 hours, so a library of 64 passing designs cost roughly 900 H100 hours against 81 hours for in-context learning. That 81-hour estimate uses the best arm, the 7-billion-parameter model with 8 loci in the prompt at temperature 1.0 ^[1]. The rate is therefore pooled and the cost is for the best arm. The baseline is Evo 1 and the in-context runs use Evo 2, so the comparison also mixes method and model.

For antibodies, the authors seeded the model with 8 de novo binders and targeted 7 antigens under a fixed budget of one GPU-hour for generation and one for scoring. They report it was the only method with at least one filter-passing design for every antigen, against baselines that included a structure-conditioned design tool and a position-specific score matrix ^[1]. A companion study on mammalian regulatory elements tested designs in cultured cells and mouse brain, but that is a different task and does not bear on the designs here.

Future directions

The paper says in-context learning emerges during pre-training but does not explore the mechanism, and it suggests that post-training could strengthen the behaviour. A proposed measurement is to run the same ΦX174 prompts through the 1-billion, 7-billion and 20-billion-parameter models, the sizes the paper mentions alongside its main 7-billion model, and report the correctness and novelty pass rates side by side. That would show whether the ability tracks model size or is already present in small models.

The Cas9 result leaves open how much of the gap comes from the method and how much from the model. The paper already varies the number of Cas9 loci from 4 to 32, so that sweep is a reported experiment. Two proposed follow-ups would sharpen the comparison: fine-tuning Evo 2 itself on the same Cas9 sequences and comparing it with in-context prompting at a matched compute budget, which would separate method from model, and reporting the pass rate of the best arm (8 loci, temperature 1.0) next to the pooled 0.14%, so that rate and cost describe the same condition. Neither has been run in the paper.

MS2 raises a different problem. The paper characterizes the amino acid identity of natural proteins in Figure S4C, and with only two natural genomes in the prompt pool the novelty filters rejected every design. A proposed test is to place the protein identity threshold against that natural identity range and ask whether any threshold separates a meaningful new MS2 variant from a near copy of a prompt genome. The paper does not report such a calibration.

The largest gap is function. The designs that pass all three ΦX174 novelty filters, and the chain-of-thought designs in particular, could be synthesised and tested for whether they produce viable phage. That would show whether the lower quality-filter rate under chain-of-thought (46.7% against 51.7%) matters biologically, and whether any of the 2.9% of designs that pass all novelty filters are alive.