How to Use Evo 2

Evo 2 is a genome language model for DNA sequence analysis and generation. The paper presents it as a long-context nucleotide model for genome modeling and design across all domains of life, which makes it useful for both interpreting local sequence changes and generating plausible new sequence in genomic style.

On Neurosnap, Mode switches between two distinct workflows. Zero-Shot Variant Scoring evaluates single-nucleotide variants from genomic coordinates using local reference context from hg19 or hg38, while Sequence Autocomplete continues a user-provided DNA prompt. That split makes Evo 2 useful both for variant-prioritization tasks and for exploratory synthetic-biology or genome-design ideation.

In scoring mode, Evo 2 treats variant effect analysis as a zero-shot language-model problem. The model compares how well it explains the reference-context sequence versus the alternate-context sequence around a candidate SNV, so researchers can rank variants by the change in model score without training a task-specific predictor for each assay.

In generation mode, the same genome model is used autoregressively to extend a DNA prompt. On Neurosnap, Model Name, Number Sequences, N Tokens, Temperature, Top K, and Top P expose the main tradeoffs between conservative continuation and broader sampling diversity. This is useful when a researcher wants multiple candidate continuations from one seed sequence rather than one deterministic completion.

Researchers should interpret Evo 2 differently depending on mode. For variant scoring, the main signal is the difference between alternate and reference scores, which is best used as a ranking prior for downstream experimental or annotation follow-up. For autocomplete, the output is better treated as a panel of plausible DNA continuations to inspect for motif content, composition, or design constraints rather than as guaranteed functional constructs.

• Dual workflow coverage: Evo 2 supports both SNV scoring from genomic coordinates and DNA prompt continuation, covering analysis and generation in the same service.
• Context-aware variant scoring: Genome selection lets researchers evaluate variants in hg19 or hg38 sequence context instead of as isolated base substitutions.
• Generation control: Sampling settings such as Temperature, Top K, Top P, continuation length, and sequence count make it practical to tune diversity for sequence-design exploration.
• Research workflow fit: Neurosnap presents score-based variant ranking or generated DNA candidates directly in a no-code workflow that can feed into downstream genomics, annotation, or synthetic-biology pipelines.