| name | evo2-genome-modeling |
| description | Run pinned Evo 2 DNA sequence scoring, embeddings, variant scoring, or controlled generation. Use for long-context genomic foundation-model analyses when compatible local GPU hardware and explicit genome/strand provenance are available. |
| license | MIT |
Evo 2 Genome Modeling
Gate
Ask once before installing pinned Evo 2/Vortex dependencies, downloading large
checkpoints, using GPU/remote compute, or sending sequence to a hosted API.
State hardware compatibility and estimated memory/time. Follow
$cx-compute-environment.
Workflow
- Record assembly, chromosome/contig, coordinates, strand, reference allele,
sequence window, organism, and intended task. Validate reference alleles and
keep reverse complements/overlapping loci together in evaluation splits.
- Pin Evo 2 version/commit, checkpoint and checksum, tokenizer, context length,
embedding layer or scoring rule, precision, hardware, seeds, and generation
parameters/budget.
- Run the official installation smoke test, then execute into
artifacts/<run-id>/evo2/; retain exact sequences, coordinate maps, logits or
embeddings, generations, configs, environment, logs, and failures.
- For variants, compare matched reference/alternate windows and both strand
conventions when appropriate. For generation, report memorization, novelty,
composition, constraints, filter failures, and nearest database neighbors.
- Record with
$science-provenance; review with $science-review.
Boundaries
- Likelihood is not pathogenicity, fitness, expression, or clinical evidence.
- Generated DNA is a computational hypothesis; do not automatically synthesize,
deploy, or optimize for harmful biological function.