Dyno Therapeutics

AI-focused gene therapy company that partnered with Anthropic on the sequence-to-function modeling and design evaluation for Claude Mythos Preview.

Evaluation Partnership

Dyno provided a medium-horizon challenge task for evaluating CB-2 capabilities. The task uses unpublished sequences (uncontaminated) and has also been administered to 57 human participants from the leading edge of the US ML-bio labor market since 2018.

Task Design

Participants (human or model) must:

  1. Analyze data and develop a model of sequence-to-function relationships from a small set of experimental measurements
  2. Predict the function of sequences in a test dataset
  3. Design novel sequences with the highest possible function

The task requires discovering non-trivial sequence attributes, engineering expressive model architectures, and making optimal design tradeoffs.

Scoring

Two automated metrics:

  • Prediction score: Spearman correlation with ground-truth function of test sequences
  • Design score: Ground-truth function of the best sequence proposed

Claude Mythos Preview Results

  • Mean performance exceeded the 75th percentile of human participants on both tasks
  • Exceeded 90th percentile on prediction score
  • Did not exceed the top human performer on either task
  • First model to nearly match leading experts in both sequence design and modeling
  • Moderate improvement over Claude Sonnet 4.6 and Opus 4.6; notable improvement over Opus 4.5 and Haiku 4.5