Skip to main content

Review round results and decide

Analyze round results​

Once your lab cycle completes, compile your results data in the same consistent format you used before, and upload the experimental data to the same project.

Your Customer Success team will coordinate a review of the round:

  • how did the new variants perform against previous rounds
  • what trends appear
  • were there data-quality issues
  • did predictions align with reality

To set expectations, round one is primarily about learning: the model is building its understanding of your fitness landscape. From round two on, predictions should align with measurements and variants should trend toward your objectives, with the pace depending on problem difficulty and data quality. If results plateau, engage your Customer Success team: it may be your assay's ceiling, a hard property, or a compounding data issue.

If you decide to proceed to the next round, think through the following:

  • Do your objectives still make sense, or has the project evolved?
  • Should any constraint thresholds change?
  • Is the mutation load right, or do you want to explore more or less boldly?

To proceed to the next round, you will create a new round in your project and repeat the training, design and review steps.

Decide whether to continue or stop​

Your exit criteria should tie back to the project goals you set. If you hit your threshold, and have enough candidates meeting your constraints, the project can be deemed as completed and archived.

If the model's suggestions stop improving, or a round looks no better than the last, it can mean you are approaching the limits of your current assay setup, that the property is harder than expected, or that a data-quality issue is compounding across rounds. Talk to your Customer Success team: they can help diagnose whether it's a data issue, a configuration adjustment, or a genuine biological ceiling.

Tip: Two signals worth watching: predicted improvement is consistently low, and models keep proposing recombinations of known mutations, rather than proposing novel mutations.

Last updated on