Helix StudioGenome IDE
← Docs
Helix CLI docs
Browse Helix CLI docs

Prime Editing Panel Benchmark

This benchmark ties Helix's prime editing outcome model to observed outcome spectra for a panel of targets and pegRNAs.

  • Spec template: templates/prime_editing_panel.helix.yml
  • Harness: benchmarks/prime_panel.py

Spec shape

The prime panel config declares:

  • Global defaults for simulation (draws, priors_profile, seed).
  • A targets list where each entry defines:
    • reference_sequence – digital DNA window around the edited site.
    • peg – pegRNA fields:
      • id, spacer, pbs, rtt
      • optional simulation overrides per target.
    • measurements – optional outcome counts table:
      • counts_file – path to a CSV.
      • label_column, count_column – per-row outcome label and integer count.
      • optional timepoint_h metadata.

Running the benchmark

Example invocation:

python3 -m benchmarks.prime_panel \
  --config templates/prime_editing_panel.helix.yml \
  --out bench-results/prime_panel.json

The harness:

  • Runs helix.prime.simulator.simulate_prime_edit on each reference window with the configured pegRNA and prime priors profile.
  • Aggregates outcome probabilities by label and, when counts are available, compares simulated vs observed distributions:
    • Per-target metrics: L1 distance and Jensen–Shannon divergence between simulated and observed label distributions.
    • Panel-level summary: mean L1 and JS divergence across all targets.

The JSON output (prime_editing_panel schema) is intended for CI drift monitoring and for supporting methods-text claims about prime editing model calibration.