CRISPR Efficiency Panel Benchmark

This benchmark ties Helix's CRISPR on-target scoring model to observed edited fractions for a panel of endogenous-like sites.

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

Spec shape

The efficiency panel config declares:

• Global cas settings (Cas system type, PAM pattern, mismatch weights).
• A targets list where each entry defines:
  • reference_sequence – digital DNA window for the locus.
  • guide – guide id, sequence, and optional strand.
  • measurements.edited_fraction – either:
    • value: inline edited fraction for quick tests, or
    • file + column: path to a TSV with an edited fraction column.
  • Optional chromatin_features (ATAC scores, histone marks, etc.) passed through to the JSON output.

Running the benchmark

Example invocation:

python -m benchmarks.crispr_efficiency_panel \
  --config templates/crispr_efficiency_panel.helix.yml \
  --out bench-results/crispr_efficiency_panel.json

The harness:

• Builds a CasSystem + GuideRNA from the panel's cas and per-target guide.
• Uses helix.crispr.simulator.predict_efficiency_for_targets (the public CRISPR engine batch API) to obtain physics-based on-target scores for each reference window.
• Compares predicted scores to observed edited fractions and reports:
  • Per-target status and absolute error.
  • Panel-level mean absolute error (MAE) and Pearson correlation (when at least two observations are available).

The JSON output (crispr_efficiency_panel schema) is suitable for CI drift tracking or methods-text calibration summaries.