CRISPR Off-target Panel Benchmark

This benchmark connects Helix's CRISPR off-target search + scoring engine to assay-like hit counts (e.g., CHANGE-seq / CIRCLE-seq).

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

Spec shape

The off-target panel config declares:

• genome.fasta – digital DNA reference used for off-target search.
• defaults – shared search parameters:
  • pam – name of a registered PAM (e.g. SpCas9-NGG).
  • max_mm, max_gap – mismatch / gap limits for enumeration.
• guides[] – per-guide entries with:
  • id, sequence, pam override (optional).
  • search – optional per-guide search params.
  • assay – optional off-target results:
    • file – path to a TSV.
    • columns – mapping of column names for chrom, start, strand, read_count.

Running the benchmark

Example invocation:

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

The harness:

• Enumerates off-target candidates using helix.crispr.score.enumerate_off_targets over the panel's genome window.
• Scores hits with helix.crispr.score.score_off_targets (default weight profile).
• Loads assay hit counts, aggregates by (strand, start), and compares:
  • Per-guide metrics: predicted vs observed site counts and a Pearson correlation between predicted scores and log1p(read_count) over overlapping sites.
  • Panel-level summary: number of guides evaluated and mean Pearson correlation.

The JSON output (crispr_offtarget_panel schema) is designed for CI drift checks and for supporting methods-text statements about off-target risk calibration.