Real CRISPR Benchmark

This report uses real public MAGeCK/Yusa CRISPR screen inputs, not synthetic reads.

• Guide library: yusa_library.csv

• FASTQ samples: ERR376998.fastq.gz, ERR376999.fastq.gz

• Extraction: --target-start 23 --target-length 19

• Assignment threshold: exact Levenshtein k=1

• Real reads benchmarked against Edlib scan: 25 extracted reads

• Real FASTQ records downloaded per sample: 25

• Correctness comparator: native Edlib exhaustive scan over the guide library

Native Real-Data Results

tool	workflow	n_reads	n_targets	len	k	seconds	reads_per_sec	candidates_per_read	verified_per_read	checksum	mismatches
dotmatch_indexed	public_crispr_yusa	25	87437	19	1	0.00	19669.60	0.84	0.84	14224824	0
edlib_native_scan	public_crispr_yusa	25	87437	19	1	1.80	13.90	87437.00	87437.00	14224824	0

Speedup

DotMatch indexed speedup vs native Edlib scan on this real-data subset: 1415.08x.

Workflow Comparator Availability

tool	status
mageck	not installed in this environment
cutadapt	not installed in this environment
bowtie2	not installed in this environment

MAGeCK, Cutadapt, and Bowtie2 are workflow comparators and are intentionally optional. Run python3 scripts/run_public_crispr_benchmark.py --run-mageck --run-cutadapt --run-bowtie2 in an environment with those tools installed to populate those rows.

Evidence Boundary

This benchmark supports known-target CRISPR guide assignment only under the listed extraction rules. Native Edlib remains the exact semantic oracle; MAGeCK/Cutadapt/Bowtie2 comparisons should be described as workflow comparisons, not identical semantic oracles.