Inspiration

CRISPR off-target scores are usually computed against one reference genome. A common variant in a named population can change the CFD score without changing the mismatch count, or leave the score unchanged because the substitution lands on the flexible N of the PAM. Both outcomes are real method findings, and the second one is easy to throw away because the number did not move.

Population-aware scoring already exists. What we kept running into was a missing record: which assumption produced the conclusion, who chose the next expensive check, and whether a null was actually a null. CRISPR-Ops started there. Don't score the guide. Audit the assumption.

What it does

CRISPR-Ops audits how reference-only scoring, mismatch counting, and simplified haplotype rules change computational off-target conclusions under a named population's common variants. It records the investigation as a re-derivable trace. Nulls are results. It makes no claim about real cutting.

You give it a gene, a gnomAD population, guides, and a note. A policy gate runs first. A personal genome or a clinical question is refused before any tool is called. If the run is allowed, an investigator agent (NVIDIA Nemotron) writes a hypothesis and picks one legal next step. Allele frequencies, CFD scores, and verdicts stay in code.

How we built it

The cheap path is deterministic: C++ enumeration on GRCh38, gnomAD v4 allele frequencies, Doench 2016 CFD, and a second in-repo checker. Nemotron sits on NeMo Agent Toolkit with an OpenAI-compatible Nebius API. Every step lands in SQLite; FastAPI streams the trace. We ship a 20-guide × 3-population frozen grid (9,639 site rows) and offline replay when APIs are down.

Challenges we ran into

Eligibility and API base URLs (avoid double /v1 in the path). Keeping the model an investigator under a budget, not a pipeline. Reporting nulls like the chr11 PAM-N case (ΔCFD = 0) as findings, not failures.