← Về thư mục
📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / skills / bioskills / bio-genome-engineering-grna-design / references / usage-guide.md

Guide RNA Design - Usage Guide

Overview

Designs and ranks CRISPR guide RNAs for Cas9/Cas12a gene knockout. The skill scans a target for the chosen nuclease's PAM, applies hard filters (Pol-III TTTT terminator, 5' G, GC), ranks on-target activity with the model that matches the delivery context, picks the cut site by exon/transcript biology, and predicts the indel/frameshift outcome -- because the indel distribution, not the cutting rate, decides whether a gene is knocked out. On-target activity and off-target specificity are two separate axes; this skill owns the activity/site-selection axis and hands off specificity to off-target-prediction.

Prerequisites

pip install biopython
# CRISPOR (the de facto standard aggregator) is web (crispor.tefor.net) or a CLI clone;
#   it selects the context-appropriate on-target score and nominates off-targets per genome.
# Outcome prediction: inDelphi, FORECasT, Lindel are web servers / authors' code.

There is no maintained pip install crisprscan package with a generic .score() API -- CRISPRscan is a web tool/model. Route real on-target scoring to CRISPOR rather than hand-rolling a scoring matrix.

Quick Start

Tell your AI agent what you want to do: - "Design guides to knock out BRCA1" - "Find SpCas9 guides in an early constitutive exon of TP53 and rank them" - "My guides are for an in-vitro-transcribed RNP injection -- which on-target score applies?" - "Design Cas12a guides for this AT-rich target" - "Which of these guides has the most frameshift-prone repair outcome?"

Example Prompts

Knockout guide design

"I have the coding sequence of KRAS. Find SpCas9 NGG guides in the first two constitutive coding exons, drop any with a TTTT or GC outside 40-70%, and give me the top 5 ranked so I can run them through CRISPOR for on-target scoring."

"Design guides to knock out MYC, but explain which exon to target and why -- I keep getting edited cells with no phenotype."

Context and nuclease choice

"My guides will be expressed from a U6 lentiviral vector in a pooled screen -- which on-target model is valid, and should I tile the functional domain instead of the 5' exon?"

"This target is AT-rich and has no good NGG. What are my options?"

"I need AAV delivery in vivo -- design SaCas9-compatible guides and note the PAM change."

Outcome and validation

"Rank these candidate guides by frameshift probability, not just cutting efficiency."

"I confirmed a frameshift but the protein is still there -- what could be happening?"

What the Agent Will Do

  1. Establish the delivery context (nuclease, U6/lentiviral vs in-vitro T7/RNP, cell type, genetic background) -- this sets which PAMs exist, which on-target score is valid, and which hard filters apply.
  2. Enumerate candidate spacers adjacent to valid PAMs on both strands.
  3. Apply hard filters (reject TTTT for U6/H1; prepend a 5' G if absent; flag GC extremes).
  4. Route on-target ranking to CRISPOR (context-appropriate model), treating the score as a shortlist signal, not an oracle.
  5. Choose the cut site by exon biology: early, constitutive, NMD-competent, away from splice sites; tile the conserved domain for screens.
  6. Predict the editing outcome (Bae out-of-frame score; inDelphi/FORECasT/Lindel) and prefer frameshift-rich, dominant-outcome guides.
  7. Return 3-6 candidates and hand off to off-target-prediction before ordering.

Tips