Cas9 pre-loaded with a 20-nt single-guide RNA (sgRNA) doesn't scan raw DNA sequence directly — it diffuses along the helix probing for short PAM motifs (5'-NGG-3' for SpCas9). Only at a PAM does it locally melt the DNA and let the guide RNA try to base-pair with the target strand, growing an R-loop from the PAM outward. Pairing in the 8–10 nt seed region right next to the PAM is checked almost perfectly — a seed mismatch collapses the R-loop before it can propagate. Mismatches further from the PAM (the distal/PAM-distal region) are tolerated more, which is exactly why off-target edits cluster there in real genomes.
P(cleave) = exp(-k_seed · m_seed) · exp(-k_distal · m_distal)
k_seed = 0.9 × fidelity (steep — seed mismatches are punished hard)
k_distal = 0.15 × fidelity (shallow — a few distal mismatches are survivable)
- Specificity (fidelity) — mirrors engineered high-fidelity variants (eSpCas9, HiFi Cas9): raising it steepens both exponentials, so off-target sites with any seed mismatch are rejected almost every time, while the perfect on-target site (0 mismatches) still cleaves with probability 1 regardless.
- Candidate PAM sites — how many NGG loci exist in this stretch of genome; more candidates means more decoys for Cas9 to sample (and reject) before it reaches the true target.
- Scan speed — the 1-D sliding rate along the backbone during facilitated diffusion; only changes how long the search clock runs, not whether a site gets cut.
Real-world relevance: this PAM-first, seed-then-distal checking order is the actual mechanism behind CRISPR-Cas9 genome editing, and it's precisely the mismatch tolerance in the distal region that off-target-prediction tools and high-fidelity Cas9 engineering both target.