bio-genome-engineering-prime-editing-design

What this skill does

## Version Compatibility

Reference examples tested with: BioPython 1.83+, PrimeDesign 1.2+ (Docker), PRIDICT2.0 (web/code).

Before using code patterns, verify installed versions match. If versions differ:
- Python: `pip show <package>` then `help(module.function)` to check signatures
- CLI: `<tool> --version` then `<tool> --help` to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

PrimeDesign is Docker-only (no pip) and takes the edit inline in a single string with exact parenthesis notation (below) -- the most-hallucinated thing in PE tooling; verify it against the repo, never reconstruct from memory. Outcome-prediction models are trained mostly on HEK293T + small edits (<=3 bp); their scores are priors, not measurements, and degrade off-distribution. The PE *system* (MMR status, expressed vs synthetic pegRNA) drives efficiency more than any oligo tweak.

# Prime Editing Design

**"Install a precise small edit without a double-strand break"** -> Establish the edit, cell type, MMR status, and delivery; choose the nick position/strand; design a *panel* of PBS x RTT combinations; pick the PE system; add the free wins (PAM-disrupting + MMR-evading silent edits, a 3' motif); rank with a model; and test.
- CLI (Docker): `PrimeDesign` generates ranked pegRNA + nicking-guide components from a reference + edit string
- Python: assemble/sweep PBS x RTT panels with `Bio.Seq`; enforce the don't-end-on-C and 5'-G rules
- Web/code: PRIDICT2.0 / DeepPrime rank candidates by intended-edit and indel rate

## The Single Most Important Modern Insight -- there is no universal PBS/RTT optimum, and the *system* choice carries the order of magnitude

Two reframes:

1. **PBS and RTT length are parameters to optimize per locus, not constants to look up.** The PBS x RTT optimum is locus-specific -- it depends on local GC (which sets the PBS annealing Tm), the edit, the nick-to-edit distance, and chromatin. A high-GC target wants a *short* PBS; a low-GC target a *long* one; the "13/15" that is perfect at one locus is useless 200 bp away. A hard-coded default produces a sequence that *looks* valid, so nothing flags it until the data come back at 2%. The correct deliverable is **a ranked panel** (a few PBS x a few RTT x the viable nicks), tested or model-ranked -- emitting a single pegRNA is the tell of someone who has never run PE.

2. **Prime editing efficiency is a cellular-genetics problem, not just oligo design.** The cell's mismatch repair (MMR; MutSalpha/MutLalpha) detects the edit:original heteroduplex and excises the *edited* strand, reverting it and spawning indels. The biggest post-2019 jump was not a better PBS -- it was **inhibiting MMR (MLH1dn -> PE4/PE5, ~7.7x average)**. The second was **stopping the pegRNA 3' end from being degraded** (epegRNA motifs; PE7's La protein). Design now means choosing the *system* (PE2 vs PE3b vs PE5max+epegRNA vs PE7) as much as the sequence. First branch: what edit, what cell type, MMR-proficient or not, expressed or synthetic.

## Mechanism (the design rules fall out of it)

The prime editor (Anzalone 2019) is **Cas9 H840A nickase + engineered M-MLV reverse transcriptase**, programmed by a **pegRNA** = sgRNA (spacer + scaffold) with a 3' extension read 5'->3' as **[RTT][PBS]**. (1) The nickase cuts the protospacer (PAM) strand ~3 nt 5' of the PAM, exposing a free 3'-OH. (2) The **PBS** anneals to that nicked 3' end (the genomic strand becomes the primer). (3) The RT extends through the **RTT**, synthesizing a new 3' DNA flap that *encodes the edit*. (4) FEN1-type nucleases preferentially excise the unedited 5' flap, favoring incorporation of the edited 3' flap; ligation seals it. (5) The resulting heteroduplex is resolved by MMR -- which preferentially reverts the edit (hence the MMR section below). Consequences: PBS length is tuned by annealing Tm; **RTT length = nick-to-edit distance + edit + 3' homology tail (~10-16 nt)**; efficiency falls as the edit moves farther from the nick; the edit must lie within the RTT.

## The PE System Stack -- orthogonal axes, not a "bigger number is better" ranking

| System | Adds over previous | Acts on | Cite |
|--------|--------------------|---------|------|
| PE1 | Cas9 H840A + **wild-type** M-MLV RT | proof of concept | Anzalone 2019 |
| **PE2** | **engineered M-MLV RT** (pentamutant) | the workhorse enzyme | Anzalone 2019 |
| PE3 | + second **nicking sgRNA** on the non-edited strand (~1.5-4x) | flap resolution / MMR strand bias -- **but raises indels** (transient near-DSB) | Anzalone 2019 |
| **PE3b** | PE3 ngRNA matching only the **edited** sequence -> nick fires *after* the edit | near-eliminates PE3's indels; **only possible when the edit makes/breaks a protospacer** | Anzalone 2019 |
| PE4 | PE2 + **MLH1dn** (dominant-negative MMR) (~7.7x avg) | MMR globally | Chen 2021 |
| PE5 | PE3 + MLH1dn | second nick + MMR | Chen 2021 |
| **PEmax** | optimized protein (codon, NLS, R221K/N394K, linker); +MLH1dn = **PE4max/PE5max** | the protein | Chen 2021 |
| PE7 | PEmax-family + **La-protein** RBD capping the pegRNA 3' end | pegRNA stability | Yan 2024 |

The expert move is to reason about which axis the problem needs: low efficiency in an MMR-active cell -> add MLH1dn; too many indels -> drop to PE2 or design **PE3b** (not PE3); short pegRNA half-life -> epegRNA/PE7. Note: in **MMR-deficient lines** (HCT116, many tumor lines) PE2 already behaves like PE4, so MLH1dn adds nothing -- benchmark numbers from such lines overstate the gain in MMR-proficient primary cells. **PE5max + epegRNA is the modern default workhorse** for hard, MMR-active contexts.

## pegRNA Parameters & the Free Wins

- **PBS (~8-17 nt; start ~11-15):** tune to annealing Tm/GC, not a fixed length. pegFinder's starting heuristic is PBS ~= 24 - (GC%/5), clamped 8-17; test a small ladder (e.g. 10/13/15/17).
- **RTT:** = nick-to-edit + edit + ~10-16 nt 3' homology. Shorter RTT is usually more efficient -- use the shortest that spans the edit with adequate homology, then test a couple.
- **Don't end the synthesized flap on a C** (a C at the +1 templated position lowers efficiency; PrimeDesign exposes `--filter_c1_extension`).
- **5' G for U6:** prepend a G if the spacer lacks one -- **prepend, do not replace** the first base (replacing creates a spacer:target mismatch).
- **PAM-disrupting silent edit (free win):** if the edit (or an added silent change) destroys the protospacer/PAM, the editor cannot re-nick the edited strand -> fewer indels, and the change doubles as an MMR-evading mismatch. Always check whether the edit can be routed to disrupt the PAM.
- **MMR-evading bystander edits (free win):** add 1-2 *silent* substitutions next to the intended edit to make a >=3-bp edited "bubble" that MMR recognizes less efficiently -> higher correct-edit yield. Trivial in coding sequence (synonymous codons); the tactic of choice before reaching for MLH1dn.

## epegRNA 3' Motifs & pegRNA Stability

The pegRNA 3' extension (RTT+PBS) is single-stranded RNA that is **exonucleolytically degraded** before it can prime RT -- an invisible failure (the molecule is made, just chewed back). **epegRNAs** append a structured pseudoknot motif to the 3' end (Nelson 2022): use **tevopreQ1** by default (~3-4x gain, no added off-target); **mpknot** is larger and benefits most from a **pegLIT**-designed linker (tevopreQ1/evopreQ1 often work linker-free). **PE7** (La protein) attacks the same degradation from the protein side and is **partly redundant** with epegRNAs (PE7's gains are largest with plain pegRNAs) -- don't stack them as if independent. For synthetic (non-expressed) pegRNAs where a folded motif is awkward, PE7 / La-optimized 3' chemistry is the lever instead.

## Outcome Prediction (rank, but still test)

| Model | Predicts | Cite |
|-------|----------|------|
| **PRIDICT / PRIDICT2.0** | intended-edit + unintended (i