bio-genome-engineering-base-editing-design

What this skill does

## Version Compatibility

Reference examples tested with: BioPython 1.83+.

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

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

BE-Hive is NOT a pip package -- it is a web tool (crisprbehive.design) plus clone-only repos (`maxwshen/be_predict_efficiency`, `be_predict_bystander`) pinning old dependencies (scikit-learn 0.20.3, BioPython 1.73). Use the web tool for a quick answer, the clone for batch use. Editing windows and outcome predictions are **editor-variant- and cell-type- specific** and are activity gradients, not hard boxes -- record which editor model was used.

# Base Editing Design

**"Install a transition mutation without a double-strand break"** -> Write the edit as a base-pair change to pick the editor family, find a PAM that lands the target base at the window peak while keeping bystanders out, read sequence context, choose the editor variant, and report the predicted genotype spectrum -- not a lone efficiency number.
- Python: window/bystander scan with `Bio.Seq` + `re`; per-base context reading
- Web/clone: BE-Hive / DeepBE for the per-base efficiency + bystander genotype spectrum
- Web: BE-Designer (guide enumeration), BE-Analyzer / CRISPResso2 (validate from NGS)

## The Single Most Important Modern Insight -- the job is product PURITY, not editing efficiency

A base editor produces a *distribution of genotypes at one site*, not a binary cut. "80% editing" can mean 80% of alleles carry the clean intended edit, or 80% carry *some* edit -- a soup where a bystander C/A two positions away is also converted half the time, so the exact desired genotype is only 30% of alleles. Both report as "80% efficient." The number that answers the biology is **precise editing: the fraction of alleles with the target edit AND no bystander edit.** Design is dominated by three coupled levers on one ~5-nt window: **(1) where the target base sits** (set by PAM/guide; drives efficiency), **(2) what else sits in the window** (bystander C's/A's; drives purity), **(3) the local dinucleotide context** of each base (drives which actually edit). The corollary trap: a *more active* editor (ABE8e over ABE7.10) raises headline efficiency while *lowering* purity (higher processivity sweeps more bystanders). Rank guides by the predicted **outcome spectrum**, never by efficiency alone.

The second field-defining insight: **base editors have THREE off-target classes, and the two that distinguish CBE from ABE are invisible to every Cas-off-target tool** (below). Most people only think of class 1 (guide-directed); the classes that matter for safety are Cas-independent and guide-invisible.

## Mechanism & the Position-Numbering Convention (get this right)

Both families are a **ssDNA deaminase fused to a Cas9 NICKASE (nCas9, D10A)**, guided by an ordinary sgRNA -- **no double-strand break**. When Cas9 binds, the non-target (PAM-containing) strand is displaced as ssDNA in the R-loop; the deaminase edits bases on that displaced strand within the window. **Canonical numbering (Komor 2016): position 1 = PAM-DISTAL (5' end of the spacer), position 20 = PAM-PROXIMAL (next to the PAM at 21-23). The editing window is ~positions 4-8, peak ~5-7.** (This is the opposite of what older tutorials sometimes say.)
- **CBE:** cytidine deaminase (rAPOBEC1) C->U, + **UGI** (blocks uracil excision -- the main CBE purity determinant), nCas9 nicks the unedited strand -> resolves to C->T (G->A other strand).
- **ABE:** lab-evolved TadA* deaminase A->inosine (read as G) -> resolves to A->G (T->C other strand). No UGI needed (inosine is not efficiently excised) -- a key reason early ABE was intrinsically cleaner than CBE.

## The Window Is a Gradient, Editor-Specific -- not a box

Activity inside the "box" is a steep gradient: a target at position 6 edits far better than one at 8; a bystander at 8 edits at a fraction of one at 5. The design move is "how close to the 5-6-7 peak is my target, and how far toward the cold edges can I push the bystanders?" The width is a property of the **editor variant**: ABE8e's high processivity widens it to ~3-11, so a guide that was bystander-clean with ABE7.10 becomes dirty when "upgraded" to ABE8e without re-checking. Carrying a window assumption across an editor switch is the most common silent failure.

## Editor Variant Selection

| Editor | Class | When |
|--------|-------|------|
| **BE4max / AncBE4max** | CBE | default modern CBE (Koblan 2018) |
| **ABEmax** | ABE | strong default ABE (Koblan 2018) |
| **ABE8e** | ABE | maximum activity (hard targets, screens) -- but WIDER window, MORE bystanders/off-target, LOSES context preference (Richter 2020; Lapinaite 2020) |
| YE1 / SECURE-BE3 | CBE | narrowed window / low RNA off-target (Grunewald 2019) |
| **TadCBE / TadDE** | CBE / dual | lowest Cas-independent DNA+RNA off-target CBE (Neugebauer 2023) |
| CGBE1 | C->G | the only transversion a base editor does well (Kurt 2021) |
| A&C-BEmax / SPACE | dual | simultaneous C->T and A->G (niche; adds bystander surface) (Grunewald 2020; Sakata 2020) |
| SpG / SpRY base editors | any | when no canonical NGG positions the base (Walton 2020) -- more class-1 off-target |

Default BE4max/ABEmax; reach for ABE8e for activity at a purity cost; a SECURE/TadCBE variant when off-target matters; a PAM-flexible editor when no NGG positions the base. The *deaminase* choice (rAPOBEC1 vs APOBEC3A vs eA3A vs evolved TadA) sets window width, context, and off-target far more than the BE3-vs-BE4 generation number.

## The Three Off-Target Classes -- ask "which of the three?" first

| Class | Mechanism | Detected by | Mitigation |
|-------|-----------|-------------|------------|
| **1. Cas-dependent DNA** | guide mismatch-tolerance, like ordinary Cas9 | GUIDE-/CIRCLE-seq; in-silico (-> off-target-prediction) | high-fidelity Cas; guide selection |
| **2. Cas-INDEPENDENT DNA** | deaminase edits transiently-exposed genomic ssDNA, **no guide** | GOTI (Zuo 2019); WGS (Jin 2019) | choose a low-activity-deaminase variant (YE1, **TadCBE**) |
| **3. RNA** | deaminase edits cellular mRNA, **no guide**, transient | RNA-seq (Grunewald/Rees/Zhou 2019) | **SECURE** variants (rAPOBEC1 R33A; ABE F148A); TadCBE |

Classes 2-3 have no protospacer, so they are **invisible to every guide-based predictor** -- the guide cannot be designed to avoid them; only a cleaner editor can. **CBE has historically been the dirtier family on classes 2-3** (Zuo/Jin: CBE >20x background Cas-independent SNVs, mostly C>T in transcribed DNA; early ABE did not) -- a genuine input to the CBE-vs-ABE choice. But this is *variant*-specific, not family-destiny: ABE8e raised it back up; TadCBE pulled CBE's down. Class-3 RNA edits are transient (RNA turns over) -- a real dose/exposure-dependent liability for chronic/AAV/therapeutic use, often tolerable for a transient cell-line transfection.

## Knockout by Base Editing (no DSB)

Base editing knocks out a gene without a double-strand break -- no indel lottery, no large deletions/translocations, no p53 response, works in **non-dividing cells**, and safe for **multiplex** (no translocations between simultaneous cut sites). Two routes:
- **Premature stop (CRISPR-STOP/iSTOP):** a CBE converts **CAA->TAA, CAG->TAG, CGA->TGA** (sense) or **TGG->stop** via the antisense strand. Only these four codons are reachable; target an **early** stop (before functional domains, NMD-competent). iSTOP precomputed sgRNAs cover 97-99% of genes (Billon 2017; Kuscu 2017).
- **Splice-site disruption:** edit the invariant splice-donor **GT** or acceptor **AG** -> mis- splicing / exon skipping (Kluesner 2021). Often the more robust KO (every gene has many junctions; only ~half have a well-placed early stop codon).

The quiet failure mode: a ba