The Smallest Red Pen

The letters on the bench are 201 bases of KRAS read from GRCh38 chromosome 12, positions 25,245,250–25,245,450, taken from Ensembl release 116 and shown on the coding strand. KRAS sits on the minus strand, so the letters you read are the reverse complement of the forward genome. The build script translates the whole coding sequence and refuses to emit a file unless it comes out as the known 188-residue KRAS protein and codon 12 reads GGT.

about 40% of colorectal cancers carry a KRAS mutation, rising to roughly 45% in metastatic cohorts, and about 82% of KRAS mutations in colorectal cancer sit in codon 12 or 13—which is why the bench opens on codon 12. Peritoneal metastasis is the setting because it is where a locally delivered editor is even thinkable. 5.7% of colorectal cancers already show peritoneal metastases at diagnosis; a further 5.5% develop them within three years.

TCGA COADREAD (218/534) and MSK-IMPACT metastatic CRC (495/1134), via cBioPortal; npj Precision Oncology, 2025; Netherlands Cancer Registry, 7,233 patients, BJS Open 2020

A mutation that creates a PAM is the clean case for allele-specific editing: the tumor copy can be opened and the healthy copy simply cannot. At KRAS codons 12 and 13 the opposite happens. The wild-type sequence reads …GCT GGT GGC…, and those GG pairs are themselves PAMs. Every common mutation here destroys one and creates none, so the only guide that can tell the copies apart by PAM alone is the one that cuts the patient and spares the cancer. Try each variant: the count in the ledger stays at zero.

With no PAM to separate them, a mutant-specific guide shares its PAM with the healthy copy and differs from it by a single mismatch. How well that works depends on where the mismatch falls: positions 2 to 8 from the PAM are the most mismatch-sensitive stretch. Position 1, hard against the PAM, is not among them — a single mismatch there is tolerated more often than one a little further along.

That is exactly the trap at codon 12. G12D and G12V put the mismatch at position 1; G12C puts it at position 2. When these guides were measured in cells, the position-1 guides cut 12–15% of healthy copies and the position-2 guides 2–3%. The most common variant in this cancer is the hardest one to target cleanly.

Rabinowitz and Offen, Molecular Therapy 2021, pooling six single-mismatch studies; Gao et al., Theranostics 2020

Editing measured on wild-type KRAS in 293T cells by mutant-specific guides, against 66% for the guide that matches wild-type exactly. The guide written for wild-type still cut a G12S tumor copy at 40%, so the sparing does not run both ways.

Gao et al., Theranostics 2020

There are 3.1 billion letters in a human genome, against 4²⁰ ≈ 1.1 trillion possible twenty-letter words — which is why a guide is twenty letters long and not twelve. On the non-N primary assembly that is about 5.9 billion places a twenty-letter word could start, against 1.1 trillion possible words, so a given twenty-letter match is expected roughly once in 186 genomes — and once in about 3,000 when you also demand a PAM. That argument explains why a perfect match is nearly unique. It says nothing about near matches, which is where real off-target cutting lives, and why this page does not pretend to run a genome-wide search.

GRCh38

Cas9 makes a break; it does not type a replacement. When the two ends are rejoined, short repeats that happen to sit on either side of the cut can pair up, and everything between them is lost. The repair panel enumerates those repeats and scores each resulting deletion with the published microhomology formula: a pattern score of 100 × e^(−L/20) × (bases + GC bases), summed, with the out-of-frame score being the share of that total belonging to deletions whose length is not a multiple of three.

The implementation here reproduces the authors’ own worked example exactly — 44 patterns, a microhomology score of 4662.9 and an out-of-frame score of 50.747 — and a test pins those numbers.

Bae, Kweon, Kim and Kim, Nature Methods 2014

If the break is the problem, the newest editors do not make one. A base editor carries a deaminase that chemically rewrites a single letter in place: an adenine editor turns A into G, a cytosine editor turns C into T. At G12D and G13D the mutation is exactly a G→A on the coding strand, so an adenine editor pointed at that letter writes the healthy G straight back — no double-strand break, no repair lottery. Switch the bench to Rewrite the letter and the window, the editor and the single swap are all computed from the same real sequence.

The catch is chemistry. A base editor only runs four of twelve point changes — the transitions. An adenine base editor rewrites an A·T pair to G·C and a cytosine base editor does C·G to T·A, both with no double-strand break. Between them they cover the four transition changes but none of the eight transversions. G12V and G12C are G→T transversions, so no base editor can touch them; the bench says so. And reaching codon 12 needs a relaxed PAM, NGG to NG, engineered SpG and near-PAMless SpRY Cas9 relax the mark the editor has to find, which is what brings the codon 12 and 13 adenines inside a base editor's narrow activity window at all.

Komor et al., Nature 2016 (CBE); Gaudelli et al., Nature 2017 (ABE); Walton et al., Science 2020

Something quietly important happens once the correction lands. The whole first half of this bench is about telling the two copies apart, because a nuclease that cuts the healthy copy does damage. But if the edit turns the tumor copy into the wild-type letter, then doing the exact same edit to the healthy copy changes a G into a G. It is a no-op. The allele-specificity problem does not get solved so much as it stops existing — there is nothing harmful left to do to the wrong copy.

Prime editing is the general case. It nicks one strand near the mutation and copies a new sequence in from a template tethered to the editor, with no double-strand break and no donor — so it can write all twelve point changes, plus small insertions and deletions, including the G12V and G12C transversions a base editor cannot. On the bench, under Write it back, a canonical NGG nick sits within reach of every variant here, and the edit is always the exact wild-type codon.

There is a way to make it work harder, and it comes with the page’s own moral attached. Nicking the other strand as well raises editing 1.5- to 4.2-fold more editing when the other strand is nicked too, in four of five sites tested. Nicks about 40 to 90 bases from the first averaged 41% editing without excess indels, and the authors recommend starting around 50. But two nicks close together on opposite strands are a double-strand break under another name, and the scramble from the first half of this bench comes back with them. Switch Write it back to Nick both strands and drag: the bench marks every second site that falls inside the recommended band, and every one that sits too near.

So why is this page not a cure? Because specificity was never the last wall. Prime editors are large, their efficiency in real tissue is modest, and getting enough editor into enough cells is the unsolved part — the same delivery problem note 15 ends on. The chemistry caught up; the plumbing has not.

Anzalone et al., Nature 2019

This is not hypothetical. In genome-wide CRISPR knockout screens, −1.55 against −0.51 is the median Chronos gene-effect score for knocking out KRAS in KRAS-mutant colorectal cell lines (n=34) versus KRAS-wild-type ones (n=29), Mann–Whitney p=1.3×10⁻⁹. A score of −1 is the median of genes every cell needs.

Two honest caveats. The wild-type median is −0.51, not zero, so plenty of KRAS-wild-type lines need KRAS too; this is a shift in a distribution, not a switch. And the effect tracks the genotype rather than the tissue — the all-lineage figure is essentially the same.

computed from DepMap 24Q4 Public, CRISPRGeneEffect.csv

The peritoneum is a bag you can fill. Filling it buys 10 to 1000 times the drug exposure inside the peritoneal cavity against the bloodstream when a drug is given into the cavity, depending on the agent; penetration into tissue is only millimetres, so it treats surfaces rather than bulk. That is a real pharmacological argument for treating this compartment locally, and it is the reason the idea is placed here rather than anywhere else in the body.

intraperitoneal pharmacology reviews, after Dedrick et al. 1978

Heating chemotherapy and washing the cavity with it is the obvious version of this idea, and the trial that tested it properly came back empty. PRODIGE 7 measured 41.7 vs 41.2 months median overall survival with and without oxaliplatin HIPEC after complete cytoreductive surgery, hazard ratio 1.00, p=0.99; late grade 3 or worse complications were more common with HIPEC, 26% vs 15%, p=0.035. The honest reading is that the surgery, not the wash, carried those 41 months — and that one drug at one dose failing does not retire the route.

PRODIGE 7, Quénet et al., Lancet Oncology 2021, n=265