Designing a CRISPR Guide RNA With AI

Guide design is two rankings pulling against each other: which guide cuts the target best, and which guide is least likely to cut anywhere else. Both are predictions from empirical models, and neither is a measurement of what happens in your cells.

What a guide has to be

For the most widely used nuclease, Cas9 from Streptococcus pyogenes, the rules that define a candidate are simple and hard. The guide RNA carries a 20-nucleotide spacer that base-pairs with one strand of the target DNA. The nuclease will only engage a site immediately followed, on the protospacer strand, by a protospacer-adjacent motif of the form NGG — any base, then two guanines. No PAM, no cut, regardless of how perfect the pairing is. Cleavage occurs about three base pairs upstream of the PAM, which is what makes the cut position predictable and therefore useful for repair-template design.

Enumerating candidates for a target region is therefore mechanical: scan both strands for NGG, take the 20 bases 5-prime of each, and you have your list. In a random sequence an NGG occurs roughly every 8 bases per strand, so a 200-base exon typically yields tens of candidates. The design problem is choosing among them, not finding them.

Other nucleases change the constants but not the shape of the problem. Different Cas9 orthologues and engineered variants recognise different PAMs, and Cas12a recognises a T-rich PAM on the other side of the protospacer and leaves staggered rather than blunt ends. Every scoring model below is fitted to a specific nuclease, and applying a model trained on one to another is a category error that produces confident-looking numbers.

On-target activity scoring

Guides targeting the same gene differ enormously in how efficiently they cut, and the difference is largely sequence-determined. On-target scoring models are supervised regressions fitted to large libraries of guides whose editing efficiency was measured experimentally — the widely used Rule Set 2 model was published by John Doench, Nicolo Fusi and colleagues in Nature Biotechnology in 2016, and the Broad Institute’s sgRNA scoring documentation describes what each of its scores is fitted to.

The features these models use are the useful part to know, because they tell you what varies:

• Position-specific nucleotide identity. Which base sits at which position in the spacer and the flanking sequence, as one-hot and dinucleotide features. Positions nearest the PAM contribute most.
• GC content of the spacer. Very low GC guides pair too weakly and very high GC guides too strongly; efficiency falls off at both ends rather than increasing monotonically.
• Runs of the same base. Four or more consecutive thymines act as a termination signal for the polymerase transcribing the guide, so a TTTT in the spacer truncates the guide itself. This is a hard exclusion, not a soft penalty.
• Secondary structure of the guide RNA. A spacer that folds back on the scaffold cannot load properly. The folding calculation here is the same energy minimisation described in RNA secondary structure prediction.

What these models do not see is the cell. Chromatin accessibility strongly affects cutting, and the same guide performs differently in different cell types and at different loci for reasons no sequence model can capture. A score is a prior over guides, and a poorly ranked guide that happens to sit in open chromatin can outperform a top-ranked one.

Enumerating the off-targets

Cas9 tolerates mismatches, so a 20-mer that appears once in the genome exactly may appear many times with a few substitutions. Off-target analysis is therefore a two-stage process: find every genomic site within a mismatch budget of the guide and followed by a valid PAM, then score each one for how likely it is to actually be cut.

The enumeration is a specialised search rather than an ordinary alignment, because the pattern is short and the mismatch budget is fixed — typically up to four. Dedicated tools do it exhaustively across a genome. Two properties of the result are worth anticipating: the number of sites grows steeply with the mismatch budget, so a guide with 3 sites at 2 mismatches may have several hundred at 4; and guides targeting repetitive or recently duplicated regions have off-target counts an order of magnitude above typical, which is usually reason enough to pick a different guide.

The enumeration also has a known blind spot. Searching for substitutions only will miss sites where the guide and the DNA can pair by bulging out an unpaired base on one side, and such bulged off-targets are real and are cut. A mismatch-only enumeration therefore has a false-negative rate it cannot report on.

Scoring one off-target, worked

Not all mismatches are equal. A mismatch near the PAM is far more disruptive than one at the distal end of the spacer, and the identity of the mismatched pair matters too — some pairings are tolerated much better than others. The Cutting Frequency Determination score encodes exactly this as a lookup table of percent-activity values, one per combination of position and mismatch type, measured from libraries of guides carrying single substitutions. The score for a site is the product of the per-mismatch factors, so it lies between 0 and 1.

guide      G A C T T G A C C T A G G T C A T C G T  (PAM: AGG)
off-target G A C A T G A C C T A G C T C A T C G A  (PAM: TGG)
                 ^                  ^            ^
             pos 4               pos 13       pos 20
                                            (PAM-proximal)

CFD = f(pos 4, T->A) x f(pos 13, G->C) x f(pos 20, T->A)

ILLUSTRATIVE factors, to show the shape of the arithmetic
only -- read the real values off the published matrix:

    0.90  x  0.60  x  0.20  =  0.108

The shape is what to take from this. A distal mismatch barely moves the score; a PAM-proximal one dominates it; and because the factors multiply, three moderate mismatches suppress a site more than one severe one. The published guidance is that sites scoring below about 0.02 are unlikely to be cleaved, which gives you a cut-off for the long tail. The factors above are placeholders chosen to illustrate the multiplication, not values from the matrix, and any real calculation must use the published table.

Individual site scores are then aggregated into one specificity number per guide. The common form divides a constant by one plus the sum of the individual off-target scores, so a guide with many weak off-targets and a guide with one strong off-target can land at similar aggregate values — which is a real weakness of any aggregate. A single off-target at 0.8 inside a tumour suppressor gene is a different risk from forty at 0.02 in intergenic space, and no scalar distinguishes them. Rank by the aggregate, then read the individual sites and their genomic annotations before choosing.

What the score cannot establish

Every number above comes from a model fitted to measurements made in particular cell lines with a particular delivery method at a particular nuclease concentration. Its predictions transfer to a new setting only approximately, and the direction of the error is not knowable in advance.

The empirical check is a separate experiment. Unbiased genome-wide methods exist precisely because computational prediction is not sufficient: some detect double-strand breaks in living cells by capturing an integrated tag at the break, others map nuclease activity on genomic DNA in vitro, and each has its own sensitivity and its own biases. These consistently find cut sites that prediction missed, and consistently find that many highly ranked predicted sites are not cut.

Beyond off-targets, several outcomes are invisible to any guide score and have to be measured: large on-target deletions and rearrangements at the cut site, chromosomal translocations when two sites are cut, the distribution of repair outcomes at the intended site, and effects on cell viability and differentiation.

So the honest framing is that guide design software prioritises candidates for testing. It is a research method. Any application in humans sits inside a regulatory framework, requires institutional and ethical oversight, and rests on experimental evidence rather than on predicted scores — and no ranking produced by these models is a safety assessment or a claim that an edit is appropriate for anyone.