🧬 CRISPR Guide RNA Off-Target Prediction

Defining the Search Space — Genome-Wide Protospacer-Adjacent Motif Scanning

Off-target prediction begins not with the guide RNA itself but with an exhaustive enumeration of every location in the reference genome where the nuclease could conceivably cut. For SpCas9 this means finding every NGG trinucleotide (the protospacer-adjacent motif, or PAM) on both strands of GRCh38, since Cas9 will not engage DNA lacking a compatible PAM immediately 3' of the target sequence regardless of spacer complementarity. This single structural requirement collapses an otherwise intractable 3.1-billion-base-pair search problem into a tractable candidate list.

• 3.1 Gb: GRCh38 genome size (both strands scanned)
• ~1 in 8 bp: NGG PAM frequency (≈2.1M sites genome-wide)
• 20 nt: Spacer length (immediately 5' of PAM)
• <3 min: Scan runtime (Bowtie2/Cas-OFFinder, indexed FM-index)

PAM identity dictates nuclease engagement before homology is even considered

Every CRISPR nuclease has a fixed PAM requirement encoded in its PI (PAM-interacting) domain, and this requirement is absolute — no amount of spacer-target complementarity rescues binding at a locus lacking the correct PAM. Streptococcus pyogenes Cas9 (SpCas9) requires 5'-NGG-3' immediately downstream of the 20nt protospacer on the non-target strand; the engineered variant SpCas9-NG relaxes this to a bare 5'-NG-3', roughly quadrupling the addressable genomic space at the cost of higher intrinsic off-target activity. Staphylococcus aureus Cas9 (SaCas9), favored for AAV-packaged in vivo delivery due to its compact 3.2 kb ORF, requires the longer 5'-NNGRRT-3' motif (R = A/G), which is more restrictive and inherently improves specificity by shrinking the candidate pool roughly 16-fold relative to NGG. Cas12a (Cpf1) orthologs use a T-rich PAM (5'-TTTV-3') positioned 5' of the protospacer rather than 3', and generate staggered rather than blunt cuts, which alters downstream off-target repair signatures.

The scanning step is implemented as an exact-match search across both genomic strands using an FM-index (Burrows-Wheeler transform) or suffix array, as in Cas-OFFinder and the CRISPOR web tool. On GRCh38, NGG motifs occur roughly every 8 base pairs by chance alone (2 fixed positions in 3, so 1/4 probability per position pair, adjusted for GC content), yielding approximately 2.1 million PAM-adjacent candidate windows genome-wide for SpCas9 — before any spacer sequence is compared. This unfiltered candidate set is the universe from which the homology search in Stage 2 draws matches; scan quality here (correct reference build, inclusion of alternate haplotypes and structural variant contigs, masking of low-complexity repeat regions) directly bounds the completeness of every downstream prediction. Tools differ in reference handling: CRISPOR defaults to the primary GRCh38 assembly, while more conservative clinical pipelines additionally scan the T2T-CHM13 telomere-to-telomere assembly to capture off-targets in previously unmapped segmental duplications and centromeric satellite arrays that are recurrently implicated in false-negative off-target calls.

Cas-OFFinder and CRISPOR — Seed-Weighted Genome-Wide Alignment

With the PAM-adjacent candidate universe defined, the spacer sequence is aligned against every candidate 20mer to identify sites of sufficient homology to support Cas9 binding and cleavage. Because Cas9 tolerates mismatches non-uniformly across the 20nt protospacer — with far less tolerance in the PAM-proximal "seed" region — a naive global edit-distance search both over- and under-predicts true off-target risk, motivating seed-aware alignment algorithms.

• 10–12 nt: Seed region length (PAM-proximal, low mismatch tolerance)
• 0–6: Max mismatches modeled (configurable in Cas-OFFinder)
• DNA + RNA: Bulge types allowed (1–2 nt insertions/deletions)
• ~347: Candidates after filter (from 2.1M PAM sites)

The seed model, bulges, and GPU-accelerated exhaustive search

Hsu et al. (Nat. Biotechnol. 2013) first established empirically that Cas9 tolerates mismatches poorly within the ~10–12nt of the protospacer proximal to the PAM (positions 9–20 counting from the PAM, in common convention) but permits multiple mismatches in the distal 8–10nt without abolishing cleavage. A single mismatch at PAM-proximal position 1 or 2 essentially eliminates cutting, whereas the same mismatch at position 18–20 (furthest from the PAM) is frequently tolerated with only modest reduction in activity. This asymmetric tolerance is a direct consequence of the RNA-DNA heteroduplex propagation mechanism: R-loop formation nucleates at the PAM and zippers 3'→5' through the seed before the distal end is even tested, so seed mismatches block R-loop nucleation entirely while distal mismatches merely destabilize an already-formed duplex.

Cas-OFFinder (Bae et al. 2014) implements an exhaustive, GPU-accelerated (OpenCL) search that enumerates every genomic locus within a user-specified total mismatch budget (typically 3–6) and, critically, also models bulges: a DNA bulge occurs when the genomic target contains an extra nucleotide relative to the spacer (causing a one-base loop-out of genomic DNA), and an RNA bulge occurs when the spacer contains an extra nucleotide relative to the genomic target (looping out the gRNA). Both bulge types were shown by Lin et al. (2014) and later systematically by CIRCLE-seq data to account for a meaningful fraction of bona fide off-target cleavage events that pure mismatch-based searches miss entirely — in some validated guide panels, 1–2nt bulge-mediated off-targets represented 15–20% of all GUIDE-seq-confirmed sites. CRISPOR layers a seed-aware two-stage filter on top of the raw Cas-OFFinder output: an initial coarse search allows up to 4 total mismatches anywhere in the 20mer plus PAM, followed by re-ranking that up-weights sites where mismatches cluster distally versus in the seed. For a typical well-designed 20nt spacer against GRCh38 with 4 mismatches allowed and 1 bulge, this two-stage search typically returns on the order of 200–500 raw candidate off-target loci, which then proceed to quantitative scoring in Stage 3.

Cutting Frequency Determination — Quantitatively Ranking Off-Target Cleavage Risk

Raw homology hits are not equally dangerous: a distal-mismatch site with a favorable PAM may cleave more efficiently than a seed-adjacent mismatch site with a suboptimal PAM. The Cutting Frequency Determination (CFD) score, derived by Doench et al. (Nat. Biotechnol. 2016) from a large-scale GUIDE-seq and biochemical cleavage dataset, converts each mismatch position/identity combination and PAM strength into a single multiplicative score between 0 and 1 that correlates strongly with observed cleavage frequency.

• 0.0 – 1.0: CFD score range (1.0 = perfect match, unweighted)
• ~19: Sites CFD >0.2 (flagged as high-risk from 347 candidates)
• r≈0.7: MIT score correlation (vs. empirical GUIDE-seq read counts)
• NGG > NAG > NGA: PAM strength weighting (multiplicative PAM penalty term)

The CFD matrix, MIT score, and modern ensemble/deep-learning scorers

The CFD score is computed as a product of per-position penalty terms: for each of the 20 spacer positions, a lookup table gives the multiplicative penalty for every possible mismatched-base pairing at that position (e.g., an rG:dA mismatch at position 14 has a very different penalty than an rC:dT mismatch at position 3), derived empirically from a saturating mismatch library assayed by deep sequencing of indel frequency at each target. These 20 position-specific penalties are multiplied together, then multiplied again by a PAM-strength term (canonical NGG scores 1.0; NAG scores roughly 0.3; other near-PAM sequences score lower still), yielding a final CFD score where 1.0 represents a perfect-match, canonical-PAM site and scores approaching 0 represent sites vanishingly unlikely to be cleaved. Empirically, sites with CFD >0.2 correlate with a substantially elevated probability of detectable GUIDE-seq tag integration, and most clinical guide-selection pipelines (e.g., those used in ex vivo CAR-T and hemoglobinopathy editing programs) flag any predicted off-target above this threshold for mandatory experimental follow-up.

The earlier MIT specificity score (Hsu et al. 2013, implemented in the original crispr.mit.edu tool) uses a related but distinct formulation: a position-specific mismatch weight vector combined with a term that penalizes closely spaced mismatches (since consecutive or near-consecutive mismatches destabilize the heteroduplex more than the same number of mismatches spread across the spacer) and a term for total mismatch count. The MIT score is typically reported per-guide as an aggregate 0–100 "specificity score" that integrates over all predicted off-targets genome-wide — a guide with many low-CFD off-targets scores lower in aggregate than one with few, even if no single off-target is individually alarming. Modern successors incorporate machine learning directly: Elevation (Listgarten et al. 2018) trains a gradient-boosted ensemble on expanded biochemical and cellular cleavage datasets and explicitly models bulges; DeepCRISPR (Chuai et al. 2018) and CRISPR-Net use convolutional/recurrent architectures trained end-to-end on GUIDE-seq and CIRCLE-seq compendia, capturing higher-order sequence context (dinucleotide and trinucleotide effects) that simple position-independent matrices cannot represent. Across independent benchmarks, CFD remains the field-standard baseline embedded in CRISPOR, Benchling, and most clinical guide-design pipelines, with deep learning scorers used as a secondary cross-check.

Epigenetic Context — Why Sequence Similarity Alone Overstates Real-World Risk

A perfect sequence match embedded in constitutive heterochromatin may never be cleaved in the relevant cell type, because Cas9 must displace or navigate nucleosomes and higher-order chromatin compaction to access DNA, exactly as transcription factors and other DNA-binding proteins do. Integrating cell-type-specific chromatin accessibility data — ATAC-seq, DNase I hypersensitivity (DHS), and histone modification ChIP-seq — into the ranking pipeline substantially improves the precision of the final candidate list by down-weighting sequence-plausible but functionally inaccessible sites.

• ~2–3%: Genome in open chromatin (typical ATAC-seq peak coverage)
• 19: Candidates after chromatin filter (from 347 sequence-based candidates)
• ~3–5×: Cleavage enrichment, open vs closed (per Verkuijl & Rots 2019 meta-analysis)
• >1,600: ENCODE cell types indexed (ATAC-seq/DNase-seq tracks)

Nucleosome occlusion, cell-type specificity, and integrated risk models

In vitro biochemical cleavage assays performed on naked, protein-free DNA systematically overpredict off-target activity relative to what is observed in living cells, because chromatinized DNA presents a fundamentally different substrate. Nucleosome-wrapped DNA sterically restricts R-loop formation: Isaac et al. (2016) and subsequent single-molecule studies showed that Cas9 binding and cleavage efficiency drop by an order of magnitude or more when a target site is positioned at the nucleosome dyad versus linker DNA, and that even accessible-looking sites can be occluded by higher-order 30nm fiber compaction in heterochromatic domains. Because chromatin state is cell-type and even cell-state specific (a site open in HEK293T may be closed in primary hematopoietic stem cells, the actual therapeutic target for many ex vivo gene-editing programs), off-target prediction pipelines increasingly require ATAC-seq or DNase-seq data generated in the specific target cell type — not a generic reference — to be biologically meaningful.

Practically, this stage cross-references each CFD/MIT-ranked candidate coordinate against ENCODE or user-generated ATAC-seq peak calls (typically MACS2-called peaks with q<0.05) and DNase I hypersensitivity tracks, along with repressive histone marks such as H3K9me3 and H3K27me3 that flag constitutive and facultative heterochromatin respectively. A composite accessibility score is computed, often as a simple binary in/out-of-peak flag or a continuous signal-fold-enrichment value, and combined with the CFD score — either as a hard filter (discard closed-chromatin sites regardless of CFD) or as a multiplicative risk-reweighting term in more sophisticated pipelines such as CRISPOR's "Cutting Frequency" chromatin overlay or the CRISPRitz tool's epigenetic annotation module. Applied to a typical 347-candidate off-target list, chromatin filtering combined with the CFD threshold from Stage 3 commonly collapses the actionable list to on the order of 15–25 loci meriting empirical follow-up — a roughly 15- to 20-fold reduction in wet-lab validation burden relative to testing every sequence-homologous hit, which is the difference between a tractable and an intractable experimental campaign for a clinical guide-selection program.

GUIDE-seq, CIRCLE-seq, and DISCOVER-seq — Ground-Truthing the Computational Predictions

No in silico model, however well-calibrated, substitutes for unbiased empirical detection of genuine double-strand breaks in the relevant genomic and cellular context. Tag-integration and biochemical cleavage assays — GUIDE-seq, CIRCLE-seq, and the more recent in vivo DISCOVER-seq — provide genome-wide, largely hypothesis-free maps of actual Cas9 cutting, which are used both to validate individual guides before clinical use and to continuously retrain the position-weight and machine-learning scoring models described in Stage 3.

• ~0.1% indel: GUIDE-seq detection limit (tag integration frequency (Tsai 2015))
• in vitro, cell-free: CIRCLE-seq sensitivity (detects sites missed in cells)
• ~30–60%: Predicted-vs-validated overlap (typical CFD top-20 hit rate)
• 6 of 19: Confirmed off-targets (this guide) (GUIDE-seq, HEK293 + primary CD34+)

Orthogonal unbiased assays and closing the computational-experimental loop

GUIDE-seq (Tsai et al., Nat. Biotechnol. 2015) works by co-delivering Cas9/gRNA ribonucleoprotein with a short double-stranded oligodeoxynucleotide (dsODN) that gets non-homologously integrated at genomic double-strand break sites by the cell's own NHEJ machinery. Genome-wide sequencing of dsODN integration junctions then maps every detectable cut site in living cells at read-count sensitivity down to roughly 0.1% modification frequency, with no prior assumption about where off-targets should occur — making it a true unbiased ground truth rather than a hypothesis-confirmation assay. Because it requires transfectable, dividing cells and reasonably efficient dsODN uptake, GUIDE-seq is typically run in HEK293T or K562 lines first, then repeated in the actual therapeutic target cell type (e.g., primary CD34+ hematopoietic stem and progenitor cells for a hemoglobinopathy program) since chromatin state and off-target spectrum can differ meaningfully between cell types, exactly as flagged in Stage 4.

CIRCLE-seq (Tsai et al., Nat. Methods 2017) instead performs the cleavage reaction entirely in vitro on circularized, sheared genomic DNA, exploiting Cas9's in vitro cleavage activity followed by adapter ligation and sequencing of the cut ends. Because it is cell-free, CIRCLE-seq is not limited by cellular chromatin accessibility or transfection efficiency and can be scaled to test many guides in parallel cheaply, but for that same reason it systematically over-calls off-targets relative to what actually occurs in intact chromatin — making it most useful as a maximal-risk candidate generator to be cross-filtered against cell-based chromatin data (Stage 4) and cell-based GUIDE-seq confirmation. DISCOVER-seq (Wienert et al., Science 2019) takes a third approach, using ChIP-seq against the DNA damage response factor MRE11 recruited to endogenous double-strand breaks, which permits off-target mapping directly in vivo (including in animal models and, in principle, patient tissue) without requiring exogenous tag delivery.

In practice, guide-selection pipelines for therapeutic editing programs run all three assays in combination and reconcile them against the CFD/chromatin-filtered computational shortlist: sites confirmed by two or more orthogonal methods are treated as high-confidence genuine off-targets requiring either guide redesign, high-fidelity Cas9 variant substitution (eSpCas9, SpCas9-HF1, HiFi Cas9), or explicit clinical risk assessment (e.g., targeted deep sequencing of that locus in the final cell product before infusion).

In the original Tsai et al. 2015 GUIDE-seq study, the widely used EMX1-targeting guide (the sequence illustrated in this simulation's canvas) was predicted by early MIT-score methods to have very few concerning off-targets, yet GUIDE-seq empirically identified several bona fide off-target sites in HEK293 cells — including one within an intron of a neighboring gene at approximately 15% of the on-target tag-integration frequency — that had been missed or under-ranked by sequence-similarity search alone. The finding was foundational in establishing that unbiased genome-wide empirical assays are not optional confirmatory steps but a required component of any rigorous off-target risk assessment, particularly before advancing a guide into a therapeutic editing program.