Figure 1: Three tiers of CRISPR off-target evidence — in silico prediction, targeted NGS validation of candidate sites, and genome-wide detection — assembled into a defensible evidence package for publication or regulatory review.

Every journal reviewer and regulatory assessor who evaluates a CRISPR-edited crop line asks the same question: were there unintended edits? Answering it well requires more than running a prediction tool and listing the top five hits. It requires a strategy — one that matches the depth of evidence to the stakes of the decision, accounts for the quirks of crop genomes, and correctly interprets results against the background of natural and tissue-culture-induced variation.

This article outlines that strategy. It covers in silico prediction, targeted experimental validation, genome-wide detection, and the assembly of an evidence package that is proportionate to the project's goals. It is the final article in this series and assumes familiarity with the methods covered in CRISPR editing validation, amplicon panel design, and lead line selection.

Why Off-Target Evidence Matters

Off-target editing — Cas9 or Cas12a cleaving at genomic sites other than the intended target — is a real biological phenomenon. Its frequency and distribution depend on the guide RNA sequence, the nuclease variant, the delivery method, and the crop species. The question is not whether off-target editing exists as a theoretical possibility, but whether it occurred at detectable levels in the specific edited line being advanced.

The stakes determine the evidence standard:

The evidence package for a publication does not need to prove the absence of every conceivable off-target edit across the genome. It needs to demonstrate that a systematic search was conducted and that the search did not find evidence of off-target editing at the predicted sites. For regulatory dossiers, the bar is higher, and WGS is the expected approach.

In Silico Prediction: The First Pass

Every off-target analysis starts with a computer. In silico prediction tools search the reference genome for sequences similar to the guide RNA and rank them by the number and position of mismatches.

Tools and Their Limitations in Crops

CRISPOR and Cas-OFFinder are the most widely used tools. Both identify genomic sites with up to four mismatches to the guide RNA sequence plus an NGG PAM (for SpCas9). For a typical 20-nucleotide guide RNA, this returns anywhere from zero to several hundred candidate off-target sites, depending on the guide sequence and the genome's repeat content.

The limitation is that these tools were developed and validated primarily on mammalian genomes. Crop genomes — larger, more repetitive, and often polyploid — pose challenges that mammalian-trained tools handle imperfectly. A guide RNA designed against the wheat A subgenome may have an off-target site in the B or D subgenome that differs by only one or two nucleotides. If the reference genome's subgenome assembly is incomplete or contains phasing errors, the prediction tool may miss it entirely.

A practical approach: run the prediction on every available assembly for the target species, including pan-genome assemblies if they exist. If the crop has a well-characterized variant database (e.g., RiceVarMap, WheatVarDB), check whether the predicted off-target sites overlap with known natural variants — a natural SNP at an off-target site may disrupt the PAM and eliminate the risk of cleavage.

What In Silico Predictions Cannot Do

Prediction tools tell you where Cas9 could cut, not where it did cut. Slaman et al. (2023) tested 89 sgRNAs with 224 predicted off-target sites in tomato protoplasts and found that off-target mutations were detected for only 13 of 89 sgRNAs, and exclusively at sites with one or two mismatches. No off-target mutations were found at any of the 138 sites with three or four mismatches. The computational prediction overestimates the actual off-target landscape — but it is the necessary first step for deciding where to look.

Targeted Validation of Predicted Sites

The next step is to sequence the predicted off-target sites in the edited line and a matched wild-type comparator.

Amplicon NGS of Candidate Sites

For each predicted off-target site, design PCR primers that amplify a 200–350 bp region centered on the predicted cut site — following the same panel design rules described in the amplicon panel design article. Pool the amplicons, barcode each sample, and sequence at 500–1,000x depth.

The output is an allele table for each predicted site, showing whether editing occurred. In most cases, the answer is no — Slaman et al. (2023) found that even for sgRNAs with detectable off-target activity, the off-target editing frequency was typically below 1% at the affected sites.

Interpreting Low-Frequency Signals

When targeted NGS detects a variant at 0.5–2% frequency at a predicted off-target site, determining whether it is a genuine off-target edit or a sequencing artifact requires care. Three checks:

• Is the variant present in the WT comparator? If the same low-frequency variant appears in the wild-type control, it is a natural polymorphism or PCR artifact, not an off-target edit.
• Does the variant match the expected off-target pattern? Off-target editing by Cas9 typically produces small indels centered on the predicted cut site, 3 bp upstream of the PAM. A random SNV 100 bp from the predicted cut site is unlikely to be a genuine off-target edit.
• Is the frequency above the technical noise floor? For amplicon NGS at 1,000x depth, variants below 0.1–0.2% may reflect PCR or sequencing errors. Set a reporting threshold appropriate to the sequencing depth.

Genome-Wide Detection Methods

For lead lines destined for regulatory review, or when the guide RNA has extensive homology to multiple genomic regions, targeted validation of predicted sites may not be sufficient. Genome-wide methods are required.

Whole-Genome Sequencing

Whole-genome sequencing of the edited line and a matched wild-type comparator at 30x coverage is the most comprehensive approach. It detects SNVs, small indels, and larger structural variants across the entire genome. Sretenovic et al. (2023) used this approach to characterize ABE8e-edited tomato lines, sequencing both the edited plants and GFP-expressing control plants regenerated through the same tissue culture pipeline. Both groups carried ~1,200–1,500 SNVs — but with no enrichment of A-to-G mutations or TA motif bias in the edited group, confirming that the observed variation was somaclonal, not editor-induced.

The key to WGS-based off-target analysis is the matched comparator. Without it, every somaclonal variant looks like a potential off-target edit. With it, the analysis becomes a comparison: which variants are unique to the edited line and located near sequences resembling the guide RNA? In practice, very few — often zero — variants meet both criteria.

In Vitro Genome-Wide Methods

When WGS is not feasible, or when a more sensitive off-target search is needed, in vitro methods that enrich for nuclease cleavage sites before sequencing can detect off-target sites present at very low frequencies:

These methods are more sensitive than WGS for detecting low-frequency off-target cleavage but are technically demanding and not yet routine in crop editing projects. For most publication-grade evidence packages, WGS of the edited line and WT comparator at each predicted off-target site is the accepted standard.

The Multiplexing Safety Threshold

An important finding for projects using multiplexed editing: Zhang et al. (2023) performed WGS on rice lines edited with Cas12a at multiple sites simultaneously and found that plants receiving more than 50 simultaneous double-strand breaks exhibited large chromosomal rearrangements, while plants with fewer than 10 DSBs did not. This establishes a practical safety threshold for multiplexed editing — keep the number of simultaneous cuts below 10 per generation, or stagger edits across generations — and highlights why WGS is important for multiplexed projects even when individual guides show clean specificity.

Building a Proportionate Evidence Package

The evidence package should match the stakes. An over-investment in off-target screening for a T0 screening project wastes budget. An under-investment for a regulatory dossier risks rejection.

Publication-Grade Package

For a typical plant journal submission:

1. In silico prediction. Run CRISPOR or Cas-OFFinder on the reference genome. Report all sites with ≤3 mismatches. If the crop genome has a pan-genome assembly, run the prediction on all representative accessions.
2. Targeted validation. Sequence the top 10–50 predicted off-target sites (all sites with ≤2 mismatches, plus the top-ranked sites with 3 mismatches) by targeted amplicon NGS in the edited line and a WT comparator. Report the results as a supplementary table.
3. Statement of limitations. Acknowledge that targeted validation cannot detect off-target editing at unpredicted sites and that WGS was not performed. Most plant journal reviewers accept this for lines that will not enter the food supply.

Regulatory-Grade Package

For a regulatory dossier, the approach described by Cai et al. (2025) in their review of transgene-free editing strategies provides a framework:

1. All of the above (in silico + targeted validation).
2. WGS (30x) of the final lead line and a matched WT comparator that went through the same tissue culture process.
3. Coverage analysis. Report the percentage of the genome covered at ≥10x, ≥20x, and ≥30x depth. Regions with zero coverage cannot be assessed for off-target edits — this is an inherent limitation of short-read WGS and should be disclosed.
4. Structural variant analysis. Check for large deletions, duplications, inversions, and translocations using read-pair and split-read analysis.
5. Integration site characterization. Confirm that no vector backbone or selectable marker sequences remain in the genome.

Movahedi et al. (2023) reviewed the biosafety implications of different CRISPR variants and concluded that high-fidelity Cas9 variants (HiFi Cas9, eSpCas9) combined with careful guide RNA selection substantially reduce the off-target burden and simplify the evidence package.

Figure 2: Evidence packages at two tiers: publication-grade (in silico prediction + targeted NGS) versus regulatory-grade (WGS + structural variant analysis + integration site characterization).

Common Misinterpretations

• "No off-targets detected" means zero off-target edits occurred. It means no off-target edits were detected by the methods used, at the sequencing depth applied, in the tissue sampled. Every detection method has a limit of detection. WGS at 30x cannot reliably detect variants present in less than ~15% of cells. For a chimeric T0 plant where an off-target edit is present in 5% of leaf cells, WGS will likely miss it.
• "The prediction tool found 200 off-target sites, so this guide RNA is unsafe." The prediction tool identifies genomic sites with sequence similarity — places where Cas9 could potentially bind. Slaman et al. (2023) showed that sites with three or four mismatches were not edited at detectable levels in tomato, even when predicted. The number of predicted sites is not itself a safety metric; it determines how many sites need to be checked experimentally.
• "Somaclonal variation is irrelevant — we only care about guide-RNA-dependent off-targets." A reviewer or regulator assessing the safety of an edited line cares about all genetic differences from the parental line, regardless of their origin. Somaclonal variation from tissue culture, if present in a gene of interest, is as biologically relevant as an off-target edit. The matched WT comparator controls for this.
• "Genome-wide off-target analysis is required for every edited plant publication." It is not. Most plant journals accept targeted validation of predicted sites as the off-target evidence standard for research publications. WGS is required for regulatory dossiers and is increasingly expected for high-profile publications, but it has not replaced targeted validation as the minimum standard.

From Evidence to Decision

Off-target evidence is not a binary pass/fail result. It is a body of data that supports a judgment about whether the edited line is suitable for its intended use. A line with no detectable off-target edits at predicted sites, a clean WGS comparison to a matched WT, and a well-designed guide RNA with low in silico off-target prediction burden is as strong a candidate as current technology can produce.

Figure 3: Decision pathway for off-target evidence: match the depth of analysis to the stakes of the project.

For researchers building an off-target evidence package, CD Genomics provides CRISPR Validation Sequencing for Agriculture, which includes targeted amplicon NGS for off-target candidate validation with matched WT comparator analysis. The Targeted Sequencing platform supports custom off-target amplicon panel design for projects requiring systematic screening of 50–500 candidate sites. Bioinformatics Analysis Services provide WGS-based off-target analysis, including SNV/indel calling, structural variant detection, and edited-line vs. WT comparator analysis with publication-ready figures.

For the step-by-step process of advancing a fully characterized edited line from screening to a publication or regulatory data package, the lead line selection and sequencing QC guide covers the full selection pipeline.

FAQ

Q1: How many off-target sites should I validate experimentally?

A: Validate all sites with 1–2 mismatches to the guide RNA (typically 0–20 sites), plus the top 10–30 ranked sites with 3 mismatches. If the guide RNA has more than 50 predicted sites with ≤3 mismatches, consider redesigning the guide — the off-target prediction burden is high enough that experimental validation becomes impractical and a cleaner guide may exist. For publications, 10–50 validated sites is typical.

Q2: Can I skip off-target analysis if I used a high-fidelity Cas9 variant?

A: High-fidelity variants (HiFi Cas9, eSpCas9) substantially reduce off-target editing but do not eliminate it entirely. For a publication, reviewers will still expect in silico prediction and targeted validation of the top predicted sites, though the list will usually be shorter. For internal line advancement with no regulatory implications, a high-fidelity variant plus in silico screening may be sufficient. For any regulatory submission, the full evidence package is expected regardless of the Cas9 variant used.

Q3: What if I find a genuine off-target edit in my lead line?

A: A genuine off-target edit at a predicted site does not automatically disqualify the line. Assess: (a) What gene is affected? If the off-target edit falls in an intergenic region or a gene with no known function related to the phenotype of interest, it may be acceptable. (b) Can it be segregated away? If the off-target edit is on a different chromosome from the on-target edit, selfing or backcrossing can separate them. (c) Is the off-target edit heterozygous? If so, selfing can produce homozygous null segregants at the off-target locus. If the off-target edit is in a functionally important gene and is linked to the on-target edit, select a different lead line.

Q4: Is GUIDE-seq or CIRCLE-seq worth doing for a plant publication?

A: For most plant publications, no. These methods add experimental complexity and cost without changing the conclusion for a line where targeted NGS of predicted sites already shows no off-target editing. They become valuable when: (a) the guide RNA has unusually high predicted off-target burden and a more sensitive search is warranted, (b) the line is heading for a high-profile publication where reviewers have specifically requested genome-wide off-target evidence, or (c) the project is a regulatory submission and in vitro off-target detection complements the WGS data.

Q5: How should I present off-target results in a manuscript?

A: As a supplementary table with these columns: (1) Off-target site genomic coordinates, (2) sequence with mismatches highlighted, (3) number of mismatches, (4) gene annotation (if any), (5) editing detected in edited line (yes/no and frequency if yes), (6) editing detected in WT comparator, (7) method used (targeted NGS or WGS). Include the in silico prediction parameters (tool, maximum mismatches, PAM sequence, reference genome version) in the Methods section. Deposit raw NGS data for the off-target validation in a public repository along with the on-target data.

References

1. Slaman, E., Lammers, M., Angenent, G. C., & de Maagd, R. A. "High-throughput sgRNA testing reveals rules for Cas9 specificity and DNA repair in tomato cells." Frontiers in Genome Editing, 2023, 5, 1196763. DOI: 10.3389/fgeed.2023.1196763
2. Sretenovic, S., Green, Y., Wu, Y., Cheng, Y., Zhang, T., Van Eck, J., & Qi, Y. "Genome- and transcriptome-wide off-target analyses of a high-efficiency adenine base editor in tomato." Plant Physiology, 2023, 193(1), 291–303. DOI: 10.1093/plphys/kiad347
3. Zhang, Y., Wu, Y., Li, G., Qi, A., Zhang, Y., Zhang, T., & Qi, Y. "Genome-wide investigation of multiplexed CRISPR-Cas12a-mediated editing in rice." The Plant Genome, 2023, 16(2), e20266. DOI: 10.1002/tpg2.20266
4. Movahedi, A., Aghaei-Dargiri, S., Li, H., Zhuge, Q., & Sun, W. "CRISPR Variants for Gene Editing in Plants: Biosafety Risks and Future Directions." International Journal of Molecular Sciences, 2023, 24(22), 16241. DOI: 10.3390/ijms242216241
5. Cai, R., Chai, N., Zhang, J., Tan, J., Liu, Y.-G., Zhu, Q., & Zeng, D. "CRISPR/Cas system-mediated transgene-free or DNA-free genome editing in plants." Theoretical and Applied Genetics, 2025, 138(9), 210. DOI: 10.1007/s00122-025-04990-0

---

This article is for Research Use Only. CD Genomics provides agricultural genomics services for research purposes; it does not provide clinical diagnosis, treatment recommendations, or regulatory approval guarantees.