Analytical Technologies for the Detection and Identification of Genome-Edited Crops: Current Capabilities, Product-Specific Feasibility, and Enforcement Readiness

Abstract

Genome-edited (GE) crops are reaching consumers faster than the analytical infrastructure designed to monitor them. Unlike transgenic crops, most GE products carry only small sequence changes and no foreign DNA, making conventional element-based polymerase chain reaction (PCR) screening, which has underpinned genetically modified organism (GMO) enforcement for two decades, largely ineffective. This review critically evaluates detection technologies not by listing them sequentially, but by comparing their performance against a shared set of enforcement-relevant criteria: sensitivity at regulatory thresholds, allele discrimination capacity, prior target knowledge requirement, and validation maturity. Building on detection/discrimination distinctions already present in ENGL guidance documents and the DETECT project, we formalize a two-axis framework separating detectability (technically achievable for most known targets in defined seed or DNA mixtures at or near the 0.1% MRPL) from identifiability (rarely achievable without developer disclosure), with detection as a necessary precondition for identification, and apply it product by product to each commercialized GE crop for which public molecular data are available. The sulfonylurea-tolerant canola (SU Canola) case, in which analytical specificity is established but forensic event specificity is contested, and the German DETECT project are examined as contrasting case studies of analytical success and attribution failure, extracting generalizable lessons for the field. A technology comparison table, a product-specific feasibility matrix, and a tiered enforcement workflow are provided as practical tools. We conclude with five research priorities for closing the detection–identification gap across near-term, mid-term and longer-term horizons.

1. Introduction: Detection, Identification, and the Layers Between

The premise of genome-edited (GE) crop regulation rests on an assumption that has gone largely unexamined: that detecting an edited allele is equivalent to identifying an engineered product. This review argues that it is not, and that the gap between these two capabilities defines the central challenge for enforcement laboratories worldwide.

1.1. Detection–Identification Framework

Three layers of identification need to be distinguished at the outset, because conflating them produces avoidable disagreement about what is analytically achievable. Molecular (allele-level) identification confirms the precise sequence change at a defined locus and is solved for most known targets. Product-level identification attributes the detected allele to a specific commercialized GE line rather than to an independently arising natural or mutagenic variant. Legal or forensic attribution adds the evidentiary sufficiency required for regulatory or judicial decisions in a particular jurisdiction.

Regulatory frameworks operate at different layers. EU official controls under Case C-528/16 [1] and the proposed new genomic techniques (NGTs) regulation [2] require closer to product-level identification for unauthorized material. US APHIS verification under the SECURE rule [3] operates closer to the molecular layer combined with pathway compliance. The Cartagena Protocol’s documentation regime for transboundary movements of living modified organisms [4] sits between these two, often invoking “may contain” provisions. References to the “detection–identification gap” throughout this review concern principally the product-level and forensic layers unless explicitly qualified.

We refer to the analytical structure underlying this gap as the Detectability–Identifiability (D–I) framework. Detection and identification are treated as two distinct, hierarchically ordered capabilities rather than degrees of a single difficulty scale: detection is a necessary precondition for identification, because identifiability can be assessed only on material that has already been detected and sequenced, but passing the detection axis carries no guarantee of passing the identification axis.

The detection axis is an analytical-chemistry property: whether a validated assay can quantify the edited allele at the regulatory threshold (the 0.1% minimum required performance limit (MRPL) of EU Regulation 619/2011 [5]) under ENGL Minimum Performance Requirements (MPRs) [6,7]. Four criteria define it: D1 (limit of detection (LOD) and limit of quantification (LOQ) at the relevant matrix), D2 (locus-level specificity against paralogs/homeologs), D3 (matrix robustness) and D4 (inter-laboratory reproducibility). D1–D3 are method-level properties of a developed assay; D4 captures inter-laboratory validation status (deployment readiness) rather than chemistry feasibility itself, and is currently unmet by any commercialized GE crop because no published ring trial exists. Statements throughout this review that detection is “tractable” or “chemistry-solved” for known targets refer to D1–D3 being met; D4 is the validation step addressed in Section 7.

The identification axis is a population-genetic and forensic property: whether the detected variant can be attributed to the suspect engineering event rather than to natural diversity, somaclonal variation, or chemical mutagenesis. Four criteria define it. I1 captures allele uniqueness: the edited allele must be absent from surveyed natural and induced-mutagenic diversity, must carry a signature whose independent spontaneous or mutagenic origin is implausible (multi-bp indel or novel insertion rather than single-base substitution), and must not lie at a residue under recurrent mutagenic or selection-driven independent occurrence (the Type I residue list defined in Section 5.2); all three conjuncts must hold for I1 to be satisfied. I2 is the presence of an informative flanking haplotype within the species’ linkage disequilibrium (LD) window. I3 is the availability of a curated reference panel. I4 is disclosure of the edited sequence and parental genotype.

To avoid ambiguity with the three identification layers introduced above, the identification axis as used in this review covers the product-level and forensic layers; the molecular (allele-level) layer is subsumed under the detection axis, since molecular identification at a defined locus is achieved by the same locus-targeted chemistry that establishes detection. I4 underpins both axes: the disclosed variant lets the detection chemistry be designed (the D side), and the disclosed parental genotype enables haplotype-based attribution (the I side). We group I4 with the identification criteria for continuity with the round-1 reviewer mapping but do not include it in the identification pass condition (which requires I1, or I2 and I3 jointly).

A product passes the detection-chemistry axis when D1–D3 are jointly met; D4 (inter-laboratory ring-trial validation) is a deployment-readiness gate, currently unmet by every commercialized GE crop and tracked separately (Section 7). A product is identifiable at the product level when the allele-uniqueness criterion (I1) is met, or when both haplotype criteria (I2 and I3) are met together. Because I1 requires implausibility of independent origin, a single-base substitution or very short indel does not satisfy I1 based on database absence alone; such an edit can pass the identification axis only through the I2/I3 haplotype route. The two axes are non-equivalent rather than independent: passing the detection axis does not entail passing the identification axis, and identification is conditional on, not independent of, prior detection. Products whose edited sequence has not been disclosed cannot be assessed on the identification axis until locus-targeted detection (Section 4.2, strategy 2) has supplied a candidate variant to evaluate. We apply the framework to every technology (Section 3), every commercialized product (Section 5), and every workflow tier (Section 6).

To anchor these definitions, three contrasting profiles bracket what the framework predicts in practice. SU Canola is highly detectable (0.05% LOQ) yet identification fails because the same single-nucleotide signature arises independently via tissue-culture-induced somaclonal variation. DETECT-rapeseed is both detectable and identifiable: a 1 bp insertion sits within an informative flanking polymorphism in an outcrossing background. DETECT-barley is the most instructive contrast: the chemistry that detects the rapeseed insertion also detects the barley deletion, but barley’s narrow elite-pool haplotype diversity leaves no informative flanking polymorphism, so identification fails for a biological rather than chemical reason. Proprietary products fail upstream of both axes: the detection chemistry cannot be designed without the disclosed variant. These four profiles correspond to the three recurrent failure modes formalized in Section 5.2 (Type I allele non-uniqueness, Type II haplotype opacity, Type III disclosure-blocked) and are encoded by the tiered triage workflow of Section 6.

This formalization builds on the prior reference-laboratory literature. The European Network of GMO Laboratories (ENGL) Working Group on new mutagenesis techniques [8], the reviews [9,10], and the German DETECT (Detection of Genome-Edited Organisms in Food and Feed) project [11] all separated detection from discrimination/identification, with DETECT supplying the first empirical demonstration that detection without identification is a documented outcome for SDN-1 edits, observed for one of the two SDN-1 lines characterized (barley) while the other (rapeseed) remained identifiable. The present review’s contribution is incremental and integrative, with five elements: (i) two hierarchically ordered axes (detection as a necessary precondition; identification as non-equivalent and conditional on prior detection) with eight explicit criteria (D1–D4 against ENGL MPR [6,7]; I1–I4 against population-genetic and disclosure conditions); (ii) layered identification (molecular/product-level/forensic) mapped to regulatory frameworks; (iii) a three-mode failure typology (Section 5.2); (iv) a four-tier enforcement workflow with an operational footprint (Section 6); and (v) uniform application to every commercialized or notified GE crop with public molecular data (Section 5; Supplementary Table S1). The combined result is a framework that can be applied at the notification stage to predict identifiability before method development begins.

A decade of rapid commercialization has brought a diverse portfolio of GE products to market. In the United States and Canada, sulfonylurea-tolerant canola (hereafter SU Canola), produced by oligonucleotide-directed mutagenesis, has been commercially grown since 2014 [12,13]. High-oleic soybean with transcription activator-like effector nuclease (TALEN)-mediated FAD2 knockouts completed US Food and Drug Administration (FDA) consultation in 2019 [14,15], and CRISPR-edited waxy corn with a 14 bp Wx1 deletion demonstrated superior field performance [16]. In Japan, four genome-edited crop varieties have completed the voluntary notification process, including a high-gamma-aminobutyric acid (GABA) tomato that was the first GE food product sold directly to consumers [17,18].

The earliest regulatory milestone was United States Department of Agriculture Animal and Plant Health Inspection Service (USDA APHIS)’s 2016 confirmation that a clustered regularly interspaced short palindromic repeats (CRISPR)-edited anti-browning mushroom would not be regulated. This notification provided unusually specific molecular detail: 1–14 bp deletions in a polyphenol oxidase gene with no integrated foreign DNA [19,20]. That level of disclosure proved exceptional. Under the US regulatory framework, APHIS exemptions for genome-edited plants require assessment of plant pest risk but do not mandate disclosure of the precise edited sequence or submission of a detection method [21,22].

Consequently, for most subsequently commercialized products the public record describes the trait category and regulatory pathway but not the specific nucleotide changes an enforcement laboratory would need to design an event-specific, validated detection assay. Examples include non-browning romaine lettuce [23] (GreenVenus, now Third Security, LLC), reduced-pungency mustard greens produced by multiplex CRISPR-Cas12a knockout of the myrosinase gene family in Brassica juncea [24] (Pairwise/Bayer), and non-browning avocado [25] (GreenVenus, Third Security, LLC, Radford, VA, USA).

This asymmetry between what developers know and what enforcement laboratories can access is not a peripheral detail. It determines whether analytical methods can be developed, validated, and deployed.

Regulatory frameworks shape disclosure. In the EU, the Court of Justice of the European Union (CJEU) ruling in Case C-528/16 (2018) classifies genome-edited organisms as GMOs requiring analytical methods for official controls [1]. The proposed new genomic techniques (NGTs) regulation (COM(2023) 411) would differentiate between SDN-1 products and those containing more complex modifications [2]. The United States, Argentina, Japan, and China employ product-based or notification-based approaches that can permit commercialization with limited sequence disclosure [3,26]. In the Republic of Korea, GE organisms fall under the Act on Transboundary Movement of Living Modified Organisms [27] (Korea LMO Act) with no separate pathway for transgene-free products; given Korea’s substantial commodity imports from GE-producing countries, the practical urgency for enforcement-capable detection methods is immediate.

Section 1.2 below describes the literature-selection methodology. Subsequent sections apply the D–I framework to the molecular landscape (Section 2), a comparative technology assessment (Section 3), bioinformatics and databases mapped to the workflow (Section 4), a product-by-product feasibility analysis with two case studies (Section 5), a tiered enforcement workflow (Section 6), and research priorities for closing the gap (Section 7).

1.2. Scope and Methodology

This review follows a structured narrative rather than PRISMA systematic protocol. Literature was identified through PubMed, Web of Science, Scopus and Google Scholar (January 2010–April 2026) using genome-editing, detection-chemistry and enforcement keywords; regulatory documents from EU EUR-Lex, USDA APHIS, Health Canada, MAFF, MOTIE, OECD and Codex were included. Eligible sources were peer-reviewed studies, validated reference-laboratory reports and policy documents governing detection requirements; medical-only and transgene-only studies were excluded. Comparative evaluation followed the D–I framework of Section 1.1; product-level entries in Section 5 and Supplementary Table S1 cover commercialized or regulatorily notified GE crops with at least partial public molecular characterization.

2. The Molecular Landscape: What Enforcement Must Detect

To understand why detection chemistry succeeds where attribution logic fails, it is necessary to examine how different editing approaches produce different molecular signatures or, more precisely, how most of them do not produce signatures at all.

2.1. Editing Outcomes Ranked by Analytical Accessibility

The site-directed nuclease (SDN) framework classifies outcomes by the cell’s repair response [28]. In SDN-1, a nuclease creates a double-strand break (DSB) that the cell repairs by non-homologous end joining (NHEJ), producing small insertions or deletions (indels) at the target site. No exogenous template is provided, and the resulting mutations are often indistinguishable from those caused by chemical mutagenesis. In SDN-2, a short oligonucleotide template directs precise base substitutions at the break site, producing sequence changes identical to natural single-nucleotide polymorphisms. Oligonucleotide-directed mutagenesis (ODM), the technique used for SU Canola, achieves a similar outcome without creating a DSB, using a chemically modified oligonucleotide to direct a specific nucleotide change. In SDN-3, a longer donor template inserts a sequence of interest, creating new DNA junctions that can be detected by established event-specific PCR methods [29,30]. SDN-3 is included here for taxonomic completeness of the SDN framework; no genome-edited crop currently commercialized or under voluntary notification uses SDN-3, so the focal regulatory challenge of this review concerns SDN-1, SDN-2, ODM, base editing, and prime editing. ODM exploits endogenous mismatch repair rather than homology-directed repair after a double-strand break; its products therefore lack the flanking deletions occasionally produced by SDN-2 and are, if anything, less distinguishable from natural variation.

Two newer platforms extend this spectrum. Base editors fuse a catalytically impaired Cas protein to a deaminase enzyme, converting specific bases (C to T via cytosine base editors, A to G via adenine base editors) within a narrow editing window (typically 4–8 nt for cytosine base editors and 4–7 nt for adenine base editors, depending on the editor variant) without creating a DSB [31,32]. The resulting changes are transition mutations identical to those arising spontaneously. Prime editors couple a Cas9 nickase to a reverse transcriptase guided by a prime editing guide RNA (pegRNA) that templates the desired change, enabling any point mutation, small insertion (typically ∼40 bp with first-generation pegRNAs in PE2/PE3/PE5 (which differ in nicking strategy rather than insertion size); insertions > 100 bp are accessible through engineered pegRNA designs (epegRNA), paired-pegRNA strategies (twinPE), and serine-integrase coupling (PASTE), some demonstrated in plant systems), or small deletion without DSBs or donor DNA [33,34,35]. Proof-of-concept studies in rice, wheat, and maize indicate imminent entry into regulatory pipelines [34,36].

From a detection standpoint, the critical divide is not between “old” and “new” techniques but between SDN-3, which preserves the established screening-then-confirmation logic through unique insertions, and everything else (SDN-1, SDN-2, ODM, base editing, prime editing), where the modification is small, contains no foreign DNA, and may be indistinguishable from standing genetic variation. The German DETECT feasibility study stated this explicitly: Cas editing does not typically leave method-associated residues, and identification is possible only when a distinguishing linked polymorphism exists in the edited line’s genetic background [11]. This empirical finding, not a theoretical argument, anchors the detection–identification gap.

2.2. A Difficulty Scale for Enforcement

Building on qualitative assessments [9,10], we formalize an enforcement-oriented difficulty scale combining both axes of the D–I framework: the detection axis anchored to ENGL MPR [6,7], and the identification axis to criteria I1 (allele uniqueness) and I2–I3 (haplotype informativity). The scale is applied to every technology (Section 3) and every product (Section 5; Figure 1). Low difficulty: SDN-3 insertions creating unique DNA junctions or incorporating recognizable transgenic elements, for which validated event-specific methods exist or are developable. Moderate: SDN-1/SDN-2 edits producing short indels (≥2 bp) at a known locus, where ddPCR or amplicon sequencing is achievable provided the allele is unique within known diversity [37]. High: Single-base changes (SDN-1, SDN-2, ODM, base editing) that are not unique or occur in paralogous gene-family members, where detection succeeds but attribution requires linked haplotype data or developer disclosure [12,38].

3. Technology Assessment: A Comparative Evaluation

Previous reviews have typically presented detection technologies sequentially. Here, we compare them directly against four enforcement-relevant criteria (

Table 1. Detection technologies scored against four enforcement-relevant criteria. Sensitivity uses the 0.1% MRPL of EU Reg. 619/2011 [5]. Allele discrimination capacity is the ability to resolve single-nucleotide variants within paralogous gene families. Prior target knowledge indicates whether the method can run without developer sequence disclosure. Validation maturity is a five-tier ENGL-MPR scale [6,7]: Fully validated (MPR met with ≥1 ring trial); Partially validated (originating lab + ≥1 independent confirmation); Research-validated (MPR-equivalent in a single project); Triage-grade (screening only); Exploratory. Validation maturity tracks analytical-method status, not routine deployability (treated in Section 6). Numerical scores: Supplementary Table S1 (0–3 rubric).

Technology	Sensitivity (LOD/LOQ)	Allele Discrimination	Prior Target Knowledge	Validation Maturity	Key Ref.
Droplet digital PCR (ddPCR)	0.1% in 1 bp rapeseed insertion; 0.9% lowest mixture tested for 1 bp barley deletion	High (known variants)	Required	Research-validated (DETECT; single project, no published ring trial)	[11,39,40]
LNA-enhanced qPCR	0.05% LOQ (SU Canola); 0.003% optimized	High (allele-specific)	Required	Partially validated (SU Canola only)	[12,38,41]
Amplicon deep sequencing	0.1% in seed mixtures; down to 0.01% in DNA mixtures (DETECT)	Very high (any variant in amplicon)	Required (locus)	Research-validated (DETECT); no enforcement protocol	[11,42,43]
WGS (short-read)	Variable (depth-dependent)	Highest in sequence content; attribution requires flanking polymorphism or parental haplotype	Not strict; parental control needed for attribution	Research-grade; not routine	[11]
Long-read sequencing (PacBio HiFi, ONT R10)	Variable; sub-1% with deep coverage	Very high; reads phase flanking haplotypes (I2–I3)	Not strict; parental control improves attribution	Research-grade (HiFi in DETECT); no enforcement validation	[11,44]
CRISPR diagnostics (DETECTR/SHERLOCK)	Attomolar in non-food research assays; not validated for plant food matrices	Position-dependent; PAM-constrained	Required	Exploratory only	[45,46,47,48,49]
High-resolution melting (HRM)	Moderate (Tm-dependent)	Low (cannot confirm identity)	Required	Triage-grade; not for compliance	[50]

): sensitivity at regulatory thresholds (typically 0.1–0.9% for seed lots), allele discrimination capacity, requirement for prior target knowledge, and validation maturity. This comparative structure reveals that the technologies are more complementary than competitive, and that the critical bottleneck is shared across all of them.

3.1. What Has Succeeded: ddPCR, LNA-qPCR, and Amplicon Sequencing

Three technologies have demonstrated enforcement-relevant sensitivity for known GE targets, and comparing their strengths reveals both a shared achievement and a shared limitation. Digital PCR achieves the highest quantitative precision of any current platform through reaction partitioning into thousands of independent compartments, enabling absolute quantification without external calibration curves [39,40]. The DETECT project exploited this property to validate ddPCR assays for a 1 bp barley deletion and a 1 bp rapeseed insertion, with both assays meeting ENGL Minimum Performance Requirements for analytical sensitivity and robustness [11]. The rapeseed insertion was detected down to 0.1% in seed mixtures, reaching the 0.1% minimum required performance limit (MRPL) established by EU Regulation 619/2011 [5] for unauthorized GM material, the regulatory anchor against which enforcement-relevant sensitivity is defined [30]. For the barley deletion the lowest mixture tested in the DETECT report was 0.9%, which reflects the minimum spike level evaluated rather than the assay limit of detection; the per-assay analytical LOD of the ddPCR barley method has not been published separately.

LNA-enhanced qPCR remains the only technology partially validated for a commercialized GE product. For SU Canola [12], locked nucleic acid modifications in primer design achieve a 0.05% quantification limit for a single G-to-T substitution in AHAS1C (paralogous to AHAS3A, where chemical mutagenesis in Clearfield varieties introduces an independently arising, functionally equivalent substitution at the conserved W574 codon). The method was independently validated by the Austrian Environment Agency. JRC reassessment, however, showed that the assay also detects canola lines carrying the same AHAS1C mutation arising from somaclonal variation during tissue culture: a failure of allele non-uniqueness rather than paralog cross-hybridization [8,38]. A recent SNV-detection survey [41] reports LNA probes reaching 0.003% sensitivity under optimized design, so the operational 0.05% LOQ reflects validation constraints rather than a chemical ceiling, and confirms that the specificity failure [38,51] arises from allele non-uniqueness rather than probe chemistry.

Amplicon deep sequencing combines the locus specificity of PCR with the quantitative depth of next-generation sequencing, and the DETECT project demonstrated its capacity to achieve 0.1% sensitivity in seed mixtures and down to 0.01% in controlled DNA mixtures [11]. Unlike allele-specific PCR, amplicon sequencing captures any variant within the amplicon, with no need to redesign a probe for every possible change; this matters in surveillance scenarios where multiple GE products may target the same gene family. At allele frequencies below 1%, where the platform’s intrinsic error rate overlaps with the biological signal, unique molecular identifier (UMI)-based consensus calling becomes essential for distinguishing true low-frequency variants from sequencing artifacts [42].

What these three successes share is a clear demonstration that sensitivity at enforcement-relevant thresholds is technically solved for known targets across multiple independent platforms and validation contexts. What they equally share is a dependence on prior knowledge of the target: all three require that the enforcement laboratory already knows which locus to interrogate, which allele to look for, and what the wild-type reference sequence is before analysis can begin. This shared limitation is the operational form of the detection–identification gap: the detection axis is chemistry-tractable for known targets (D1–D3); the identification axis is conditionally tractable depending on edit class, mating system, and reference-data availability, and is unresolved in many cases (Section 5).

3.2. What Has Not Succeeded: WGS for Attribution and CRISPR Diagnostics for Enforcement

Whole-genome sequencing provides the most comprehensive molecular view of a sample. The DETECT project used PacBio HiFi WGS to confirm a targeted 1 bp deletion in barley, rule out integrated transgene residues, and survey off-target sites [11], yet identification still failed: no linked polymorphism in the flanking background distinguished the edited line from a hypothetical spontaneous mutant. WGS thus confirms what sequence is present but cannot establish how it came to be there. Practical barriers compound this limitation, including sequencing costs one to two orders of magnitude above qPCR per sample, variant-calling complexity in polyploid crops, reference-genome dependency, and the need for matched parental controls rarely available for trade samples.

Third-generation (long-read) platforms, PacBio HiFi and Oxford Nanopore R10, complement short-read WGS by directly phasing flanking-polymorphism windows within single 10–25 kb (HiFi) or longer (ONT) reads, addressing criteria I2–I3 that short-read assemblies cannot resolve unambiguously [44]. DETECT used PacBio HiFi to characterize the edited barley genome and rule out off-targets [11]; the same platforms underpin plant pan-genome and haplotype-panel construction for Tier 4. The current bottleneck is throughput cost and the absence of validated enforcement-grade pipelines for single-molecule error profiles; long-read sequencing is the most likely near-term route to making criterion I2 operationally available once these fall.

CRISPR-based diagnostic platforms, particularly Cas12a-driven DETECTR and Cas13a-driven SHERLOCK [45], offer programmable, sequence-aware detection with isothermal amplification and lateral-flow readouts that could enable rapid checkpoint screening without laboratory infrastructure [46,47,48]. No CRISPR method has yet been validated for a GE crop in food matrices, and several barriers remain: position-dependent variability in single-nucleotide discrimination [45,49], PAM-site constraints [46], and absence of standardized inter-laboratory protocols or proficiency testing. CRISPR diagnostics remain a promising platform for future known-target field screening but are premature for routine enforcement deployment.

Neither greater depth (WGS) nor greater convenience (CRISPR diagnostics) addresses attribution at its root. The bottleneck is the logical gap between observing a variant and proving it was placed there by a specific engineering process in a specific commercial product.

3.3. Base and Prime Editing: The Sharpest Expression of the Detection–Identification Gap

Base editing and prime editing are the categories in which the detection–identification gap is sharpest, and their imminent entry into crop regulatory pipelines [34,36] makes their enforcement implications worth stating explicitly.

For base-edited crops, the analytical challenge is not sensitivity. LNA-enhanced qPCR can detect C-to-T or A-to-G transitions at sub-0.1% allele frequencies in appropriate sequence contexts [12,41]. The challenge is attribution: the changes are transition mutations indistinguishable in kind from those produced by deamination, oxidative damage, or chemical mutagenesis. The narrow editing window (typically 4–8 nt for CBEs, 4–7 nt for ABEs) creates no insertion, no junction, and no foreign sequence, only a point substitution that existing diversity databases may or may not record as a natural polymorphism. Detection is therefore technically achievable for a known target. Identification is blocked unless either (a) a population-genetics argument can establish the edited variant’s absence from natural diversity, or (b) the developer provides parental genotype information that enables haplotype-based attribution.

For prime-edited crops, the situation is analogous but can be somewhat more favorable when the event introduces a short novel insertion. A short novel insertion (typically tens of bp with first-generation pegRNAs; >100 bp accessible through engineered pegRNA, twinPE, or PASTE designs) not present in any known natural allele of the target gene would provide a detectable signal under amplicon sequencing, and population-level querying could in principle support an absence-from-diversity argument. For prime editing events that introduce only point mutations or small deletions, however, the attribution problem is identical to that of base editing.

The enforcement implication is conditional rather than categorical. For base- and prime-edited crops in narrow-elite-pool species (typically but not always selfing crops; see Section 5.2) producing single-nucleotide changes whose independent spontaneous occurrence cannot be excluded (the structurally blocked regime in Section 5.2), the detection axis remains tractable but the identification axis fails because criterion I1 (allele uniqueness) cannot be satisfied and I2 (informative flanking haplotype) is unavailable; better chemistry will not close this gap. For outcrossing species (rapeseed, rye, maize variable), prime-edit insertions outside known natural alleles, or breeding contexts preserving informative flanking LD, identification is currently impractical rather than structurally impossible and may become tractable as curated haplotype panels (I3) and disclosed parental genotypes (I4) accumulate. Identification feasibility is therefore a joint function of edit class, species mating system, and reference-data availability, not a property of editing technology alone.

3.4. What HRM Teaches About Triage

High-resolution melting (HRM) occupies a distinct niche: it is the cheapest and fastest variant screening tool but has the lowest specificity [50]. Its inability to confirm identity makes it unsuitable for enforcement decisions, but its speed makes it valuable as a first-pass filter to prioritize samples for more expensive confirmation. This triage role, flagging, not concluding, may be the template for how CRISPR diagnostics eventually integrate into enforcement workflows.

• Section 3 Key Takeaways

• Detection sensitivity at or below the 0.1% MRPL is technically achievable for known targets across ddPCR, LNA-qPCR, and amplicon deep sequencing.
• All three platforms require prior knowledge of the target locus and allele; this dependency, not chemistry, is the operational bottleneck.
• WGS adds analytical depth but does not, by itself, deliver attribution; CRISPR diagnostics remain exploratory for food-matrix enforcement.
• For base- and prime-edited products, identification feasibility is conditional on edit class and on species mating system rather than on assay performance.
• HRM is a triage tool, not a compliance tool; CRISPR diagnostics may eventually occupy a similar triage niche.

4. Bioinformatics and Database Infrastructure

Even comprehensive molecular characterization does not, by itself, deliver attribution; the DETECT-barley negative result (Section 5.2) is the clearest demonstration. Enforcement laboratories nevertheless require dedicated computational infrastructure to determine which candidate variants enter the identification pipeline at all (low-frequency detection, Tier 2/3), to gate Tier 3/Tier 4 escalation on the criterion I1 conjuncts, and to assemble haplotype evidence at Tier 4. The bioinformatic methods and database resources surveyed here are components of the tiered workflow defined in Section 6. Section 4.1 (low-frequency variant detection) supports Tiers 2–3 on the detection axis. Section 4.2 covers the identification axis. Strategy 1 (population-level filtering) screens criterion I1 (allele uniqueness) at the Tier 3/Tier 4 boundary: its outcome decides whether the workflow can conclude at Tier 3 or must escalate to Tier 4. Strategy 2 (functional-annotation filtering) supports panel design (Tier 2) and prior probability (Tier 4). Strategy 3 (haplotype fingerprinting) addresses criteria I2–I3 and provides the principal attribution evidence inside Tier 4. Section 4.3 (databases) supplies inputs across all tiers. Table 2 maps every method and database to its tier and D–I axis.

4.1. Computational Requirements for Low-Frequency Variant Detection

Standard variant-calling pipelines require GE-specific modifications. Low-AF variant callers originally developed for cancer-genomics applications (Mutect2, VarScan2, LoFreq, Strelka2 [52,53,54,55]) can be adapted for plant low-frequency variant detection but should be benchmarked against spike-in seed–DNA mixtures at the target AF before deployment in an enforcement workflow. Neural network callers such as DeepVariant [56] and DeepSomatic [57] offer improved performance in repetitive and polyploid plant genomes.

For amplicon sequencing below 1% allele frequency, UMI-based consensus calling is mandatory. A minimum family size of three or more reads per UMI is standard, with five or more recommended for AF estimation at 0.1%; this reduces false positives from PCR jackpotting below 1% at typical sequencing depths [42,43,58].

Subgenome-aware alignment is essential for polyploid crops. PolyCat [59] handles Brassica napus (AACC); HomeoRoq [60] handles hexaploid wheat (AABBDD). Both should be integrated into alignment workflows for SU Canola, waxy wheat, and related targets. Subgenome-specific k-mer filtering provides a practical fallback when dedicated tools are unavailable.

Together, these methods support Tier 2 (targeted detection) and Tier 3 (comprehensive characterization) of the workflow defined in Section 6. UMI consensus calling and low-AF callers operate on Tier 2 amplicon panels; neural callers and subgenome-aware alignment are required for Tier 3 WGS or long-read confirmation in polyploid backgrounds.

• Minimum Bioinformatic Requirements for Enforcement Laboratories

An enforcement-capable amplicon-sequencing pipeline for genome-edited targets should at minimum implement the following:

• Read processing: Adapter and quality trimming with documented parameters; mapping to a versioned, species-appropriate reference (subgenome-aware tools such as PolyCat [59] or HomeoRoq [60] for polyploid crops, or k-mer-based subgenome assignment as a fallback).
• Variant calling: For known editing targets at <1% AF, a low-AF variant caller [52,53,54,55] or DeepSomatic [57], benchmarked against spike-in controls at the target AF before deployment; DeepVariant [56] retained for parental-line and homozygous-control genotyping only.
• Unique molecular identifiers: UMI-based consensus calling for any analysis below 1% AF, with a minimum family size of three reads (five strongly recommended for AF ≤ 0.1%, with empirical justification documented per panel) [42,43,58].
• Empirical error-floor characterization per amplicon and per sequencing platform; the platform-specific noise model is locus-dependent and cannot be ported across instruments.
• Reference data: Locus-specific coordinates and wild-type sequences for each target gene; population-level allele frequency data from at least one diversity panel (SoyBase [61], MaizeGDB [62], 3000 Rice Genomes [63], Wheat 10+ [64]); when available, parental-line genotypes.
• Reporting: Explicit confidence qualifiers on calls below the platform’s empirically determined error floor; auditable parameter records and pipeline versioning consistent with ENGL documentation expectations [6,7].

4.2. Three Strategies for Moving from Detection to Attribution

The first strategy, population-level filtering, leverages public SNP databases and varietal diversity panels (SoyBase [61], MaizeGDB and the Maize HapMap [62], 3000 Rice Genomes [63], Wheat 10+ [64]) to flag variants at known editing-target loci absent from surveyed diversity as stronger candidates for editing origin, recognizing that database absence is not proof of artificial origin, since rare natural variants may be unsampled. This strategy screens criterion I1 at the Tier 3/Tier 4 boundary, and is the canonical statement of the I1 gate that Section 5 and Section 6 and Figure 2 reference rather than restate.

The gate has three conditions, any one of which sends the case from Tier 3 to Tier 4 for haplotype-based comparison between the edited line and a non-edited carrier: (i) the allele is present in the relevant diversity panels; (ii) the allele sits at a residue under recurrent mutagenic or selection-driven independent occurrence (the Type I residue list defined in Section 5.2); or (iii) the allele carries a signature whose independent spontaneous origin cannot be excluded, namely a single-base substitution or a very short indel. Only a variant that meets none of these three conditions, namely absent from those panels and not located at a Type I residue and carrying a signature implausible to arise spontaneously (multi-bp indels, prime-edit insertions outside known alleles), allows Tier 3 to issue a probabilistic attribution conclusion without escalation.

Each of the three conditions corresponds to the failure of one I1 conjunct (Section 1.1): condition (i) to diversity-panel absence, condition (ii) to Type-I-locus exclusion, and condition (iii) to signature implausibility. The gate is therefore exactly the criterion I1 gate, with no precautionary buffer beyond I1. It addresses Type I failures (Section 5.2).

The second strategy, functional-annotation filtering, restricts variant analysis to exonic positions within documented editing-target genes (FAD2, AHAS, ALS, Wx, PPO, GAD, MLO; EU-SAGE [65]), reducing candidate variants by orders of magnitude. It is invoked at Tier 2 (panel design) and Tier 4 (prior probability that a variant is editing-related).

The third and most powerful strategy is haplotype fingerprinting, which compares linked-SNP patterns flanking the edited locus to reference panels for known GE lines and parental varieties. DETECT demonstrated feasibility for rapeseed (informative flanking polymorphisms identified) but reported a critical negative result for barley (no distinguishing linked polymorphism) [11]. Fingerprinting therefore makes identification conditionally possible when the background is informative, but requires curated haplotype panels that do not yet exist at the global scale for any commercialized GE crop. It is the principal source of attribution evidence at Tier 4, addressing criteria I2–I3 and Type II failures (Section 5.2).

A recent proof of concept [66] concretizes this strategy at enforcement-relevant sensitivity, combining targeted high-throughput sequencing with genome-database-mined genetic fingerprints to identify a genome-edited rice line spiked into a food-chain matrix at 0.1%, with applicability noted to be contingent on sufficient public genomic information for the edited line. This demonstrates that haplotype fingerprinting is feasible in principle where sequence disclosure has occurred, and simultaneously marks its current boundary: in the absence of such disclosure, the approach fails by construction.

Table 2. Mapping of bioinformatic methods and database resources (Section 4) to the tiered enforcement workflow (Section 6) and to the Detectability–Identifiability axis each supports. Tier numbers refer to Figure 2: T1, intelligence-guided triage; T2, targeted detection; T3, comprehensive molecular characterization; T4, haplotype-based attribution.

Method/Resource	Section	Primary Tier(s)	Function in Workflow	D–I Axis
Low-AF variant callers, cancer-genomics origin (Mutect2, VarScan2, LoFreq, Strelka2)	Section 4.1	T2–T3	Variant calling on amplicon/WGS data at <1% AF in plant material	Detection
Neural callers (DeepVariant, DeepSomatic)	Section 4.1	T3	Variant calling in repetitive and polyploid regions for first-detection events	Detection
UMI consensus calling (≥3–5 reads/family)	Section 4.1	T2	Suppressing PCR/sequencer error below 0.1% AF in amplicon panels	Detection
Subgenome-aware alignment (PolyCat, HomeoRoq)	Section 4.1	T2–T3	Homeolog-specific read assignment in B. napus (AACC) and hexaploid wheat (AABBDD)	Detection
ddPCR (event-specific)	Section 3.1	T2	Quantitative confirmation at ≤0.1% MRPL for known SDN-3/SDN-1 targets	Detection
LNA-enhanced qPCR	Section 3.1	T2	Allele-specific quantification of known SNV/indel edits	Detection
Amplicon deep sequencing (multiplex panels)	Section 3.1	T2	Locus-level capture of any variant in known editing target genes	Detection
HRM	Section 3.4	T1 (triage flagging)	Cheap pre-screen to prioritize Tier 2; insufficient specificity for compliance	Detection
CRISPR diagnostics (DETECTR/SHERLOCK)	Section 3.2	T1 (future)	Field-deployable known-target screening once validated for food matrices	Detection
WGS/PacBio HiFi long-read	Section 3.2	T3	Confirms allele structure, rules out off-targets, supplies flanking haplotype	Detection + I1/I2
Population-level filtering (SoyBase, 3000RG, MaizeGDB, Wheat 10+)	Section 4.2 (strategy 1)	T3/T4 boundary	Tests allele uniqueness against surveyed natural diversity (criterion I1); gates escalation from Tier 3 to Tier 4	Identification
Functional-annotation filtering (EU-SAGE catalog of editing targets)	Section 4.2 (strategy 2)	T2 + T4	Panel design and prior-probability calculation, restricting variant search to documented editing target genes	Both
Haplotype fingerprinting (proof of concept [66]; DETECT-rapeseed reference)	Section 4.2 (strategy 3)	T4	Probabilistic attribution against curated reference panels (criteria I2 and I3)	Identification
EUginius molecular characterization registry	Section 4.3	T1–T4	Method retrieval (T1–T2), sequence query (T3), event registry (T4)	Both
GMOMETHODS, GMO-Matrix, GMO-Amplicons	Section 4.3	T2	Validated reference methods and screening strategy optimization	Detection
Biosafety Clearing-House and regulatory records	Section 4.3	T1	Origin and regulatory-status gating; intelligence-guided triage	Detection

Table 2. Mapping of bioinformatic methods and database resources (Section 4) to the tiered enforcement workflow (Section 6) and to the Detectability–Identifiability axis each supports. Tier numbers refer to Figure 2: T1, intelligence-guided triage; T2, targeted detection; T3, comprehensive molecular characterization; T4, haplotype-based attribution.

Method/Resource	Section	Primary Tier(s)	Function in Workflow	D–I Axis
Low-AF variant callers, cancer-genomics origin (Mutect2, VarScan2, LoFreq, Strelka2)	Section 4.1	T2–T3	Variant calling on amplicon/WGS data at <1% AF in plant material	Detection
Neural callers (DeepVariant, DeepSomatic)	Section 4.1	T3	Variant calling in repetitive and polyploid regions for first-detection events	Detection
UMI consensus calling (≥3–5 reads/family)	Section 4.1	T2	Suppressing PCR/sequencer error below 0.1% AF in amplicon panels	Detection
Subgenome-aware alignment (PolyCat, HomeoRoq)	Section 4.1	T2–T3	Homeolog-specific read assignment in B. napus (AACC) and hexaploid wheat (AABBDD)	Detection
ddPCR (event-specific)	Section 3.1	T2	Quantitative confirmation at ≤0.1% MRPL for known SDN-3/SDN-1 targets	Detection
LNA-enhanced qPCR	Section 3.1	T2	Allele-specific quantification of known SNV/indel edits	Detection
Amplicon deep sequencing (multiplex panels)	Section 3.1	T2	Locus-level capture of any variant in known editing target genes	Detection
HRM	Section 3.4	T1 (triage flagging)	Cheap pre-screen to prioritize Tier 2; insufficient specificity for compliance	Detection
CRISPR diagnostics (DETECTR/SHERLOCK)	Section 3.2	T1 (future)	Field-deployable known-target screening once validated for food matrices	Detection
WGS/PacBio HiFi long-read	Section 3.2	T3	Confirms allele structure, rules out off-targets, supplies flanking haplotype	Detection + I1/I2
Population-level filtering (SoyBase, 3000RG, MaizeGDB, Wheat 10+)	Section 4.2 (strategy 1)	T3/T4 boundary	Tests allele uniqueness against surveyed natural diversity (criterion I1); gates escalation from Tier 3 to Tier 4	Identification
Functional-annotation filtering (EU-SAGE catalog of editing targets)	Section 4.2 (strategy 2)	T2 + T4	Panel design and prior-probability calculation, restricting variant search to documented editing target genes	Both
Haplotype fingerprinting (proof of concept [66]; DETECT-rapeseed reference)	Section 4.2 (strategy 3)	T4	Probabilistic attribution against curated reference panels (criteria I2 and I3)	Identification
EUginius molecular characterization registry	Section 4.3	T1–T4	Method retrieval (T1–T2), sequence query (T3), event registry (T4)	Both
GMOMETHODS, GMO-Matrix, GMO-Amplicons	Section 4.3	T2	Validated reference methods and screening strategy optimization	Detection
Biosafety Clearing-House and regulatory records	Section 4.3	T1	Origin and regulatory-status gating; intelligence-guided triage	Detection

The rapeseed versus barley contrast in DETECT reflects population-genetic architecture, not database coverage. The underlying biology, namely mating system, ploidy, and LD structure, is treated under the Type II haplotype-opacity mode in the typology of Section 5.2. The general implication is that haplotype-based identification is structurally feasible for outcrossing crops but requires external disclosure or modeling for selfing crops. Applied in combination, these three strategies progressively narrow the space of candidate variants and increase the prior probability that a detected change is editing-related. The output, however, remains a probabilistic assessment rather than a definitive attribution. Without parental genotype information, population-scale allele frequency data at target loci, and curated haplotype panels, none of which is currently available for most commercialized GE products, even the most carefully constructed bioinformatics pipeline cannot close the gap between observing a variant and proving its engineered origin.

4.3. Database Resources and Their Limitations

EUginius provides structured molecular characterization for over 960 GMO events using GMO-GET standardized nomenclature [67,68]. GMOMETHODS curates validated reference methods [69]. JRC GMO-Matrix and GMO-Amplicons support screening strategy optimization [70,71]. EU-SAGE catalogs published editing experiments [65]. The Biosafety Clearing-House provides regulatory decision records. Despite this infrastructure, no centralized database currently provides the combination of edited sequence, parental genotype, and validated detection method for all commercialized GE crops.

This is the informational gap that ultimately limits what any analytical pipeline can deliver, and it corresponds to the disclosure-blocked failures detailed in the typology of Section 5.2; database resources support all four tiers of the workflow as summarized in Table 2.

• Section 4 Key Takeaways

• Low-frequency variant calling for GE enforcement requires UMI-based consensus calling, subgenome-aware alignment for polyploids, and benchmarking of low-AF callers against spike-in mixtures at the target frequency.
• Variant calling at sub-1% AF requires both UMI consensus calling AND empirical error-floor characterization per amplicon and per sequencing platform; the platform-specific noise model is locus-dependent and cannot be ported across instruments.
• Three complementary attribution strategies, population-level filtering, functional-annotation filtering, and haplotype fingerprinting, progressively narrow candidate variants but cannot, in combination, substitute for disclosed parental genotypes and curated haplotype panels.
• Existing databases (EUginius, GMOMETHODS, JRC GMO-Matrix, EU-SAGE, BCH) provide useful infrastructure but do not yet integrate an edited sequence, parental genotype, and validated detection method for any commercialized GE crop.

5. Product-Specific Feasibility: Applying the Detection–Identification Framework

5.1. A Two-Axis Assessment

Applying the D–I framework at the product level, we score every commercialized GE crop with available public molecular data against the three method-level detection criteria (D1–D3, with D4 ring-trial status reported separately) and the four identification criteria (I1–I4). Detectability (chemistry) passes when D1–D3 are jointly met; D4 is reported separately and is presently unmet field-wide (Section 7). Identifiability passes when I1 holds, or when I2 and I3 jointly hold; I4 acts as the practical enabler, supplying the disclosed variant required for an event-specific validated detection method and the parental genotype required for haplotype-based identification. A detailed enforcement readiness assessment scoring each product against seven criteria (mutation disclosure status, method availability, inter-laboratory validation, certified reference material availability, identifiability, enforcement readiness, and key barrier) is provided in Supplementary Table S1, with a product-by-product summary in

Table 3. Detection and identification feasibility for commercialized and regulatorily notified genome-edited crops, assessed along two axes: detectability (whether a validated assay can find the edited allele in a mixed sample at regulatory thresholds) and identifiability (whether the analytical signal can be uniquely attributed to the specific commercial product rather than to an independently arising natural or mutagenic variant). Assessments are based on the public molecular record as of 2026. Identifiability terms follow the I1–I4 criteria of the Detectability–Identifiability framework defined in Section 1.1: Established (allele unique in the sense of criterion I1, namely absent from diversity panels, not located at a Type I locus, and carrying a signature implausible to arise spontaneously, or informative flanking haplotype with curated panel and full disclosure), Achievable (likely to pass with reference-panel curation), Uncertain (cannot be assessed without disclosure or population-scale data), Contested (contradictory evidence on allele uniqueness; analytical performance not in dispute), Not achievable (the I1 path is closed: allele present in surveyed diversity, OR signature plausible to arise spontaneously, OR locus on the Type I residue list, AND no informative flanking haplotype within the species’ LD window), and Not applicable (detection axis itself blocked by proprietary sequence withholding). Products for which molecular details remain proprietary are annotated accordingly. A more detailed scoring against seven enforcement-readiness criteria is provided in Supplementary Table S1. Abbreviations: ODM, oligonucleotide-directed mutagenesis; TALEN, transcription activator-like effector nuclease; LOQ, limit of quantification; ddPCR, droplet digital PCR; WGS, whole-genome sequencing. Reference numbers correspond to the main reference list.

Product	Developer (2026)	Editing Approach	Target Gene	Mutation Disclosed?	Detectability	Identifiability	Overall Assessment	Key Ref.
SU Canola (sulfonylurea-tolerant canola)	Cibus	ODM	AHAS1C (G1676T)	Yes	Demonstrated (0.05% LOQ, LNA-qPCR)	Contested at the forensic layer (Type I; allele non-unique; somaclonal-origin and edited-origin alleles indistinguishable; analytical performance of the assay is not in dispute)	Detection solved; identification unresolved	[12,13,38]
Waxy corn	Corteva (formerly DuPont Pioneer)	CRISPR-Cas9	Wx1 (14 bp deletion)	Yes	Technically developable (amplicon sequencing, ddPCR)	Likely achievable (novel deletion size)	Developable; not yet validated	[16]
High-oleic soybean	Cibus (formerly Calyxt; merger 2023)	TALEN	FAD2-1A, FAD2-1B	Partial	Developed and in-house-validated (real-time PCR for FAD2 variants)	Uncertain (multigene family)	Method developed; independent confirmation pending	[14,15,51]
High-GABA tomato	Sanatech Seed	CRISPR-Cas9	SlGAD3	Yes	Technically developable	Uncertain (depends on allele uniqueness)	Developable; not yet validated	[17,18]
Anti-browning mushroom	Penn State	CRISPR-Cas9	PPO (1–14 bp deletions)	Yes (unusually detailed)	Technically developable	Uncertain	Developable; no commercial detection method	[19,20]
Non-browning romaine lettuce	GreenVenus (Third Security, LLC; formerly Intrexon/Precigen)	CRISPR-Cas9	PPO	No (proprietary)	PPO gene-panel surveillance developable (locus known); event-specific validated method blocked by undisclosed variant	Not applicable until detection supplies a candidate variant	Locus-targeted screening possible; event-specific detection and identification blocked by disclosure gap	[23]
Reduced-pungency mustard greens	Pairwise/Bayer	Multiplex CRISPR-Cas12a	Myrosinase gene family (B. juncea)	No (proprietary)	Myrosinase-family gene-panel surveillance developable (loci known); event-specific validated method blocked by undisclosed variants	Not applicable until detection supplies candidate variants	Locus-targeted screening possible; event-specific detection and identification blocked by disclosure gap	[24]
Non-browning avocado	GreenVenus (Third Security, LLC)	CRISPR-Cas9	PPO	No (proprietary)	PPO gene-panel surveillance developable (locus known); event-specific validated method blocked by undisclosed variant	Not applicable until detection supplies a candidate variant	Locus-targeted screening possible; event-specific detection and identification blocked by disclosure gap	[25]
DETECT barley (research reference)	JKI Germany	CRISPR-Cas9	ENGase (1 bp deletion)	Yes (project disclosure)	Demonstrated (ddPCR tested down to 0.9% lowest mixture, LOD not published; amplicon seq; WGS)	Not achievable (Type II compound; I1 conjunct 2 fails on 1 bp signature, I2 unavailable)	Case study: detection yes, identification no	[11]
DETECT rapeseed (research reference)	JKI Germany	CRISPR-Cas9	RPS5a (1 bp insertion)	Yes (project disclosure)	Demonstrated (ddPCR)	Achievable (informative flanking polymorphisms)	Case study: detection and identification both tractable	[11]

Note: GreenVenus was originally developed under Intrexon Corporation (Germantown, MD, USA; 2019 [23]) and renamed Precigen in 2020; it was divested to Third Security, LLC (Radford, VA, USA) in July 2023 [25]. Calyxt merged with Cellectis and Cibus in 2023 to form the current Cibus, Inc. (San Diego, CA, USA).

. For products where the edited sequence has been publicly disclosed, detection methods are either already demonstrated (SU Canola) or technically developable using existing platforms (waxy corn, high-GABA tomato, high-oleic soybean), yet identifiability remains contested or unresolved in every case. SU Canola is the only product for which a detection method has undergone even partial independent validation, and that method’s event specificity is contested by the very reference laboratories that would need to deploy it. The dispute is forensic, not analytical: it turns on the population-genetic non-uniqueness of the targeted allele rather than on the performance of the assay chemistry. For products where molecular details remain proprietary, such as the GreenVenus non-browning lettuce and avocado (Third Security, LLC; formerly Intrexon/Precigen) and Pairwise/Bayer reduced-pungency mustard greens, the analytical barrier is partial rather than absolute: gene-panel surveillance over the disclosed target locus or gene family (PPO, myrosinase) is developable based on locus knowledge alone, but event-specific validated detection and product-level identification are blocked by the absence of the variant-level sequence information that would allow method validation to begin.

5.2. Two Case Studies That Bracket the Field

The SU Canola and DETECT cases, examined together, bracket the current boundaries of what the field can and cannot deliver.

The SU Canola case represents a scenario in which the detection chemistry has been solved but forensic attribution remains disputed. The LNA-qPCR method [12] achieves a 0.05% quantification limit for a single G-to-T substitution in AHAS1C, a technically demanding target in a multigene family. The Austrian Environment Agency independently validated the method against several ENGL performance criteria. Yet when JRC scientists tested the same method against canola lines carrying the same AHAS1C G1676T mutation arising independently via somaclonal variation during tissue culture, rather than via deliberate editing, the assay could not distinguish between the two origins [38].

The limitation is therefore best understood as forensic and regulatory rather than analytical: the LNA-qPCR chemistry performs as specified, but “event specificity” is a population-genetic property of the targeted allele rather than a property of the assay, and what counts as evidentiary sufficiency may differ across jurisdictions depending on how their enforcement frameworks weigh allele-level confirmation against product-level attribution. The DETECT project approaches the same boundary from the opposite direction: comprehensive characterization was achieved, but identification still failed.

The project validated ddPCR assays for 1 bp edits in barley and rapeseed and demonstrated amplicon-sequencing sensitivity down to 0.01% in DNA mixtures. PacBio HiFi whole-genome sequencing characterized the edited barley genome, confirmed the intended deletion, and ruled out off-target alterations and transgene residues [11]. Despite this analytical depth, the project concluded that identification of the barley mutant was not possible because no linked polymorphism in the surrounding genetic background distinguished it from a hypothetical spontaneous mutant carrying the same deletion. The lesson here is that even the most comprehensive molecular characterization cannot create attribution where the underlying biology does not provide a distinguishing signal; the limitation is not the sequencing platform but the genetic context of the edit itself. The two cases are structurally parallel but analytically distinct. The SU Canola failure is one of allele non-uniqueness: the same AHAS1C G1676T arises independently via somaclonal variation, so no assay targeting only that site can discriminate engineered from non-engineered origin. The DETECT-barley failure is one of genetic-background opacity: the edited allele cannot be linked to a unique background because no informative flanking haplotype exists. These require different remedies, population-scale allele frequency data for SU Canola, curated haplotype reference panels for DETECT-type cases. The two failure modes therefore call for different reference datasets.

Together, the two cases prefigure the failure-mode typology of Section 5.2: SU Canola exemplifies allele non-uniqueness driven by parallel somaclonal origin (Type I), and DETECT-barley exemplifies haplotype opacity from narrow elite-pool diversity (Type II). Identification feasibility is therefore the joint outcome of crop biology, allele uniqueness within surveyed natural and induced-mutagenic variation, and reference-data availability, rather than of editing chemistry alone.

5.2.1. Failure-Mode Typology

The failure modes generalize into three predictive types.

Type I (allele non-uniqueness) affects edits at residues under recurrent selection in mutation-breeding programs, herbicide-resistance loci (AHAS/ALS, EPSPS, HPPD), FAD2, and MLO, where chemical mutagenesis or somaclonal variation generate functionally equivalent alleles. SU Canola is the canonical case. The remedy depends on the source of the parallel allele: when it arises in standing natural diversity, population-scale allele-frequency data from varietal panels is the appropriate remedy; when it arises via tissue-culture-induced somaclonal variation or chemical mutagenesis and is therefore absent from natural-diversity catalogues (the SU Canola situation, where the same AHAS1C G1676T arises during tissue culture independently of editing), the remedy is flanking-haplotype discrimination outside the edited locus, i.e., the Type II remedy applied to a Type I failure. SU Canola is therefore best understood as a Type I + Type II compound case rather than as a pure Type I case.

Type II (haplotype opacity) affects edits in crops whose elite breeding pools carry low haplotype diversity and few informative flanking polymorphisms at the edited locus. Selfing or narrow-base crops (barley, wheat, rice, soybean, common bean) are most affected, but the controlling variable is elite-pool haplotype diversity rather than mating system per se: Brassica napus is partially selfing, yet DETECT-rapeseed retained an informative flanking polymorphism. DETECT-barley is the canonical case. The remedy is either developer-disclosed parental haplotype or high-density varietal panels enabling identity-by-descent inference. The empirical base is thin: only the two DETECT lines (barley and rapeseed) ground the typology, so broader generalization would benefit from additional case studies.

Type III (disclosure-blocked) affects products that would likely become identifiable if the edited sequence and parental genotype were disclosed and locus-targeted detection were performed (consistent with the framework rule that identifiability can be assessed only on detected and sequenced material): GreenVenus non-browning lettuce/avocado, Pairwise/Bayer reduced-pungency mustard greens. The remedy is regulatory disclosure mandates rather than analytical investment.

Types I and II are both I1 failures by different generative mechanisms (allele overlap with diversity vs. lack of informative flanking polymorphism) and can co-occur within a single case; Type III is an I4 (disclosure) failure on the cross-axis precondition. The three labels are therefore descriptive aids rather than strictly disjoint partitions, and the underlying framework variables are the three I1 conjuncts (panel absence, signature implausibility, Type-I-locus exclusion), the I2/I3 haplotype variables, and the I4 disclosure variable. The modes can compound. A hypothetical CRISPR-edited wheat AHAS line is simultaneously Type I and Type II, making identification structurally infeasible by sequence alone and tractable only with disclosure (criterion I4); SU Canola is similarly a Type I + Type II compound case as noted above. Mapping a new product to this typology at the notification stage lets enforcement laboratories predict, before method development, whether validation can plausibly close the identification axis.

5.2.2. Section 5 Key Takeaways

• Across commercialized GE crops, event-specific validated detection requires the disclosed edited variant; for proprietary products with known target loci (PPO, myrosinase) gene-panel surveillance over the locus or gene family remains developable on locus knowledge alone and can capture an undisclosed variant in screening mode, but event-specific attribution is blocked by the disclosure gap.
• Among disclosed products, only SU Canola has a partially independently validated method, and its event specificity is contested at the forensic layer rather than at the analytical layer.
• No certified reference material currently exists for any commercialized GE crop; the gap is structural (developer cooperation, single-nucleotide CRM metrology) rather than purely technical.
• The two DETECT case studies (barley, rapeseed) bracket the field: identification feasibility tracks elite-pool haplotype diversity (typically narrower in selfing/narrow-base crops, but with exceptions, as discussed in Section 5.2) and the availability of informative flanking polymorphisms.
• Three failure modes (Type I allele non-uniqueness, Type II haplotype opacity, Type III disclosure-blocked) can be anticipated at the notification stage; their remedies overlap rather than being strictly distinct (Type I and II are both I1 failures and can co-occur, as in SU Canola; Type III is an I4 disclosure failure on the cross-axis precondition).

5.3. No Certified Reference Materials

No certified reference material (CRM) exists for any commercialized GE crop through Joint Research Centre (JRC-Geel, formerly IRMM) or American Oil Chemists’ Society (AOCS) channels. CRM development is complicated by proprietary restrictions on seed material, the metrological challenge of certifying standards defined by single-nucleotide changes, and the absence of international agreements requiring developers to provide material. The information asymmetry between developers and enforcement laboratories remains the primary structural barrier: developers possess edited sequences, parental genotypes, and internal detection assays but are generally under no obligation to share them. The EU’s proposed NGT regulation may require disclosure for Category 2 plants but would exempt most SDN-1 products [2].

6. A Tiered Enforcement Workflow

We propose a four-tier decision framework (Figure 2) that enforcement laboratories can adopt with currently available resources while explicitly acknowledging where each tier reaches its analytical limit.

Tier definitions at a glance.

• Tier 1: Intelligence-guided triage. Commodity origin, exporting-country regulatory status, traceability documentation, and DNA-quality assessment against ENGL MPR benchmarks. Decides whether a sample warrants analytical work; reagent cost essentially zero; and turnaround in hours.
• Tier 2: Targeted detection. ddPCR, LNA-qPCR, or amplicon deep sequencing against the suspected edited locus, validated to the 0.1% MRPL. Event-specific ddPCR and LNA-qPCR require the disclosed variant; amplicon deep sequencing over the implicated locus or gene family runs on locus knowledge alone and can capture an undisclosed variant in surveillance mode. Low-frequency positives at or near the MRPL are confirmed by ddPCR replication or UMI-based amplicon resequencing; Sanger sequencing is used only on pure-positive control material or strongly enriched amplicons, since its effective detection limit (10–20% variant allele frequency) is well above the 0.1% MRPL.
• Tier 3: Comprehensive molecular characterization. WGS or long-read sequencing for first-detection events, contested results, or suspected unintended integrations. Confirms the allele and rules out off-targets but does not by itself establish product-level attribution.
• Tier 4: Haplotype-based attribution. Comparison of flanking-SNP haplotypes against curated reference panels for known GE lines. Addresses criteria I2 and I3 of the D–I framework; functional only where reference panels exist.

Tier 1 flags samples from countries with active GE commercialization (US, Canada, Japan, Argentina, Brazil, China) for targeted analysis after assessing DNA quality against ENGL MPR benchmarks [6,7]. Tier 2 applies methods matched to the suspected modification type: conventional element screening plus enrichment sequencing for SDN-3 [29], and ddPCR or amplicon deep sequencing with LNA probes for SDN-1/2 with known target loci. Low-frequency positives near the MRPL are confirmed by orthogonal Tier 2 chemistry (ddPCR replicate, or UMI-amplicon resequencing on an independent extract); Sanger sequencing is reserved for pure-positive controls and strongly enriched amplicons because its effective limit of detection (10–20% VAF) sits well above the 0.1% MRPL and cannot resolve low-frequency mixture positives. Under EU Regulation 619/2011, the MRPL for unauthorized GM material is 0.1% mass fraction per ingredient. Tier 2 assays should be validated to this threshold, and any confirmed positive at or above the MRPL constitutes a non-compliance trigger under Regulation (EC) No 1829/2003 [6,72]. Borderline or contested results where measurement uncertainty spans 0.1% are escalated to Tier 3.

Tier 3 (WGS or long-read sequencing) compares the sample to the species reference and, when available, the parental variety. As DETECT demonstrated for barley, this does not automatically confer attribution without a distinguishing linked polymorphism [11]. The Tier 3 exit applies the criterion I1 screen exactly as defined in Section 4.2, strategy 1: a probabilistic attribution conclusion at Tier 3 requires the variant to be absent from the relevant diversity panels, not located at a Type I locus, and to carry a signature implausible to arise spontaneously; if any of these three conditions fails, the workflow escalates to Tier 4 for haplotype-based discrimination. Tier 4 reports findings with explicit confidence qualifiers (“variant detected; flanking haplotype consistent/inconsistent/uninformative for attribution to [product]”). When no curated reference panel exists, the workflow terminates at detection with a finding escalated through regulatory channels.

Operational footprint per tier (indicative 2026 estimates; reagent pricing and labor rates vary across jurisdictions and require periodic recalibration): Tier 1: Documentation review and standard DNA-quality checks; no added reagent cost; and turnaround in hours. Tier 2 (ddPCR or amplicon deep sequencing on existing GMO-surveillance instrumentation): Tens of euros per ddPCR sample and one hundred to several hundred euros per amplicon panel; turnaround 2–5 working days. Tier 3 (WGS or long-read with parental comparison): Several hundred to low thousands of euros per sample and turnaround 1–3 weeks; subgenome-aware bioinformatics for polyploid targets exceeds the staffing of most national control laboratories without dedicated genomics teams. Tier 4 (haplotype fingerprinting): Marginal per-sample cost over Tier 3 once a panel exists; the binding cost is upfront panel construction and curation. Tiers 1–2 are deployable today across ENGL labs; Tier 3 is feasible at regional reference centers or via outsourcing; Tier 4 remains emerging in the absence of public haplotype panels. No proficiency-testing scheme (FAPAS or JRC ring-trial equivalents) yet exists for any GE crop, which is a deployment barrier independent of cost.

The framework is designed to be practical, operationally explicit, and honest about its limits. For each tier it states what is required in instrumentation and staff, what can be accomplished at the corresponding analytical depth, and where conclusions must stop without the disclosed sequences, certified reference materials, and haplotype panels that current enforcement infrastructure does not yet provide for most commercialized GE products.

7. Outlook: Research Priorities and the Feasibility–Achievability Boundary

The detection–identification gap documented in the preceding sections will not close through technology development alone, because the limiting factors are informational and institutional rather than analytical. Closing this gap will require coordinated progress across five research priorities, each operating on a different time horizon. Figure 3 maps these five priorities onto near-term, mid-term, and longer-term horizons and indicates their relative dependencies. The dependencies shown are enabling rather than strictly gating: each priority benefits from but is not hard-gated by the others, except for priority 5 (harmonized standards), whose substance depends on the technical priorities producing the evidence base that international agreements would codify.

Two distinct dependencies need to be separated. Scientific and technical feasibility concerns whether a priority can be delivered given current chemistry, sequencing, bioinformatics, and metrology; this is a question of resourcing, coordination, and time. Political achievability concerns whether the required international agreements and disclosure obligations can be negotiated.

These two dependencies operate through different channels. Mandatory sequence disclosure interacts with the Cartagena Protocol [4], the WTO TBT Agreement (Article 2.2) [73], and the WTO SPS Agreement. CRM development is constrained by developer cooperation under material transfer law. Harmonized evidentiary standards depend on bilateral and plurilateral mutual recognition negotiation. Each priority below is labeled by its dominant dependency, technical, institutional, or political.

The foundational requirement is a global molecular registry for commercialized GE crops, near-term and dominated by political dependencies. EUginius already shows registry feasibility at scale (over 960 events with GMO-GET nomenclature [67]); extension to sequence-level data would supply the target definitions that enforcement laboratories currently lack. The binding constraints are regulatory authority, IP treatment, and data-sharing terms.

Authority could rest on Cartagena Protocol [4] disclosure, domestic pre-market notification, or trade-facilitation-linked voluntary submission. IP is the contested axis. Edited sequences may be in the patent record, but parental genotypes and event nomenclature are typically trade secrets. This calls for stratified access: open metadata; controlled-access sequence data for credentialed enforcement labs; and MTA-gated parental genotypes, analogous to controlled-access genomic-data models. Coordinating bodies could include OECD [74] and FAO Codex [75]. Compatibility with WTO TBT Article 2.2 [73] is an open question any registry must address. A registry alone is insufficient without the physical materials needed to validate detection methods. Certified reference materials (CRMs) for at least three priority commercialized GE crops constitute a mid-term priority, with SU Canola, waxy corn, and high-oleic soybean as logical initial candidates; high-GABA tomato and emerging GE products are extended targets. This priority is dominated by institutional dependencies: CRM metrology for single-nucleotide changes is technically tractable, and the binding constraint is developer cooperation under material transfer agreements. The DETECT project’s well-characterized seed lots and DNA preparations with accompanying SOPs provides a working prototype [11]; JRC-Geel and the American Oil Chemists’ Society (AOCS) are natural starting institutions. Progress would be indicated by internationally distributed CRMs covering multiple commercialized GE products.

The legal scaffolding is non-trivial. Authenticated seed material is generally subject to plant variety protection or contractual seed-use restrictions, and certified DNA preparations may incorporate patent-claimed sequences. CRM production therefore typically requires bilateral MTAs specifying permitted analytical uses, restrictions on redistribution, publication conditions, and indemnity provisions. Where developer cooperation cannot be obtained, regulatory authorities may need to consider compulsory deposition of authenticated material as a condition of pre-market notification, which is a policy rather than an analytical decision. Single-nucleotide CRM metrology is additionally constrained by gravimetric mixing uncertainty, which approaches the 0.1% MRPL near the detection limit and must be characterized per lot.

In parallel with CRM development, validated GE gene panels for commodity crop surveillance constitute a near-term priority dominated by technical and coordination dependencies, and the most directly deliverable of the five. These multiplex amplicon-sequencing assays would cover known editing targets (FAD2, AHAS, Wx, PPO, GAD, MLO, ALS) across soybean, maize, canola, rice, wheat, and tomato. The format is analogous to multiplex marker-assisted selection (MAS) panels and SNP arrays used in plant breeding programs, which simultaneously interrogate many loci in a single workflow. Applied to enforcement, such panels would enable parallel screening of multiple editing loci in a single run, providing a scalable middle ground between single-locus qPCR and untargeted WGS. The priority is technically feasible with existing infrastructure; what is needed is coordinated panel design, inter-laboratory validation, and protocol distribution through ENGL or equivalent networks. Progress would be indicated by a published panel with inter-laboratory proficiency data from multiple ENGL member laboratories.

The attribution challenge requires haplotype reference databases for commercialized GE lines, a longer-term priority dominated by institutional and political dependencies. Construction is technically tractable for outcrossing crops, but deposition by developers and cross-jurisdictional access depend on disclosure agreements. These databases would curate flanking-SNP haplotype patterns for each registered GE product and its parental variety, functioning analogously to simple sequence repeat (SSR) marker and SNP-array reference panels already used for plant cultivar identification under International Union for the Protection of New Varieties of Plants (UPOV) plant breeders’ rights and seed certification systems. Making such profiles queryable through EUginius or a dedicated platform would transform fingerprinting from an ad hoc research exercise (as demonstrated for DETECT-rapeseed) into a routine workflow component. A proof of concept [66] (Section 4.2) demonstrates haplotype-based identification at 0.1% food-chain sensitivity under full disclosure and should inform the reference-panel schema.

IP and data-sharing conditions for haplotype panels differ from those for the edited sequence. The relevant data are phased flanking haplotypes within the species’ LD window, potentially covered by plant variety or breeders’ rights independent of editing-event patents. A workable database needs three elements: (i) a deposition agreement separating the edit-flanking haplotype (deposited) from the broader parental genome (not deposited openly); (ii) an access model that returns match/no-match results to enforcement labs without releasing the reference panel, analogous to the controlled-query interfaces used in seed-certification varietal-identification systems; and (iii) explicit terms governing publication of aggregate haplotype frequencies without disclosing proprietary backgrounds. Progress would be indicated by deposition of haplotype panels for multiple commercialized outcrossing GE crops through EUginius extension or equivalent infrastructure operating under such terms.

Finally, the regulatory ambiguity exposed by the SU Canola case must be resolved through harmonized legal standards distinguishing detection and identification requirements. The SU Canola dispute is one of evidentiary sufficiency at the forensic and regulatory layers rather than of analytical performance, so its resolution is a longer-term priority dominated by political dependencies. The fundamental question is whether enforcement requires detection of the edited allele or attribution to a specific engineering event. The first is feasible for most known targets; the second is often not, without disclosure and population-genetic context. Without clarity on what constitutes adequate analytical evidence, even laboratories with the most advanced platforms cannot produce reports regulators and courts will accept as definitive. Progress would be indicated by agreements among major commodity-trading jurisdictions (EU, US, Japan, China, Korea) on what analytical evidence establishes event identity. The SU Canola dispute has been open since 2014; the timeline depends on political alignment that scientific evidence can inform but not dictate. If these priorities are addressed through sustained international collaboration, the enforcement landscape could be fundamentally transformed. The current situation is one in which only a single GE product has a partially validated and contested detection method, with no certified reference material available. The target state is one in which systematic, risk-based analytical surveillance is possible for the majority of commercially traded genome-edited crops.

8. Conclusions

This review addressed a central question for enforcement in international trade: can genome-edited crops be reliably detected and identified? The evidence supports a two-part answer. First, detection of known edited alleles is analytically tractable.

Across ddPCR, LNA-enhanced qPCR, and amplicon deep sequencing, multiple independent studies have demonstrated sensitivity at or below 0.1% in defined seed or DNA mixtures (D1–D3); no published ring trial has yet established inter-laboratory reproducibility (D4) for any commercialized GE crop, and that gap is addressed in Section 7. This holds even for demanding targets such as single-nucleotide variants within multigene families. Generalization to heterogeneous, real-world samples remains to be demonstrated.

Second, product-level identification, defined as attributing a detected variant to a specific commercial product rather than to an independently arising natural or mutagenic variant, remains unresolved in many but not all cases. The failure mode is conditional rather than uniform. Identification is structurally blocked when single-nucleotide edits in narrow-elite-pool species (typically but not always selfing crops; see Section 5.2) produce a signature whose independent spontaneous occurrence cannot be excluded (so criterion I1 is not satisfied) and no informative flanking haplotype distinguishes the edited line. This is the Type II (haplotype-opacity) failure mode (Section 5.2), exemplified by DETECT genome-edited barley. Identification is currently impractical but conditionally achievable in three settings: when the edit class introduces a novel signature (larger indels, prime-edit insertions outside known alleles); when the species’ mating system preserves extended LD (rapeseed, rye, maize variable); or when curated reference panels and parental genotypes can be assembled. Allele-level molecular identification at known targets is, by contrast, largely tractable. The gap between detection and identification is therefore not purely technical. It reflects a deficit in shared data infrastructure, standardized reference materials, and harmonized evidentiary criteria. The proposed solutions of Section 7 (global sequence registries, certified reference materials, validated gene panels, haplotype-based attribution resources, and aligned regulatory standards) are institutional investments whose implementation will also require continued advances in sequencing chemistry, bioinformatics, and cross-jurisdictional harmonization.

A cautious and context-dependent interpretation is warranted. Genome-edited changes can often be detected with high sensitivity, but product-level identifiability cannot currently be established with comparable confidence, except where crop biology, allele uniqueness, or curated reference datasets provide a discriminating signal. The size of this gap depends on edit class, species mating system, and reference-data availability. Closing it will require both methodological progress and the institutional infrastructure outlined in Section 7, together with clearer articulation of the conditions under which identification is and is not achievable.