Editing DNA is only the first technical challenge in a genome-editing therapy. Developers also need to show where editing occurred, whether unintended changes appeared elsewhere, and whether the intended target acquired unexpected rearrangements. As therapies become more precise and sometimes more individualized, safety assessment is increasingly a problem of sequencing, experimental design, and bioinformatics.
In April 2026, the U.S. Food and Drug Administration issued draft guidance focused on next-generation sequencing, or NGS, methods used in nonclinical safety studies for human genome-editing products. The document is nonbinding and was released for comment, but it provides a useful view of the evidence regulators expect developers to consider before and during clinical development.
Off-Target Editing Is Only One Risk
A programmable editor is designed to change a selected DNA sequence. Similar sequences elsewhere in the genome can sometimes be edited unintentionally. Those are off-target changes. Their importance depends on the type of edit, its frequency, the affected cell population, and whether the altered region has a biological function.
Safety assessment also needs to examine the intended site. Editing can produce small insertions or deletions, larger deletions, unexpected insertions, chromosomal rearrangements, or other changes that a narrow assay might miss. FDA’s draft guidance therefore discusses both off-target editing and loss of genome integrity rather than treating one list of predicted sites as the whole problem.
Finding Candidate Sites and Measuring Them Are Different Jobs
One group of methods discovers places in the genome that could be affected. Computational searches can find sequences similar to the intended target, while experimental assays can expose DNA or cells to an editor and identify cleavage or editing events. These methods generate candidate sites.
A second stage measures whether those sites are edited in relevant samples and at what frequency. Targeted deep sequencing can examine many reads at selected locations. Whole-genome or other broad sequencing may reveal classes of changes outside a candidate list, although coverage, sensitivity, cost, and interpretation differ.
No single method is automatically complete. A purely computational screen may miss biology that is not captured by sequence similarity. An experimental discovery assay may use a cell type or exposure condition that differs from the therapy. A targeted confirmation assay only sees the regions it was designed to inspect.
The Biological Model Determines What the Data Mean
Genome editing can be performed outside the body on collected cells and returned to a patient, or delivered directly into tissues. The relevant safety model depends on the product. Cell type, delivery system, editing duration, dose, donor template, and patient genetics can all influence the pattern of changes.
A convenient laboratory cell line may be useful for method development but may not reproduce chromatin state or repair pathways in the therapeutic cell. Developers need to explain why a model is representative and where it is not. This is similar to the validation problem discussed in our overview of AI-assisted drug discovery: a technically impressive result has limited value if the evaluation system does not match the intended use.
Sequencing Depth Does Not Equal Unlimited Sensitivity
Reading a region many times can help detect a rare variant, but the practical limit also depends on sample size, DNA quality, amplification error, library preparation, sequencing error, and the analysis pipeline. A stated limit of detection should come from validation with known controls rather than from read depth alone.
Rare events are especially difficult when only a small number of cells are available. If one potentially harmful edited cell is not sampled, no amount of computation can recover it. Conversely, an apparent low-frequency change may be a technical artifact or a naturally occurring variant. Replicates, positive and negative controls, orthogonal confirmation, and careful background comparison help separate signal from noise.
Bioinformatics Is Part of the Assay
NGS does not produce a direct answer. Software aligns reads to a reference, identifies variants, filters artifacts, estimates frequencies, and annotates possible biological relevance. Each step has parameters and assumptions. Different pipelines can produce different results from the same raw data.
A credible study documents software versions, reference assemblies, thresholds, quality controls, and handling of ambiguous reads. It also validates the pipeline against samples containing known variants of relevant types and frequencies. Reproducibility includes code and configuration, not only the sequencing instrument.
This need becomes more important for larger structural changes. Short-read sequencing can be strong for small variants but may struggle across repeats or complex rearrangements. Long-read sequencing, optical mapping, cytogenetic methods, or other assays may provide complementary evidence depending on the risk.
Patient Diversity Complicates Reference Comparisons
Human genomes differ naturally. A sequence near a target site may vary among patients, creating or removing potential off-target matches. A reference genome cannot represent every haplotype. For a therapy intended for a diverse population, developers need a strategy for evaluating common and clinically relevant variation.
Individualized therapies make this more direct. The editor may be designed for one person’s variant, as described in our article on one-patient CRISPR development. FDA’s 2026 draft explicitly notes that its recommendations can apply to individualized products. A platform may reuse manufacturing and analytical knowledge, but the patient-specific guide and target still require a reasoned safety assessment.
Results Need Biological Interpretation
Not every detected edit has the same consequence. A change in a noncoding region may be harmless, regulatory, or simply poorly understood. A rare alteration in a cancer-related gene may deserve more attention than a more frequent change in a region with no known function. Frequency, cell type, persistence, and clonal expansion all matter.
Bioinformatic annotation can prioritize findings, but it cannot prove that an unfamiliar variant is safe. Developers combine sequencing with functional assays, animal or cellular studies, product characterization, and clinical monitoring. The FDA’s January 2024 final guidance places these data within a broader framework covering product design, manufacturing, nonclinical studies, and clinical-trial planning.
What the Draft Guidance Does Not Mean
The document does not declare a universal sequencing panel or a single acceptable threshold for every therapy. It is draft guidance, not a final binding rule, and the appropriate study depends on the product and proposed clinical use. It also does not imply that sequencing can predict every long-term outcome.
Ordinary consumer DNA tests are not substitutes for this work. Clinical genome-editing safety studies require product-specific samples, validated assays, and regulatory-quality documentation. The article is an explanation of the technology and regulatory direction, not medical advice.
What to Watch Next
Watch for the final form of FDA’s guidance, reference materials that help laboratories compare detection limits, better methods for large structural variants, and shared benchmarks for bioinformatics pipelines. Also watch how regulators handle prior knowledge when many individualized products use the same editing platform.
More realistic laboratory models, including the organ-on-chip systems now entering regulatory workflows, may complement sequencing by showing how edited cells behave in a tissue-like context. The strongest safety package will connect molecular measurements with biological function rather than treating a clean sequencing table as the final answer.
Leave a Reply