Measurement and clinical interpretation of CRISPR off-targets

Abstract

CRISPR genetic therapies are revolutionizing the landscape of preclinical research and clinical studies, providing new potential routes for curative intervention for a range of previously untreatable diseases. As with any therapy, the therapeutic benefits and risks must be weighed against consideration of the disease threat. Genome-related adverse events are an inherent risk of CRISPR genetic therapies, including off-target edits. The perception that CRISPR therapies ought to have near-zero off-targets belies clinical medicine, therapy development and biology, which demonstrate that ‘perfect’ therapeutics do not exist. Given that not all genomic off-target events are equal, we provide a practical framework to evaluate and assess off-target safety based on the tools available today and ones that will be developed in the future. With the comprehensive information and assessment gathered using these guidelines, we aim to streamline the transition of CRISPR therapeutics from bench to bedside.

With any new therapy, there is appropriate concern for potential adverse effects. CRISPR and other genetic therapies raise the possibility of genome-related adverse events. As these therapeutics evolve, there is also a need to develop tools that can predict and detect these off-target events. Researchers, clinicians developing CRISPR therapies and the regulatory bodies overseeing these new therapies aim for safe and effective therapies. Despite efforts to develop off-target-free CRISPR therapies, developers and regulators are likely to be faced with making clinically relevant decisions about both predictable and unpredictable unintended genomic changes resulting from CRISPR therapies. Here, we aim to dispel the notion that CRISPR must be perfect to be clinically viable and to create a clinically relevant framework for evaluating off-targets to facilitate the creation of lifesaving therapies. We employ lessons from other therapeutic modalities and current clinical practice to examine CRISPR genome-related adverse event profiling in the broader context of responsible medical therapy development.

Since the landmark discovery by Doudna, Charpentier and colleagues in 2012 (ref. 1), CRISPR has been transforming the therapeutic landscape, particularly for previously ‘untreatable’ and fatal genetic diseases^2,3. CRISPR gene therapies are used clinically for ex vivo editing^4-7, with the first landmark Food and Drug Administration (FDA) approval of a CRISPR therapy⁵ for sickle cell disease in 2023 and for β-thalassemia in 2024. More recently, in vivo genetic therapies are being tested in clinical trials^8-15. For instance, a customized CRISPR treatment was designed, tested and administered to a neonate with a rare genetic disease within a critical time window before fatal or permanent brain damage¹⁴, a milestone achievement for precision medicine. Precise genetic edits are achieved using a guide RNA that homes to a specific target DNA sequence within the genome, allowing its cognate Cas enzyme to create desired genomic changes, such as cuts, base edits or gene rewriting. It is now increasingly feasible to target nearly any genomic loci and thus design novel therapies that remove or correct disease-causing defects. As the Cas enzyme scans DNA for the perfect match to its guide RNA, it can also create undesired genomic off-target changes. Off-target edits typically occur at sites that share sequence similarity with the target region, but here we use the term to apply to any unintended genomic change, including alterations that begin at the on-target site, such as structural variant changes and chromosomal damage. After a DNA change is induced by CRISPR, the DNA repair process is dictated by the edited cell and varies depending on the cell type², which can introduce additional intended or unwanted changes at the editing site. CRISPR base editors can also incur guide RNA-independent off-target editing events via their deaminase domains^16,17.

A standard practice when developing and examining novel therapies is to define the potential risks and weigh them against the potential benefits^18,19. Due to the nature of gene editing, unintended damage to the genome is a potential side effect, such as chromosomal translocations and truncations, aneuploidy, chromothripsis, rare loss of heterozygosity and small or large inversions and deletions²⁰ (Fig. 1). The main concerns regarding off-target genomic changes are that a single ill-placed off-target change could result in cancer or that an accumulation of genomic off-targets could result in genotoxicity^2,16,17,20. These secondary cancer and genotoxicity side effects are not unique to CRISPR-based genetic therapies but are present for other nongenetic standard-of-care therapies, such as chemotherapy²¹, radiation²¹ and gene replacement therapies^22,23. We do not believe that CRISPR genetic therapies should be held to different benefit–risk ratio considerations throughout the therapeutic development pipeline than these previously developed lifesaving therapeutics, which were at one time experimental. Indeed, delaying potentially lifesaving therapies in an effort to achieve perfection results in patient suffering and death.

Fig. 1 ∣. Genomic on-target and off-target alterations that can occur after CRISPR editing.

CRISPR-induced genomic off-targets may lead to adverse events including secondary cancer, cellular dysfunction or cell death. Through preclinical testing, the risk of each potential unintended genomic alteration is identified and evaluated. Some off-targets may be well tolerated, such as those occurring in genomic regions with no functional impact on coding or regulatory elements, while others could be deleterious, such as cancer-promoting mutations, large structural variants or multiple simultaneous off-target edits leading to cell death. Each off-target must be characterized by type and evaluated for risk to cellular and patient health to appropriately weigh off-target risk against potential benefit of the therapy.

Here, we discuss lessons from the development of other therapeutics, such as available off-target detection methods and the standard practice of patients and their physicians in weighing treatment risks and benefits to guide how we approach, measure and discuss genomic off-targets.

Calculating risk tolerance in relation to disease threat

High risk tolerance for therapies addressing life-threatening diseases

Patients and physicians routinely calibrate their willingness to assume risk of new or experimental therapies to the threat posed by disease. When faced with a life-threatening disease, the risk tolerance of patients and physicians often shifts to include therapies with inherent risks^18,19. This is because the cost of doing nothing (certain death) is part of the calculus. For example, the overwhelming majority of cancer treatments risk secondary cancers due to DNA damage or immune disruption, which is even as high as 20% in some cases, but provide a chance for curative intervention²¹. Indeed, approximately 4–16% of patients receiving newer chimeric antigen receptor T cell therapies, which engineer blood stem cells to target cancer cells, develop a secondary cancer²². Nonetheless, the FDA has recently approved seven T cell gene therapies²². Gene replacement delivery vectors can also incur toxic side effects in addition to secondary cancer risks, such as damage to healthy tissues and cellular dysfunction from immune response²⁴. Lentiviral gene therapies for adrenoleukodystrophy improve bleak survival rates but risk myelodysplasia²³. When chemotherapy, radiation and antiviral medications were first tested in humans, they were often poorly tolerated and toxic. This raises the question of whether radiation or chemotherapy, if discovered in the current regulatory and scientific climate, might have been deemed too toxic to try? The history of experimental medicine necessarily progresses with brave patients and clinicians willing to incur risk to push the boundaries of medicine. The appropriate patient population to trial new CRISPR therapeutics is those for whom the other side of the scale is weighted by death from genetic disease and who understand the known and unknown risks of CRISPR. Even with one or more known genomic off-targets or the risk of secondary cancer, a potentially curative CRISPR therapeutic may be preferred by patients over death.

In our collective clinical experience, we have noticed less risk tolerance in treating late-onset genetic diseases, even if fatal, compared to diseases that manifest from birth. For example, fatal genetic diseases that manifest in late childhood, adulthood or late in life are often met with less aggressive interventions and more fatalism than aggressive diseases apparent from birth. Why does being ‘born sick’ with a fatal or debilitating genetic condition tilt the risk calculus toward intervention compared to acquiring a fatal diagnosis in a ‘born healthy’ patient? We advocate for a paradigm shift in the way we consider patients with severe inherited genetic disorders, starting with a reexamination of the benefit–risk calculus for the development of new genetic therapies targeting diseases across the lifespan. We recommend advancing those CRISPR therapies to clinic with safety profiles that outweigh the risk of fatal genetic disease, even if that profile includes some risk that both patient and physician are willing to incur.

Mid-to-high risk tolerance for therapies targeting life-altering diseases

The benefit–risk calculus is more measured for therapies that are not lifesaving but can instead modify high disease burden. For many decades, various oncologic treatments fell into this category by increasing life expectancy by months or years or offering palliation without eradicating disease. It is possible that emerging CRISPR therapeutics may also tread the path from palliation to remission to cure as CRISPR technology and diagnostic and clinical trial landscapes for genetic diseases evolve together. The benefit–risk ratio in those cases, in which the outcome is ultimately not curative but involves extending or preserving quality of life, is weighed against the substantially impaired quality of life. For CRISPR, such a scenario may result, for example, from intervening when the disease has advanced beyond the ability of a genetic intervention to halt the progression, such as late-stage, metastatic cancer, fulminant dementia or modifying a physiologic pathway that ameliorates but does not cure the disease. In these cases, preserving function or extending life may be desired by patients. Alzheimer’s disease is an apt example with high disease morbidity, mortality and no cure. Recently approved anti-amyloid antibody treatments have become the first disease-modifying treatments available. Clinical effects of anti-amyloid therapies are modest with a mild slowing in disease progression^25,26. As many as one in three patients experience adverse brain bleeding or swelling^25,26. Despite the high rate of adverse events, anti-amyloid therapy is offered to patients even without curative intent. We hold CRISPR therapeutics that have palliative rather than curative intent to a higher safety standard than CRISPR therapeutics targeting imminently deadly diseases with curative intent. Therapeutics that modify disease without curing them should tilt toward safer benefit–risk ratios.

Given that the risks of new therapies are not fully understood until after controlled clinical trials have been performed and their use by broader patient populations after approval, we recommend that novel CRISPR therapies be first tested in patients for whom the risk of disease threat outweighs the risk of unknown adverse events. Thus, until CRISPR genetic therapy safety profiles are well established, we recommend testing in patients with life-threatening or life-altering diseases before proceeding to moderate or low disease threat. When the cost of doing nothing is certain death, patients and their physicians are often willing to assume greater risk, an ethos that continues into the modern age of precision medicine.

Recommendations on testing and interpreting CRISPR-induced genomic off-targets

What data are needed to inform patients, clinicians and regulatory bodies of the genomic safety of a potential new CRISPR therapeutics? The US FDA and the European Medicines Agency provide general recommendations for monitoring genotoxicity in preclinical and clinical studies^27,28, recommending that researchers optimize the genetic engineering system and delivery method ‘to reduce the potential for off-target genome modification, to the extent possible’ (ref. 28). Design factors that can change on-target and off-target rates, such as delivery components, CRISPR reagents and the time window of editing, are an integral part of this analysis (Fig. 2).

Fig. 2 ∣. On-target and off-target screening as integral parts of preclinical and clinical evaluation of CRISPR-based therapies.

On-target and off-target editing rates can be affected by the delivery method, the Cas system, guide RNA sequence, window of editing (usually measured via the half-life of the CRISPR system), cell type and associated repair pathway. CRISPR drug product biodistribution affects potential off-organ effects.

To detect and profile off-targets, the FDA recommends using ‘multiple methods (e.g., in silico, biochemical, cellular-based assays) that include a genome-wide analysis…to reduce bias in identification of potential off-target sites’ (ref. 28). A wide range of possible tools and methods to identify off-targets are available. Others have covered off-target detection methods in detail^{16,17,20,29-31}; here, we aim to briefly summarize current practices in the field for techniques enabling investigational new drugs. The general approach of many preclinical studies is to first nominate the most likely off-targets using a combination of in silico and empirical methods and to then verify which of the nominated events are actually present, and, if so, at what frequency, in a relevant in vitro human cell line model or in primary human cells that have been exposed to the editor in question in representative conditions.

Off-target nomination using in silico and experimental methods

In silico methods (Table 1) are used to computationally predict potential off-target sites based on guide RNA sequence similarity, typically measured against a reference genome. This strategy is most applicable to highly conserved genomic regions but may fail for on-target and off-target genomic sites with variation among humans³². While patients may have meaningful differences not captured in the reference genome, sequencing each patient in the development phase of a therapy intended for large patient populations is also currently impractical. This dichotomy was addressed during the development of the CRISPR genetic therapy, Casgevy, now approved by the FDA. The original in silico predicted off-targets were not observed in cell-based assays⁵, but additional off-targets were observed and validated using a variant-aware in silico off-target prediction tool called CRISPRme that can account for polymorphisms in human populations³² and in vitro cell lines with the single-nucleotide variant that differed from the reference genome.

Table 1 ∣.

On-target and off-target detection methods

Nomination				Empirical verification
In silico	Naked DNA		Cell based	Cell based	Structural variants
Cas-OFFinder33	Digested DNA/chromatin	Library based	Repair product based	Sequencing based	Sequencing based
CRISPOR34	Digenome-seq38, DIG-seq40	ONE-seq39	GUIDE-seq47, GUIDE-seq-2 (ref. 42)	rhAmpSeq48, WGS, cancer mutation panels, ChIP–seq, ATAC-seq, RNA-seq	WGS, LR-PCR, CAST-Seq49
CRISPRme32	Circularized DNA	Tagging based			Other
	CHANGE-seq41,42, CIRCLE-seq43	SITE-Seq37	Double-strand break based		ddPCR, karyotype
			INDUCE-seq45,DISCOVER-Seq46		optical genome mapping50

A non-exhaustive list of on-target and off-target edit detection methods commonly employed in investigational new drug-enabling studies for CRISPR genetic therapies. ATAC-seq, assay for transposase-accessible chromatin using sequencing; CAST-Seq, chromosomal aberration analysis by single targeted linker-mediated PCR sequencing; ChIP–seq, chromatin immunoprecipitation followed by sequencing; ddPCR, droplet digital PCR; LR-PCR, long-range PCR; RNA-seq, RNA sequencing; WGS, whole-genome sequencing.

Typically, guide RNAs are designed to have only one perfect match in the genome: the on-target site. But guide RNAs may have up to thousands of potential binding sites in the genome that tolerate a mismatch between the guide RNA and the off-target binding site. Alignment-based (Cas-OFFinder³³), scoring (CRISPOR³⁴), machine learning (Elevation³⁵), deep learning (DeepCRISPR³⁶) and other in silico methods are used to identify and rank potential off-target sites. However, it is challenging to analyze hundreds to thousands of computationally predicted potential off-targets to find true off-targets, a process that requires empiric testing.

Accurate profiling of human targets must be performed using human samples. Animal models are not suitable for genomic off-target detection due to considerable sequence differences, particularly in noncoding regions, between humans and other species². Naked DNA methods support off-target nomination by providing evidence that CRISPR can alter DNA at a predicted site with methods based on cutting or alteration of naked DNA^37-39, chromatin⁴⁰ or circularized naked DNA^41-43 (Table 1). However, naked DNA methods overestimate the number of true off-targets, sometimes by hundreds or thousands of sites⁴⁴. This is because altering naked DNA is easier than altering DNA in a cell, and random damage of unprotected naked DNA can mimic an off-target. Apparent false positive discovery rates for naked DNA methods have been reported between 27% and 95%^16,29, while cellular methods tend to have less than 20%^16,29. Therefore, testing or validating off-targets in human cellular systems is critical.

Experimental testing to verify off-targets

After in silico and cell-free methods predict potential off-target sites, true on-target and off-target editing rates must be confirmed empirically (Table 1). Cell-based assays can provide empirical nomination and verification with additional information regarding rates of how often these changes occur, albeit in a typically simplified cellular system that may or may not recapitulate the performance of the editor in the human body. Cell-based methods^45,46, such as GUIDE-seq^42,47, which nominate off-target locations, are followed up with additional sequencing-based experiments⁴⁸. To verify off-targets, nominated genomic regions can be amplified by PCR and sequenced to increase sequencing depth and thus detection rates of potential off-target regions. Genomic rearrangements may be detected in cellulo⁴⁹ or with methods independent of sequencing, such as optical genome mapping⁵⁰. Modified detection methods for noncanonical CRISPR systems can detect RNA edits when performing base editing^16,17, changes in chromatin structure for epigenetic editing³¹ and tagging methods for prime editing^51-53.

Cell-based assays are more difficult to optimize, more expensive and currently often use cell types (such as HEK or U2OS cancer cell lines) initially optimized for blood-related disorders (T cell editing)⁴¹ that may not represent other clinical settings. Testing in clinically relevant cell or tissue types may not yet be possible if the cells or tissues are difficult to grow in a laboratory or susceptible to the cytotoxicity of cell-based assays. Current methodologies may also prioritize off-targets that are favored by the repair mechanism of the tested cell, which may differ from the repair outcome of the target tissues in patients.

Ideally, unbiased whole-genome methodologies that have high depth (number of times a base pair is read) and coverage (percentage of the genome that is read) should be used to survey the entire sequenced genome for off-targets. In practice, the major limitation to adopting genome-wide surveys is the cost of sequencing and the bioinformatic tools to quantify changes on a genome-wide scale. However, as the cost of sequencing continues to decrease and the depth and coverage of sequencing increase, it is likely that there will be a move to unbiased whole-genome off-target sequencing methodologies in the future. As our ability to accurately characterize editing outcomes continues to increase through technology and assay advancements, it will be possible to further understand the implications of unintended editing outcomes, such as loss of heterozygosity, large deletions or structural rearrangements, on cellular function. Unbiased detection methods including whole-genome sequencing also have the additional advantage of capturing unpredictable guide RNA-independent off-targets that are not related to the guide RNA sequence, such as edits that might result from guide RNA contamination during manufacturing.

In addition to guidance on somatic off-target risk, the FDA has taken a special interest in germline editing. A clinical hold was placed on Verve Therapeutics’ pioneering in vivo PCSK9-targeted base-editing trial¹⁰ for the treatment of familial hypercholesterolemia when the FDA requested quantification of germline editing. In this trial and in similar trailblazing preclinical studies by Intellia Therapeutics for hereditary angioedema¹² and transthyretin amyloidosis⁹, no editing was detected in treated male nonhuman primate sperm^9,10,12. Although editing was observed in bulk ovarian tissue^9,12, it was not observed in progeny from gene-edited female mice^10,12, confirming a lack of germline transmission of the introduced edits. Lack of germline editing more likely reflects the inability of the drug delivery vehicles, lipid nanoparticles in these cases, to penetrate dense tissue, rather than innate protection from germline editing. As delivery vectors become more potent, the issue of germline editing will remain a concern, and therefore there is as yet no clear guidance on how potential germline editing will impact regulatory approval of future therapies. Germline heritability of CRISPR edits is a sensitive issue given our incomplete understanding of many genomic changes across lifespan and because offspring of the treated individual may incur the risk of an off-target without consenting to the therapy. In cases of novel CRISPR treatments that pose a risk of germline editing, it is our opinion that lifesaving treatments for otherwise untreatable diseases should not be withheld or substantially delayed due to a potential risk of germline editing. Instead, we should employ clinical measures to mitigate the impact of germline mutagenesis, as routinely employed in patients of childbearing age receiving chemotherapy, radiation and teratogenic medications^18,19,21. Clinical mitigation measures include appropriate counseling, options for fertility preservation (such as oocyte, sperm and embryo banking), reproductive planning and appropriate birth control. In contrast to life-threatening diseases, for which we recommend that the risk of germline editing not unduly delay or prevent a lifesaving CRISPR therapy, for low-risk and non-life-threatening diseases, it may be appropriate to demonstrate minimal germline editing risk, even if clearing this hurdle delays the therapy. Furthermore, learning about the potential for germline editing in human patients who have already been treated with in vivo CRISPR therapies would be beneficial. This can readily be achieved by measuring editing events in sperm and donated oocytes from patients who have participated in clinical trials. Such monitoring may become routine as CRISPR editing reagents and delivery vehicles become more potent, increasing the chance that a future editor could penetrate gametes.

Recommendations for evaluating the risks of predicted or known off-targets

As off-target profiling methods continue to advance, we will be able to detect low-frequency or rare off-target events with increased efficiency. Achieving CRISPR-based genetic therapies with zero off-targets may not be possible for many viable therapeutic candidates owing to the tolerance of the Cas enzyme for mismatches between guide RNA and target DNA. In Table 2, we offer a practical framework rooted in the available sequencing technologies for detecting, interpreting and prioritizing genomic off-targets for emerging CRISPR therapeutics.

Table 2 ∣.

Recommendations for genomic off-target identification and evaluation

Question	Assessment strategy	Relevant models
IDENTIFICATION
Characterize on-target edit for undesired genomic alterationsa	Sequencing-based methods (rate, type (including structural variants)a)	Human samples/cells, multiple donorsa (primary, transformed, iPSC-derived cells)
Identify off-target edits via nomination and empirical verification	Comprehensive nomination/prediction (in silico, naked DNA) Verify (cell-based sequencing methods)	Human cells ± human naked DNA Cell type/methods relevant to repair system
Consider patient population genomic variants	Variant-aware models, verify via cell lines with variant	Patient population genetic data Human cellsa
EVALUATION
Characterization
Potency/persistence: is the edit probable at clinical doses/duration of editor expression?a	Characterize edit types Test at supersaturated and physiological doses	Relevant human cells or animal modelb Human cell types for edit characterization
Are there any effects on DNA structure?	Structural variant detection methods	Human cells
Cancer risk
Is the edit in or near a known cancer-causing gene?a	Catalogue of Somatic Mutations56, increase sequencing depth Tolerate no known oncogene/tumor-suppressor off-targets	Human cells ± human naked DNA
Functional risk
Does the edit affect a genomic site likely to have functional impact?	Evaluate coding vs. noncoding Ensembl Variant Effect Predictor57 Combined Annotation-Dependent Depletion score58	Relevant human cells in vitro Relevant in vivo animal modelb
What regions of the body receive nontargeted editing? At what frequency?	Ex vivo: labeled cells In vivo: sensitive/quantitative methods	Animal modelb
Does an on-target edit in a nontargeted tissue affect function?	Cellular function changes (e.g., RNA/protein changes)	Relevant human cells in vitro In vivo humanized model or animal modelb
Additional considerations
Does the delivery system give rise to genomic integration of additional material?	Test for integrations (WGS, amplicon sequencing, etc.)	Human cells
What long-term off-target effects are present?	Extended post-clinical trial monitoring	Monitor in humans

^aFDA concurs

^bAnimal model with relevant guide homology or humanized animal model. iPSC, induced pluripotent stem cell.

Ideally, effective on-target guide RNAs are desired that do not cause cancer, do not disrupt untargeted protein function and do not give rise to long-term adverse events. In accordance with the FDA²⁸, we recommend unbiased sequencing methods that are appropriate for the target cell or tissue to detect off-targets in genomic loci associated with cancer risk or loci that may disrupt cellular function. While detection and interpretation of off-targets is rapidly evolving given advancements in sequencing and other technologies and discoveries in the biological understanding of human genomes, the principles we outline here to select appropriate tools for off-target detection and the framework to interpret findings are applicable to both current and future tools.

Most experimental assays to characterize off-target effects currently use supersaturating doses of the editor, which overestimate off-target rates, and cells that may not use the same type of DNA repair system as the intended tissues. We recommend that on-target and off-target edit outcome profiling is performed in the relevant cell type that uses the matching predicted in vivo DNA repair system when possible and at physiological doses to replicate the most physiologically likely off-target outcomes. Quantification of small DNA changes and large structural changes at both the on-target site and possible off-target sites using cells with the same DNA repair system as the tissues receiving the final product should be included in this analysis.

One of the most important considerations is to limit the risk of secondary cancer. We recommend that careful attention is paid to editing of known oncogenes and tumor-suppressor genes, which could increase the risk of secondary cancer. Such editing can occur from the editor itself or from the delivery vehicle such as AAV integration at on-target or off-target cut sites^54,55. By referencing the Catalogue of Somatic Mutations⁵⁶ and deep sequencing to lower the threshold of detection of rare but meaningful events, off-target cancer mutation detection may be standardized across platforms and targets. Over time, knowledge from standardized testing will increase our ability to infer whether specific genomic changes, such as to unmapped or poorly characterized regulatory elements, will increase the risk of oncogenesis. It is also recommended to perform some of these experiments at supersaturating doses to increase detection of low-probability off-targets that could produce cancer-causing mutations. We recommend the establishment of a field-wide intolerance for any known cancer-promoting edit at adequate, low thresholds of detection.

To evaluate the risk of the CRISPR therapy disrupting cellular function, we recommend profiling whether an off-target occurs in a site with known functional impact such as either a coding sequence or a regulatory noncoding region^57,58. Avoiding approaches or guide RNAs that result in changes to coding sequences or known regulatory elements is prudent. To do this, one must know which organs and cell types will be edited and whether the genomic change alters cellular function in those tissues. It is important here to distinguish genomic off-targets (which may occur in any cell type or only subsets of cell types) from off-organ editing. On-target or off-target editing outside the target organ may or may not have clinical relevance. To evaluate the risk that CRISPR induces cellular dysfunction, care must be taken to consider both on-target and off-target editing in tissues receiving the editor, both on-organ and off-organ. For example, if an edit, whether it be on-target or off-target, could conceivably disrupt function of the lung epithelium, but the editor only encounters the eye, brain, blood and liver, then the potential edit is unlikely to disrupt cellular function and have clinical relevance. One caveat is edits in tumor suppressors and oncogenes, for which we recommend a zero tolerance for off-target editing as described above, even if that edit is predicted to only occur in a tissue not reliably encountered by the editor. Organs involved in blood filtration, such as the kidneys and liver, may be more susceptible to on-target and off-target gene editing as the editors are cleared from the body, even if these organs were not therapeutically targeted. Thus, CRISPR effects on these organs should be carefully profiled. While human cellular models are best suited for surveying the entirety of the human genome, transgenic or humanized animal models may be suitable for detecting on-target DNA editing in off-organs.

As we learn more about the biological consequences of genomic changes, we recommend that these be cataloged in a field-wide manner to create a map of genomic changes with functional consequences that we should avoid. A database of such changes⁵⁹ could be used to fuel machine learning and other high-throughput analysis, although careful attention must be paid to protect individual patient genetic information in such field-wide or open-source endeavors. The opportunity to monitor off-targets in patients participating in clinical trials or during post-marketing is desirable, although it should be noted that most in vivo targets will be inaccessible in living patients, apart from biopsies and blood cells. While blood cells are easier to obtain than solid organ biopsies, genomic changes in blood cells may not reflect editing outcomes in other organs. Cell-free DNA should be evaluated for its sensitivity to capture off-targets from solid organs in living patients. Evaluation of off-targets in post-mortem tissue from clinical trial participants would also benefit the field. Currently, there is no regulatory requirement or funded initiative to deposit such information.

The FDA currently recommends 15-year long-term follow up²⁸. While this recommendation is important in a new field such as CRISPR, the field is grappling with how to fund it, who will perform the long-term clinical monitoring and how reports of adverse events will be made and evaluated. Long-term follow-up and reporting are typically done in well-funded clinical trials or in government-sponsored programs such as those for vaccine safety monitoring. Challenges to long-term monitoring after CRISPR therapy include small numbers of participants in rare disease trials, increased risk of loss of follow-up in a single-dose therapy trial and difficulty discerning adverse events from small sample sizes. It is conceivable that many CRISPR trials may be bespoke or enroll only a few patients with rare mutations and thus never receive the level of funding of clinical trials that enroll thousands of patients and would permit such monitoring. In addition, patients in clinical trials are often lost to follow-up after the conclusion of the trial, for example, by moving away from the study site, and may not be incentivized to return for monitoring once the therapy is complete. Because many CRISPR therapies will be single dose, as opposed to trials in which repeat dosing necessitates proximity to the trial site, patients will be more likely to participate from remote locations, even crossing borders to enroll. Lastly, monitoring the outcomes of very small trials with one or a few patients will not be as sensitive at detecting uncommon adverse events as aggregate data across populations. To overcome these challenges, a centralized and government-sponsored reporting system similar to vaccine safety monitoring in which a patient, clinical trialist or primary physician could report health events even years after the therapy is administered is needed. Aggregating data across editors, delivery vehicles and on-target and off-target events from small trials is the only practical way to detect infrequent but meaningful clinical events.

Concluding remarks

We are currently in the midst of a CRISPR-based gene therapy revolution. With the possibility to address devastating disorders, including previously untreatable and fatal diseases, the field is quickly working toward providing novel therapies to patients. Here, we highlight the notion that no effective therapy is ‘perfect’ and that medicines have inherent side effects. Therefore, instead of setting the bar for CRISPR therapies at perfection, we should instead focus on fully quantifying and assessing the benefit–risk ratio for a given condition. ‘Doing nothing’ has a morbidity, mortality and healthcare expenditure cost; therefore, patients and physicians tolerate greater risks for therapies targeting life-threatening diseases. Chemotherapeutics, radiation and teratogenic medications can result in secondary cancer or genomic damage, which can be transmitted to future generations, concerns shared by CRISPR therapies; however, we would not halt the use of these medications. In concurrence with this standard, it is our opinion that CRISPR therapies that increase survival should not be withheld for the pursuit of zero off-target effects. Instead, the rigorous steps outlined here should be taken to detect CRISPR off-targets and assess their associated benefit–risk profiles, including avoiding potential cancerous activity using high-sensitivity sequencing and other methodologies to detect and characterize genomic changes, zero tolerance for potentially oncogenic off-targets and minimizing off-targets that lead to deleterious functional changes or genotoxicity. By adopting an increased benefit–risk-appropriate view of off-target effects, novel CRISPR therapies will not only benefit high-risk patients with severe and fatal diseases but will also generate the outcome data necessary to be able to begin modeling benefit–risk profiles for patients with moderate- and low-risk conditions. In the longer term, these efforts will contribute to the design of safe and effective CRISPR therapies.