Monitoring Biological Effects of Somatic Cell Genome Editing

Abstract

CRISPR-based genome editing therapeutics are entering the clinic, bringing great potential benefit but also a risk of adverse events. Preclinical-to-clinical toolkits are needed to assess the safety and efficacy of these new therapies and accelerate advancement. Emerging technologies to monitor the biological effects of genome editors cover a range of biological scales, from the direct measurement of editing outcomes, to human microphysiological systems, to non-invasive imaging in vivo. Measurements of on-target and off-target editing outcomes, including sequences unique to humans, provide essential benchmarks to understand functional responses. Microphysiological systems, such as human organoids and organs-on-chips, enable phenotypic evaluations of editing strategies in varied organ lineages and disease states. Non-invasive in vivo imaging modalities track the biodistribution and activities of genome editors in living organisms. Collectively, these technologies provide complementary insights across different scales, from the single nucleotide to the whole organism, bridging preclinical therapeutics development with clinical trials.

I. Introduction

The therapeutic potential of CRISPR-based genome editing systems, which modify DNA sequences at specific locations in the genome, is rapidly materializing¹. The first CRISPR-based cell therapy, CASGEVY (exagamglogene autotemcel), was recently approved by several regulatory agencies, including Health Canada, the European Commission, and the US Food and Drug Administration, for the treatment of sickle cell disease or beta thalassemia^2,3. Numerous chimeric antigen receptor (CAR) T-cell therapies, genome edited to enhance their efficacy and compatibility, are currently advancing through clinical trials^4–6. Beyond cell therapies that undergo editing procedures ex vivo, a novel class of injectable therapeutics capable of directly editing DNA in vivo is entering clinical trials, with the liver, retina, and skeletal muscle among the first organs or tissues to be targeted^7–10. These trials build on preclinical studies conducted over the past decade^11–16. In a recent example, an infant patient with carbamoyl-phosphate synthetase 1 deficiency was treated with a bespoke CRISPR base editing lipid nanoparticle (LNP) formulation called kayjayguran abengcemeran (k-abe), resulting in lessening of symptoms¹⁷.

Although gene editing therapeutics offer great potential benefit, their path to the clinic can be challenging¹. FDA provides guidance and recommendations for information that should be included in an Investigational New Drug (IND) application for these therapeutics¹⁸. Much remains unknown about these relatively complex biological drug products, which need to be appropriately dosed to maximize efficacy while minimizing the risk of adverse events^19,20. Essential organs targeted by genome editors are susceptible to acute physiological damage²¹, driver mutations that cause tumorigenesis²², inflammation and chronic immune responses^23,24, which can be triggered by the delivery of genome editing therapeutics or resulting editing events^25,26. Substantial adverse events, including at least one fatality, have arisen in some recent trials^8,14,27. which can halt clinical development and could ultimately have a detrimental effect on future developments.

Compared to earlier viral-vector based gene therapy approaches, genome editing enables more precise and predictable editing events, with the potential to reduce off-target genotoxicity. However, several considerations can limit efficacy, including inherent inefficiency of the genome editing enzyme versus DNA repair processes, incomplete specificity for the desired editing event at the target locus, physical properties such as size and hydrophobicity which restrict delivery of therapeutics into cells, and difficulties in accurately delivering the therapy to the desired site of action within the body. The delivery of genome editors relies upon either conventional gene therapy vectors, for example, adeno-associated virus (AAV) or lentivirus vectors, or more novel formulations such as lipid-based nanoparticles (LNP), whose risk profiles are still evolving²⁸. The clinical application of CRISPR approaches therefore requires careful selection of doses, editors, and monitoring, as well as a clear communication with patients regarding the risks associated with their use, which may not yet be fully understood.

For clinical trials to succeed, in vitro and in vivo tools and comparative approaches are needed that provide reasonable predictions of human outcomes. Traditional systems to assess the biological effects of genome editing include immortalized cell lines and mouse models. However, these models do not always accurately represent human biology. Recent FDA guidelines recommend that, “whenever possible, the potency assays be performed in the target cells or tissues¹⁸.” Since 2018, the US National Institutes of Health (NIH) Somatic Cell Genome Editing (SCGE) Consortium has sought to accelerate the development of methods to conduct and assess genome editing in disease-relevant lineages and to advance genome editing therapeutics²⁹. The SCGE Consortium has prioritized development of three emerging technologies: methods to measure editing outcome, microphysiological systems and non-invasive imaging of live organisms.

Collectively, these three technologies provide complementary insights across different biological scales, bridging preclinical therapeutics development with clinical trials. At each scale, the definition of an ‘on-target’ and ‘off-target’ effect is context dependent and represents a specific parameter that can be assessed by the appropriate technology: editing measurements indicate whether the desired change within the DNA sequence has occurred; human microphysiological systems provide insight into whether the therapy produces its target phenotypic effect; and in vivo imaging can reveal in real time whether a therapeutic reaches its target organ or has more widespread biodistribution. By integrating these distinct inputs and outputs, a holistic picture of the therapy emerges (Figure 1).

Figure 1. Assessing on-target and off-target biological effects of genome editing therapies.

Editing measurements, microphysiological systems such as organoids and organs-on-chips, and in vivo imaging techniques present a spectrum of information across different scales, from single nucleotides in vitro to the whole body in vivo. The precise meaning of ‘on-target’ and ‘off-target’ is specific to each scale and assay, each focusing on a particular parameter (red type), for which examples are provided (white and black boxes, respectively, for on- and off-target). Comparative analysis of results from different scales, together with inputs from traditional model systems (Box 1), can produce a holistic understanding of the therapy. Ultimately, this leads to more facile regulatory processes and successful outcomes in the clinic.

Here, we review key technological approaches to assess the biological effects of genome editing in different organ systems and disease contexts, as well as related challenges. We summarize lessons learned from the SCGE Biological Effects Initiative, a working group of ten multi-laboratory investigative teams, and propose a framework for the incorporation of these emerging technologies into the development of genome editing therapies. In an accompanying SCGE Toolkit (https://scge.mcw.edu/toolkit/), we provide public access to raw datasets, including protocols, models and approaches to capture any unexpected outcomes of gene editors and their delivery systems²⁹.

II. Assessing edits in DNA

Editing events can be classified as ‘on-target’, meaning that they arise at the target nucleotide sequence, or ‘off-target,’ meaning that they arise elsewhere in the genome. On-target effects can be further classified as ‘intended’, referring to the desired editing outcome, and ‘unintended’, referring to imprecise or deleterious editing outcomes at the target site. Evaluating the intended on-target impact of a genome editor provides a critical metric of therapeutic potency and efficacy. Moreover, genome editors are associated with substantial risk of unintended editing outcomes that must be carefully evaluated, including both unintended on-target as well as off-target effects. In the case of allele-specific editing, off-target effects can include the non-targeted allele (‘allelic off-target’).

Evaluating on-target editing events

Ideally, the on-target locus is the major DNA sequence affected by genome editing therapeutics. Beyond the intended edit, a variety of unintended editing outcomes can arise. Double-stranded breaks (DSBs), which are associated with Cas9 but are also a side effect of prime and base editors (perhaps based on their DNA nicking activity), can produce chromosomal rearrangements, large deletions, concatemeric insertions of AAV vectors, and other undesired events^30–32. A variety of tools have been developed to identify and quantify the rates of both intended and unintended on-target editing events, with various advantages and disadvantages, ranging from cost, time, throughput, and accuracy (Table 1). Each technology is typically designed to detect specific types of edits, as described below.

Table 1.

Assays commonly used to quantify on-target and off-target editing

	Sanger Sequencing	Surveyor, T7E1 and IDAA	ddPCR	Amplicon Sequencing	SMRT	Nanopore
Small Deletions, Insertions, Substitutions	Yes	Yes	Yes	Yes	Yes	Yes
Inversions	Yes	No	Yes	No	Yes	Yes
Large Integrations or Deletions	No	No	Yes	No	Yes	Yes
Structural Genomic Rearrangements (Translocations or Chromothripsis)	No	No	Depends on probe and primer design	When using a transposon to prime off of51	Yes	Yes
Transcriptome Editing	No	No	No	No	No	No
Single-Nucleotide Resolution	Yes	No	No	Yes	Yes	Yes
Max. Read Length (bp)	600	N/A	N/A	600	> 1,500,000	> 4,000,000
Quantifies Individual Editing Outcomes	No	No	Limited	Yes	Yes	Yes
Quantitative Accuracy	++++	+	+++	+++	++	+
Requires DNA Amplification	Yes	Yes	Yes	Yes	No	No
Ability to Detect Unexpected Outcomes	Yes	No	No	When using a transposon to prime off of51	Yes	Yes
~LOD (%)	15	0.1	5	0.1	0.5	1
References	39–41	34–36	48–50	75,76,78,81,197	49,52,56	53–55

Non-sequencing assays

Using editing platforms that introduce a DNA double-stranded break, such as CRISPR-Cas9, often begins with screening multiple guide RNAs (gRNAs) for editing efficiency. Protocols such as the Surveyor nuclease S1 (CEL1)³³ or the T7E1 (T7 endonuclease 1) mismatch detection assays,³⁴ or capillary electrophoresis-based sizing methods such as the IDAA (Indel Detection by Amplicon Analysis) assay³⁵ are typically less expensive and faster than sequencing-based methods, and have a lower LOD³⁶. For instance, T7E1 has been used to rapidly screen for editing events in kidney organoids subjected to AAV-delivered CRISPR-Cas9, using human embryonic kidney (HEK) cells as a positive control. Surveyor, T7E1, and IDAA can detect small insertions or deletions (indels) but do not provide sequence-level information, thus they have limited use to detect base editing events, and are less quantitative than other methods. They are also ineffective at identifying and quantifying all editing outcomes, such as the insertion or deletion of a large sequence (> 500 bp), inversions, translocations or AAV integrations, all of which could pose safety risks.

Another popular tool for detecting editing events is a fluorescent reporter, including assays such as fluorophore conversion through sequence replacement,³⁷ as well as traffic light reporters³⁸. These reporter systems can be introduced via transgenic cassettes into cells or model organisms, enabling detection of genome editing events with relatively high throughput. Such reporters, however, are indirect measures, and may be limited to intended on-target editing events, whereas unintended on-target events or off-target events may not be captured at all. Depending on the target lineage, the reporter system may also need to be engineered, which may require considerable effort, or encounter obstacles such as transgene silencing. Thus, while fluorescent reporter systems can be very useful for rapidly screening and optimizing editors, editing readouts agnostic to the target sequence are generally needed to demonstrate proof of concept for a therapeutic strategy.

Sanger sequencing

Standard Sanger sequencing of PCR amplified target regions (amplicons), which is based on PCR chain termination using dideoxyribonucleotides followed by size separation via gel electrophoresis, provides ~ 600 bp of continuous sequence and without complex assembly or computation, and is a mainstay for detecting mutations in DNA owing to its low cost, convenience, very high accuracy, and simplicity³⁹. For genome editing, Sanger may be adequate for assessing on-target base editing. For example, in base-edited induced pluripotent stem cells (iPSCs) with clinically relevant nonsense mutations in PKD1 and PKD2, homozygotes, heterozygotes, and non-mutant controls can be readily distinguished with Sanger sequencing, and the peaks at each mutant nucleotide correlate directly with gene dosage⁴⁰. Software such as EditR and BEAT (Base Editing Analysis Tool) can be further utilized to quantify base editing efficiency from Sanger chromatograms^41,42.

A major drawback is that Sanger chromatograms only provide an averaged output for each base position. With a limit of detection (LOD) of ~15 %, low-abundance mutations, such as off-target base editing events, are difficult to detect relative to the ‘consensus’ sequence⁴³. In addition, indel mutations exhibit overlapping Sanger sequencing traces, which must be deconvolved either by hand, or using computational software such as Tracking of Indels by DEcomposition (TIDE) and Inference of CRISPR Edits (ICE), to infer the presence and abundance of multiple mutations^44,45. In clonal populations, where only two alleles are present for any given locus, such deconvolution is feasible, although best practice is to confirm the sequence of each individual allele with topoisomerase-based cloning, for instance to establish a new iPSC mutant line derived using CRISPR-Cas9^46,47. In a mixed population, where diverse sequences are present, each at relatively low abundance, Sanger sequencing is unlikely to be accurate, even when coupled with sophisticated computational deconvolution tools.

Digital droplet PCR

Digital droplet PCR (ddPCR) can identify and quantify a wide range of editing outcomes, including single nucleotide changes and insertions and deletions of a wide range (1 bp - 1 Mb). With high precision and sensitivity (LOD of 5%), ddPCR can quantify the number of molecules that contain a target sequence as defined by designed probes and primers. The digital nature of the output overcomes the bias in quantifying differentially sized amplicons by standard PCR-based methods, including next-generation sequencing (NGS). For example, droplet digital PCR eXcision Reporter was able to quantify deletions up to 172 kb in length with a high degree of precision^48,49. Multiplexing of probes enables more complex assays, such as measurement of the allele specificity of a given editing event^48,50. Newer instruments, such as Qiagen’s QIAcuity One or BioRad’s QX600, incorporate additional fluorescent channels to facilitate detection of multiple editing outcomes in a single multiplexed reaction. While ddPCR can be designed to detect essentially any category of editing outcome, its primary limitation is that each assay is limited to detect specified editing outcomes based on the design of the primers and probes. For example, deletions larger than those expected or complex rearrangements that disrupt primer or probe binding will not be detected.

Targeted amplicon sequencing

NGS of amplicons provides direct reads of individual outcomes, which increases sensitivity and accuracy compared with inferred or averaged reads. High-throughput sequencing can generate millions of short sequence reads for many amplicons, providing extensive coverage of a target region. This deep coverage enables the detection and quantification of rare genetic variants, including low-frequency editing events, with high sensitivity (LOD of ~0.1). The most common approach, PCR-based amplicon sequencing, involves PCR amplification of the target regions using primers designed to flank the regions of interest. These amplicons are then sequenced using NGS platforms such as the Illumina MiSeq, NextSeq, and NovaSeq systems. Amplicon sequencing provides quantitative analysis of editing outcomes, offering insights into editing efficiency, specificity, and the prevalence of different types of genetic alterations.

Similar to ddPCR, NGS relies on specific primer sequences to detect editing outcomes, and therefore cannot detect large structural genomic rearrangements if one or both of the sequences complementary to the primers is lost. Furthermore, the inherent PCR bias for differentially sized amplicons makes NGS suboptimal for quantifying larger deletions (> 100 bp) and other more complex rearrangements. To overcome this challenge, an assay called UDiTaS (Uni-Directional Targeted Sequencing methodology) uses a transposase to fragment the genome and tag each fragment with a transposon sequence, which is then used as a template for priming, allowing for less-biased amplicon sequencing⁵¹. This strategy has been used to evaluate multiple editing outcomes, including deletions, inversions, and insertions of a specific sequence, and has identified unintended editing events including large sequence rearrangements and integration of the delivery vector^15,16,51.

Long-read sequencing

Long-read sequencing, including single-molecule real-time (SMRT) sequencing (by Pacific Biosciences) and nanopore sequencing (by Oxford Nanopore Technologies), is useful for sequencing large DNA molecules (>1,500 kb) in order to identify large deletions, insertions, or structural rearrangements without relying on amplification using sequence-specific primers, which limits the types of editing outcomes that can be detected^48,49. For example, generating long sequencing reads enables direct observation of the breakpoints generated by genome editors and precise mapping of deleted regions without the need for extensive computational inference or assembly. Furthermore, long-read sequencing accuracy now rivals short-read sequencing platforms, with PacBio’s Sequel II platform achieving over 99% accuracy compared to Illumina NextSeq 500’s > 99.9% accuracy, with a low LOD of 0.5%, although the cost of long-read sequencing remains a considerable barrier to this technology^49,52. The accuracy of nanopore sequencing is generally estimated to be 95–99% which remains lower than that of NGS and SMRT sequencing, in part because the technology is not sufficiently sensitive to detect individual nucleotides passing through the pore, thus the LOD is ~1%⁵³. Nevertheless, nanopore sequencing provides precise nucleotide sequences within individual DNA molecules and generally archives the longest reads of DNA (in the low Mb range) of any sequencing platform⁵⁴. In addition to standard base-calling, nanopore sequencing can also detect epigenetic modifications, such as DNA methylation, or even changes in chromatin structure^55,56. Since DNA passes through the pore as a single strand, adding a hairpin to the original double-stranded DNA enables a form of duplex sequencing, which can greatly increase sequencing accuracy⁵⁷. For genome editing events, which may occur with low frequency and require numerous reads to quantify, nanopore sequencing has emerged as an option that offers many of the strengths of NGS with the convenience and low cost of Sanger sequencing.

Detecting off-target editing events

Off-target editing events can be deleterious, even at low frequencies. For instance, ‘driver’ mutations, such as chromosomal translocations, in even 100 kidney cells can lead to renal clear cell carcinomas (RCC) years later²². The risk of such events may be linked to ‘hotspots’ that are more susceptible to chromosomal rearrangements, such as telomeres ⁵⁸. Furthermore, at least in CRISPR systems, successful gene editing events are favoured in cells with impaired DNA damage repair pathways such as P53 mutations, which may promote the selective survival of cells with imbalanced chromosomes or aneuploidy^59,60. Such events take years to manifest clinically and are difficult to model in mouse genomes⁶¹. In an example of gene therapy gone awry, hematopoietic stem cells subjected to retroviral gene transfer ex vivo were initially hailed as a successful treatment for X-linked severe combined immunodeficiency, but gave give rise to leukemias in young patients just a few years later in 2003, due to off-target integrations activating proto-oncogenes in the genome^62–64. Although rare, immunotherapeutic CAR T-cells have also recently been shown to be able to transform into T cell lymphomas⁶⁵. Similarly, 7 of 67 patients developed hematological cancer after gene therapy for cerebral adrenoleukodystrophy, perhaps in association with the strong ubiquituous promoter used in the viral vector ⁶⁶.

Allelic off-target edits can also pose risks. For instance, at the C9orf72 locus, where a hexanucleotide repeat expansion can cause amyotrophic lateral sclerosis (ALS), off-target disruption of the non-mutant allele increases expression of the pathological mutant allele.⁶⁷ This may explain the worsening of symptoms observed in some patients in a clinical trial of an antisense oligonucleotide targeting this locus.⁶⁸

Cas enzymes function by scanning the genome and interrogating sequences containing a 2–6 bp protospacer adjacent motif (PAM), followed by DNA cleavage upon binding of the gRNA to a target sequence⁶⁹. The risk of off-target activity is related to PAM recognition ability and the degree of similarity to the gRNA sequence. A variety of genome-wide analysis methods exist to predict, or nominate, off-target editing events.

Computational assays

Sites at high risk for off-target activity can be nominated using computational (in silico) prediction based on sequence analysis software (COSMID⁷⁰, CCTop⁷¹, Cas-OFFinder⁷²). Artificial intelligence tools employing machine learning techniques can further enhance in silico prediction of off-target events by training neural networks on large-scale, empirical datasets of cleavage events representing multiplex gRNAs as ground truth, including Elevation⁶⁹, DeepCRISPR⁷⁰, and CRISPR-Net⁷¹. In silico prediction tools generate risk scores for a particular gRNA associated with off-target sites, an approach that is rapid and low-cost, but does not involve physical testing, which may be insufficient to obtain regulatory approval for a candidate therapy.

Biochemical assays

Biochemical assays, including Digenome-seq and SITE-seq^73,74, CIRCLE-seq and CHANGE-seq^75,76, and ONE-seq⁷⁷, identify sites that undergo DNA cleavage in vitro, and can be performed at genome-wide scale. A variant of CHANGE-seq, CHANGE-seq BE, has been developed to nominate base-edited sequences⁷⁸. Such assays provide a physical basis for off-target nomination, but remain limited by the relative simplicity of the reaction conditions, which do not fully recapitulate the relative concentrations or complexity of molecules in the cellular environment, chromatin structure, or influence of endogenous DNA repair pathways on editing outcomes^76,79. However, biochemical assays with subsequent validation may suffice to advance a therapeutic into the clinic, as was the case with k-abe which was tested using ONE-seq and CHANGE-seq BE prior to administration into a high-risk infant¹⁷. Biochemical assays have advantages for characterizing in vivo genome editing approaches, as they can be used as a single nomination experiment, rather than separate cellular nomination assays for all possible affected cell types. They can be highly sensitive, although the limitation is that many candidate off-target sites that are edited in biochemical conditions may not result in detectable editing outcomes in cells.

Cell-based assays

To nominate off-targets in a more physiologically relevant context, genome-wide cellular assays have been developed, including BLISS (break-labeling in situ sequencing), GUIDE-seq (Genome-wide, Unbiased Identification of DSBs Enabled by Sequencing), and DISCOVER-seq (Discovery of In Situ Cas Off-targets and Verification by Sequencing)^80–83. GUIDE-seq relies on the integration of short double-stranded oligonucleotides during the cellular repair of nuclease cleavage, followed by targeted amplification using the captured oligonucleotide sequence as an anchor⁸¹. BLISS employs Tn5 transposase to simultaneously fragment DNA and ligate sequencing adapters, a process known as tagmentation, for labeling and sequencing sites of DNA double-strand breaks in cells and tissues, including those induced by Cas nucleases⁸⁰. DISCOVER-seq is a variation of chromatin-immunoprecipitation followed by sequencing, using antibodies directed against DNA repair proteins such as MRE11 to identify sites with active DNA repair⁸². Furthermore, DISCOVER-seq and a modification of GUIDE-seq (GUIDE-tag) have been applied to detect off-target activity of gene editors in vivo⁸⁴.

Additionally, tools such as CAST-seq can be used to identify and quantify chromosomal rearrangements at off-target sites⁸⁵. To nominate off-targets in the context of prime editing, cell-based assays called TAPE-seq (TAgmentation of Prime Editor sequencing )⁸⁵ and PE-tag (prime editor off-target detection system)⁷⁶ have been developed, which utilize tagmentation to assess off-target events of prime editing⁸⁶.

Validation of off-targets

Importantly, no prediction method can directly quantify editing outcomes. Any nominated off-target edits must be evaluated empirically, using methods for measurement of editing outcomes such as those described above (see Evaluating editing outcomes). For clinical studies, all reproducibly detected off-target candidates are ideally gathered for follow-up validation. As the total number of off-target events may be impractical to assess individually, a subset of nominated sites can be selected for further validation in an appropriate cell type. These sites can be selected based on their predicted likelihood for mutagenesis or based on other properties, such as the risk of genotoxicity (for example, in a tumour suppressor gene such as BRCA1). For instance, in the case of k-abe, 21 nominated off-target sites were prioritized for amplicon-based validation, which revealed few off-target effects, boosting confidence that the therapy was safe¹⁷.

A validation analysis comparing different prediction methods was recently performed with NGS in primary haematopoietic stem cells, based on a side-by-side analysis of computational tools (COSMID, CCTop, and Cas-OFFinder) and previously published data by different investigators for a range of physical techniques (CHANGE-Seq, CIRCLE-Seq, DISCOVER-Seq, GUIDE-Seq, and SITE-Seq)⁸⁷. 200 nominated sites were selected by NGS validation for each of 11 different guide RNAs. The total number of nominated sites varied widely amongst the different methods, but no single approach clearly outperformed all of the others in predicting the final NGS analysis, with COSMID, DISCOVER-Seq, and GUIDE-Seq seemingly producing the best results⁸⁷. As there is some variation between the outputs, clinical development programmes should use two or three different methods to nominate off-targets^7,11.

Challenges for Off-Target Sequencing

Off-target nomination techniques have certain caveats. The full landscape of possible off-target effects is vast, and we are still discovering it. For instance, certain base editors can cause gRNA-independent off-target mutations^88,89 as well as transcriptome-wide RNA editing⁹⁰ which can be challenging to predict or detect with existing methods. As different vulnerabilities can differ between individuals, assays should ideally be conducted based on more than one reference genome. Tools such as CRISPRme integrate data on single-nucleotide polymorphisms and indel genetic variants to help predict off-targets in the broader population, beyond reference genomes⁹¹.

Similarly, whether different types of somatic cells exhibit different off-target effects remains unclear. To determine whether these methods are accurate, it will be important to develop validation systems that match the predictive inputs, both in terms of reference genome and cellular identity, and to test them in diverse contexts. For instance, k-abe induced an off-target edit in a cell line derived from a hepatocellular carcinoma, but not in primary hepatocytes from three different donors¹⁷.

III. Assessing effects in human disease models

Beyond their genotypic effects, it is critical to understand the phenotypic effects of genome editing therapeutics. Historically, such information has been gathered through a combination of animal models, which naturally lack the precise target sequences found in humans, and simpler human monolayer cultures. Microphysiological systems, which can broadly be considered to include tissue explants, organoids and organ-on-chip devices, can help bridge the gap between simple 2D cell cultures and complex animal models (Box 1)⁹². Occupying the intermediate range of the biological scale between sequence-driven readouts and in vivo studies, microphysiological systems are designed to replicate key features of living tissues, such as heterocellular patterning and disease-specific phenotypes.

Box 1. Traditional disease models versus microphysiological systems.

Microphysiological systems have clear advantages over 2D cell cultures, which include monolayers of immortalized cell lines such as HEK cells. 2D cultures can be transfected or transduced with high efficiency after dissociation and replating to screen different formulations for on-target and off-target editing events^76,231. However, such systems are artificial compared to in vivo tissues, which are not immortalized or dissociated prior to treatment, and express differentiated characteristics, heterocellular interactions, and 3D morphology that can reduce efficiencies of transfection and transduction. In contrast to 2D cultures, microphysiological systems can recapitulate 3D features of target organ architecture and function, to simulate effects on the specific target cell types and target mutations. Importantly, this includes the potential to produce disease phenotypes in vitro. Because microphysiological systems are intricate structures that simulate the tissue microenvironment, they are more likely to reflect the challenges of editing tissues in vivo compared to immortalized cells. Microphysiological systems may also enhance our ability to predict human-specific adverse events that could be missed in other models^29,47,129.

Although transgenic mouse models can be generated with human sequences inserted into the mouse genome, the generation of such models is low-throughput, and may not fully recapitulate human phenotypes due to physiological differences between species. For example, some disease genes or risk variants, such as APOL1, lack natural orthologues in model organisms¹³¹. Moreover, disease states or specific phenotypes, such as autosomal recessive polycystic kidney disease (ARPKD) or nephropathic cystinosis, fail to reproduce in animal models despite the conservation of orthologous genes and/or mutations^232,233. Infectious and/or delivery agents may readily infect people and human organoids but not animal models (for example, SARS-CoV-2 did not robustly infect rodent models due to divergence of ACE2 receptor sequence)¹³³. Finally, adverse events can be specific to the human immune system²³⁴. Species-specific differences in genome structure and sequences also influence the pharmacodynamics of genome editing therapeutics, whereas differences in viral receptor and drug transporter expression affect the pharmacokinetics of payload delivery. Thus, microphysiological systems provide a valuable complementary system with which to assess genome editing interventions. Microphysiological systems can achieve higher throughput than corresponding animal models, and require smaller drug quantities for testing.

In the context of microphysiological systems, an ‘on target’ effect indicates the desired impact on the organ or tissue, for instance reversal of a disease phenotype, while ‘off target’ refers to adverse effects, such as incidental cytotoxicity (Figure 1). Use of microphysiological systems, while not required to advance therapeutics into the clinic, has been encouraged by federal agencies, with a recent guidance stating that “In the long-term (3–5 years), FDA will aim to make animal studies the exception rather than the norm for pre-clinical safety/toxicity testing.”⁹³

Importantly, microphysiological systems can be generated from conventional primary or immortalized human cells, or from human iPSCs, which can differentiate into a wide variety of organ-specific lineages, often forming complex organoids (Figure 2)⁹⁴. Microphysiological systems can be derived from individual patients with naturally occurring genetic mutations, including the same patients that could be treated in a clinical trial, or can undergo genome editing to produce additional mutations or reporter lines. As such, microphysiological systems contain the precise types of cells and genes that genome editing therapeutics are designed to target and present unique opportunities to assess multidimensional readouts of editing efficiency, phenotypic efficacy, and safety of genome editors (Figure 2).

Figure 2. Human organoids and organs-on-chips enable studies of gene editing.

Examples of a microfluidic device containing human liver cells (primary or iPSC -derived hepatic epithelial and stromal cells), or alternatively kidney organoids, are depicted for studying CRISPR-Cas9 mediated gene editing via a stepwise workflow. The platform facilitates assessment of transduction efficiency (multiple delivery systems can be tested), and subsequent phenotyping including efficacy, editing efficiency, and potential toxicity in a human organotypic microenvironment. Outcomes inform decisions about whether to advance a therapeutic into the clinic (IND), potentially customized for the original patient from whom the MPS was derived.

Microphysiological systems are relatively new and still suffer from significant limitations in the context of CRISPR-based interventions. Nonetheless, proof of concept has recently been demonstrated for the use of microphysiological systems to test gene therapy applications in a variety of organ systems (Table 2). We briefly describe advances in intestinal, liver, kidney and retina microphysiological systems as examples of editable epithelial lineages, as well as relevant limitations of this technology.

Table 2.

Microphysiological systems with representative genes for editing applications for different organ systems

Organ	System	Disease	Cell Types	Genes	Editor	Vector	References
Liver	Liver-on-chip Adult liver organoids	Transthyretin amyloidosis, α1 antitrypsin, Hepatocellular carcinoma	Hepatocytes, Kupffer cells	TTR A1AT EMX TP53	Cas9	AAV RNP lentivirus	104,108,109 98,103
Kidney	Kidney Organoids	Polycystic kidney disease (ADPKD, ARPKD), Ciliopathies, Focal and segmental glomerulosclerosis, Cystinosis	Proximal tubules, Distal tubules, Podocytes	PKD1 PKD2 PKHD1 KIF3A KIF3B APOL1 AAVS1 CTNS CRTC2	Cas9 ABE CBE	AAV RNP	117,127,198,199 40,46,116,129,130,132,200
Eye	Retinal pigment epithelium Retinal Organoids	Best disease, LCA16 disease Retinitis pigmentosa	Retinal pigment epithelium photoreceptor cells	BEST1 KCNJ13 CRB1 PRPF31 RPGR	Cas9 ABE	lentivirus LNP	12,13,201–203
Cardiac and Skeletal Muscle	iPSC cardiomyocytes,Muscle-on-chip	Cardiomyopathy, Muscular dystrophy, Pompe	Cardiomyocytes, Myoblasts	DMD GAA	Cas9 (exon deletion) ABE (disrupting splicing)	AAV	16,26,204–210 165,211–214
Blood	Primary hematopoietic cells	HIV, CTLA4 deficiency, Combined immunodeficiency, Autoimmune lymphoproliferative syndrome, Systemic lupus erythematosis, Acute lymphoblastic leukemia	T-cells	AAVS1 CCR5 CTLA4 CXCR4 FAS LAG3 PDCD1 PTPN2 PTPN6 TRAC TRBC1	Cas9	RNP	76,81,215,216
Brain & Central Nervous System	iPSC neurons, cerebral organoids	Autism spectrum disorder, Neuropsychiatric disorder	Neurons (excitatory, lower motor, dentate gyrus-like), Astrocytes	CNTNAP2 CDK5RAP 3 STRAP DRD2	Cas9	RNP LNP	79,217,218
Gut	Intestinal stem cell organoids (enteroids), rectal organoids	Cystic fibrosis Familial Adenomatous Polyposis	Intestinal epithelial cells Colonic epithelial stem cells	CFTR APC	Cas9	AAV Electroporation	95,97,103,219

AAV, adeno-associated virus; LNP, lipid-based nanoparticle; RNP, ribonucleoprotein. Table is limited by space constraints and does not include all organ lineages or genes of interest.

Intestinal organoids

Organoids derived from adult stem cells, which can be cultivated extensively in vitro and differentiate into complex structures, provide a convenient form of microphysiological system in which to test genome editing therapeutics. Intestinal organoids (enteroids), for instance, can be derived from individual intestinal stem cells, which can be expanded in 3-D culture and passaged to produce clonal organoids. This passaging step provides an opportunity to transfect or transduce cells with relatively high efficiency, to conduct genome editing^95,96.

When stimulated with cyclic AMP, intestinal organoids swell via a mechanism involving the cystic fibrosis transmembrane conductance regulator (CFTR). Transfection of gene editing agents to correct CFTR mutations, either via homologous recombination or base editing, can rescue this function of CFTR in organoids derived from patients with cystic fibrosis^95,97. In another example, prime editing was applied to intestinal organoids to correct mutations in diacylglycerol-acyltransferase 1. This improved survival of the organoids upon exposure to fatty acids, which this enzyme is required to metabolize⁹⁸.

A caveat to these experiments is that the genome editing step is performed on dissociated intestinal stem cells, which are subsequently grown into organoids. To improve efficiency, a selection step (e.g. with an antibiotic) may be incorporated into the workflow⁹⁵. This is suitable for scenarios in which gene editing will be conducted ex vivo, on a cell population that can be reintroduced into the body. For direct applications of gene editing in vivo, however, this may not represent the native properties of the intestine or its resident stem cells.

Liver organoids and liver-on-a-chip

The liver is a major destination for drug metabolism and clearance, and has emerged as an important target organ for gene editing, with clinical trials ongoing for transthyretin amyloidosis (NCT04601051, phase III), hereditary angioedema (NCT05120830, phase II), and PCSK9-associated familial hypercholesterolemia (NCT06164730, phase Ib)^7,10,14,27. Despite the ease of engineering and programmability that has revolutionized genome editing with CRISPR, the delivery of CRISPR itself or its vehicles could yield unintended adverse consequences on liver function^99–101, with the risk of genome-related hepatotoxicity and an immune response to gene editing raising substantial safety concerns¹⁰².

Liver organoids representing either ductal or hepatocyte lineages can be maintained in 3-D culture conditions, and subjected to gene editing, in a manner similar to intestinal organoids. Using prime editing, mutations in HEK3, CTNNB1, and ABCB11 have been generated in liver ductal organoids⁹⁸. Efficient knock-in of larger gene cassettes could also be achieved in either liver ductal or hepatocyte organoids using a CRISPR approach based on non-homologous end-joining¹⁰³. Similarly, CRISPR has also been applied in iPSCs to introduce gene regulatory networks, which improved differentiation and vascularization of derived liver organoids¹⁰⁴.

Liver organoids lack higher-order tissue patterning and have limited capacity to introduce flow throughout the structure. A 3D microfluidic liver-on-chip platform has been developed to overcome such limitations^105–107. This platform enables the investigation of gene-editing tools in a physiologically relevant human model in vitro, narrowing the translational gap between mouse studies and human clinical trials (Figure 2). The successful implementation of gene editing within the liver-on-chip system hinges upon the meticulous design and optimization of the experimental setup, including crucial CRISPR components, such as the gRNA, Cas9 nuclease, and delivery system^108,109.

Although promising, liver organoids and liver-on-chip technologies have certain limitations that must be considered. As with intestinal organoids, gene editing steps in liver organoid systems have largely been conducted in dissociated precursor cells, which are subjected to selection techniques to enrich for edited cells prior to being expanded into organoids^98,103. Transduction efficiency can vary and is influenced by the experimental setup and optimization of the CRISPR components¹⁰³. Editing efficiency may vary across different genes, target sites, and cell types inherent to the liver, and is highly dependent on the donor genetic and medical backgrounds. Variations between different batches of primary human cells can affect the reproducibility of results, necessitating careful consideration, replicates, and controls to ensure reliable and consistent findings. These factors could potentially lead to significant differences between outcomes in vitro and in vivo.

Kidney organoids and tubuloids

As kidneys are a critical pathway for drug clearance, understanding how gene therapies are processed by these vital organs is important to predict pharmacodynamics and the potential for adverse events. Kidneys are also direct targets of gene therapy, as hundreds of genes have been identified as monogenic causes of inheritable glomerulopathies, tubulopathies, ciliopathies, and other renal disorders, altogether accounting for 24% of chronic kidney disease¹¹⁰.

Nephron formation is restricted to embryogenesis and brief periods after birth in mammalian species. Mimicking this developmental process, iPSCs can be differentiated into human kidney organoids. These complex cultures contain contiguous, nephron-like segments of distinct epithelial cell types along a proximal-to-distal axis, surrounded by stromal and endothelial cells. Compared to the mammalian kidney, these organoids have analogous multicellular architecture, cell type-specific gene expression patterns, transport functions, responses to injury, and genetic disease phenotypes (Figure 2)^111–114. Organoids can be generated in high-throughput formats or organoid-on-chip platforms to facilitate therapeutic screening and drug delivery assessments^115–117.

Polycystic kidney disease (PKD) is an inherited ciliopathy affecting 1 in ~1,000 people in which microscopic tubules expand progressively into large, fluid-filled cysts, caused by loss-of-function mutations in PKD1 or PKD2. In mice, only homozygotes manifest significant cystic disease, whereas in humans, germline heterozygotes are invariably affected. This has led to a two-hit hypothesis for human PKD, wherein cysts arise focally due to somatic mutations to the healthy allele in trans, which can be detected by whole-genome sequencing of PKD cysts¹¹⁸. Human kidney organoids with biallelic loss-of-function mutations in PKD or ciliogenesis genes spontaneously form cyst-like structures, reconstituting the pathognomonic phenotype^{40,46,111,116,119}. Cystogenesis is completely rescued when homozygous nonsense mutant iPSCs, generated using a cytosine base editor (CBE), are reverted to heterozygotes using an adenine base editor (ABE), suggesting that partial restoration of wild-type PKD1 or PKD2 gene expression could be sufficient to prevent disease⁴⁰.

A caveat to these experiments is that the editing has been performed in iPSCs, rather than in the kidney organoids themselves. As an alternative to organoids, primary renal tissue cultures grown in Matrigel with specialty media form ‘tubuloids’ that retain a substantial degree of tubule-specific gene expression and functionality^120,121. This system resembles the adult stem cell-derived intestinal and liver organoid systems, and offers an opportunity to conduct genome editing during the passaging steps. Under these circumstances, kidney tubuloids can also exhibit a cystic phenotype after CRISPR editing to disrupt PKD1 or PKD2¹²⁰. These experiments highlight the utility of organoids for studying gene editing therapies for PKD.

In addition to efficacy studies, systems to predict kidney injury are crucial¹²⁴. Given their high blood flow and abundant expression of drug transporters and viral receptors, kidneys may be a major site of cytotoxicity in response to genome editing therapies. Organoid proximal tubules exhibit specific injury responses, including DNA damage, to known nephrotoxicants such as gentamicin and cisplatin^{111,112,114,125,126}. These studies have contributed to the identification of homologous recombination and cyclin G1 pathways in the tubular DNA damage and repair, relevant for predicting and modulating the response to genome editing-related DNA damage^127,128. Kidney organoids can also mount an innate immune-like response involving interferon-γ signaling and APOL1, which can result in endothelial cell death due to pyroptosis, and changes in gene expression amongst epithelial cells^129,130. APOL1 has evolved only in certain primates, and risk variants found only in humans can cause kidney disease when triggered by an inflammation-inducing event such as viral infection¹³¹. An interferon response signature was also observed in kidney organoids infected with SARS-CoV-2, which specifically infects proximal tubules amongst the various nephron segments in these cultures^132,133. Whether genome editing also triggers such inflammatory responses is not yet known.

A lack of a functional vasculature to achieve glomerular filtration of editors and perfusion of the tubules remains a notable limitation of kidney organoids for genome editing studies. This issue contributes to the inefficiency of editing in organoids and limits their utility for studying phenotypic effects. Organoids implanted in vivo into immunodeficient animals become vascularized from the host and form pseudo-glomeruli capable of filtration, which indicates the potential for maturation^134,135. Efforts towards incorporating flow into kidney organoids in vitro may therefore lead to more advanced studies of genome editing to predict effects on the kidneys^116,117.

Eye organoids

Improved understanding of the genetic basis of inherited retinal diseases and the development of efficient tools for genetic diagnosis have driven major development in ocular gene therapy in recent years, with many existing and emerging gene, base and RNA editing technologies. In contrast to other organs and tissues, the eye serves as a convenient delivery target that benefits from protection against systemic immune reactions, a lower likelihood of deleterious off-target effects, a small treatment area, prolonged and localized retention of therapeutic product, and powerful, noninvasive, functional testing procedures to monitor treatment effects and efficacy in both animal models and humans¹³⁶.

The retina, which lines the posterior pole of the eye, is a translucent tissue that captures light, transforms it into an electrical signal, and transmits it to the visual cortex of the brain. The ‘outer retina’ comprises the rod and cone photoreceptors, which detect light, and retinal pigment epithelium (RPE), which maintains the health and function of these photoreceptors. Inherited vision impairments involving the retina - most often due to monogenic mutations that lead to retinal degeneration - affect over 1 in 1,400 individuals, most often caused by monogenic mutations that lead to retinal degeneration, and can be categorized based on the primary cell type they affect (e.g. rods, cones, RPE or a combination thereof) and whether the initial vision loss is central or peripheral¹³⁷.

Voretigene neparvovec (Luxturna), an AAV encoding human RPE65 for treatment of the inherited retinal disorder Leber congenital amaurosis (RPE65-LCA), was the first in vivo gene therapy approved by the FDA¹³⁸. More recently, a clinical trial of CRISPR-based gene-editing therapeutic EDIT-101 was conducted, targeting LCA10 (NCT03872479, phase I–II)⁹. Biopsy samples from human cadaveric retinal explants were cultured and used in the preclinical validation of EDIT-101, demonstrating high levels of transduction in photoreceptors and ~40% editing efficiencies¹¹.

Human iPSC-derived retinal cells and organoids have also provided crucial “disease-in-a-dish” models to understand cellular mechanisms underlying mutation-specific loss of vision and to develop and test gene therapies. Retinal organoids consist of laminated, well-organized layers comprised of cone and rod photoreceptors and other neural-retinal cell types. iPSC-derived RPE cells can be purified and cultured as polarized monolayers for structural and functional interrogation using a range of imaging techniques and assays^12,13. For example, iPSC-derived RPE models demonstrate the potential for gene augmentation to overcome the defect in ion channel function caused by some mutations in BEST1-underlying the RPE dysfunction in Best vitelliform macular dystrophy (also known as Best disease), with gene editing offering an alternative approach for mutations that are unresponsive to gene augmentation¹². To assess for potential off-target effects, single cell RNA sequencing analysis of edited vs. unedited cells was performed, based on global expression as well as curated gene sets, and did not find any significant differences¹². Similarly, a base editing strategy for the early-onset blindness disorder LCA16, in which RPE ion channel activity is also affected, successfully restored cellular function¹³.

While the field is promising, challenges remain. Differentiation of photoreceptors from iPSCs is a lengthy process, requiring up to 12 months of culture, and produces dense, complex organoids that can be difficult to transduce. Recent studies have focused on RPE, which are more readily differentiated, and can be purified from organoids and replated shortly before treatment^12,13. This results in substantial editing efficiency (~17% for LCA16), but the process of administration differs from in vivo. As there is no way to gauge sight in microphysiological systems, assays of therapeutic efficacy are limited to readouts such as electrophysiologal patch clamping of the endogenous channels^12,13. Indeed, measurements of visual acuity from the first 14 patients dosed with EDIT-101 showed only partial improvements, leading to the trial being paused⁹. Whether editing levels in these patients have achieved the editing efficiencies of ~40% observed in explant punches is not known, which highlights the fundamental challenge of bridging the translational gap between preclinical models and clinical trials.

General applications and limitations

In addition to the aforementioned lineages, a diversity of microphysiological systems representing other lineages, including muscle, blood, and brain, are available for the study of genome editing therapeutics (Table 2). For instance, iPSC-derived motor neurons subjected to genome editing at the aforementioned C9orf72 locus revealed both corrective (on-target) and detrimental (off-target) effects on electrophysiological function, relevant to ALS pathology.⁶⁷

Much of the focus over the past decade has been in developing these microphysiological systems to model phenotypic features of human physiology and disease. Considerably less effort has been invested in applying genome editing approaches to these systems. Many of the published studies have conducted the genome editing steps in precursor cells or iPSCs, rather than in the tissue-like constructs themselves. The size and complexity of intact microphysiological systems poses physical barriers to particle delivery, and these systems lack natural conduits such as blood vessels that might facilitate delivery of genome editors deep into internal structures or across membranes. Given these limitations, it is important that the use of microphysiological systems be complemented with additional experimental systems that enable visualization of genome editors in vivo.

IV. Assessing delivery and incorporation by in vivo imaging

While microphysiological systems can be useful in assessing phenotypic effects, they still have limited ability to provide insight into many dynamic processes that occur in vivo, including local distribution, inflammation, and higher-order physiological functions. Non-invasive in vivo imaging occupies the ‘macro’ range of the biological scale, providing critical insight into the outcome of genome editing therapies after they are administered into a living subject. Such imaging modalities have emerged as important preclinical and clinical tools for tracking complex therapeutics, such as CAR T-cells and early gene therapies, in living organisms, addressing on- and off-target events, and are thus similarly appropriate for genome editing therapies¹³⁹. On the macroscale level, imaging can be used to evaluate the overall therapeutic outcome, such as myocardial contractility or kidney function, as well as assessments of inflammation and T-cell trafficking, and importantly also the correlation between the delivery and therapeutic outcome, e.g. tumor shrinkage. These approaches can furthermore be used during clinical trials to monitor agents immediately after administration or during follow-up to determine the durability of therapeutics in the body and their correlation with observed biological effects.

Ideally, these technologies also provide real-time biomarkers for editing events and off-target effects of genome editing. In the context of tracking therapeutics in vivo, ‘on-target’ refers to successful delivery and incorporation into the organ or cell population that requires treatment, whereas ‘off-target’ refers to uptake elsewhere in the body that is not relevant to the disease being treated. These technologies must first be developed and demonstrated in animal model systems. Once established, clinical trials of genome editing therapeutics in humans can incorporate such assays for real-time readouts related to safety and efficacy.

Clinical genome editing is conducted either ex vivo in cells prior to administration, or in vivo by direct administration of editing agents into the body (Figure 3). Two major strategies for labelling therapeutics exist: direct labelling incorporates molecular tracers into the therapeutic agent, for example, radioactive isotopes or magnetic nanoparticles; and indirect labelling incorporates imaging reporter genes, for instance, by engineering cells or multi-cistronic vectors to express fluorescent proteins.

Figure 3. Tracking gene editing therapies with in vivo imaging.

Genome editors can be introduced into target cells through various methods, including viral vectors and DNA-, mRNA- or RNP-based carriers such as lipid-based nanoparticles (LNP), polymeric nanoparticles (PNP) or virus-like particles/enveloped delivery vehicles (VLP/EDV). Imaging agents/radiolabels can be directly incorporated into delivery vehicles, or cells edited ex vivo, prior to administration. Alternatively, indirect labelling can be achieved by integrating imaging reporter genes into viral or nonviral vector constructs. Non-invasive molecular imaging techniques can then be employed for short- or long-term tracking, depending on the chosen labelling strategy.

Direct labelling has advantages of quantitation and sensitivity and in some cases the imaging agent can be loaded into cells at high concentration. However, in proliferating populations such as CAR T-cells, this strategy may be limited to short-term use, due to dilution of the imaging label or rapid decay of radioactive tracers such as ⁶⁴Cu or ⁸⁹Zr ^140–143. By contrast, transduction with reporter genes allows for stable integration and expression, permitting long-term tracking of engineered cells or editing agents. Such reporter genes can be incorporated into CRISPR editing vectors, such as lentiviral vectors or AAV serotypes. However, indirect labelling can be more cumbersome than direct labelling, as gene-edited cells are now carrying a dual transgene (the reporter gene and the transgene) or a transient radiolabel¹⁴⁴. In addition to providing experimental challenges, a dual vector imaging approach may require additional regulatory oversight if existing vectors in translational development are modified. Specialized scanning methodologies may also be developed to directly image therapeutics without exogenous labelling¹⁴⁵.

A suite of preclinical and clinical imaging systems are used, each with specific features (Table 3). MRI (magnetic resonance imaging), CT (computed tomography), SPECT (single photon emission computed tomography) and PET (positron emission tomography) enable human imaging up to the scale of the whole body^146,147. Ultrasound and photoacoustic imaging can also be performed in patients, and are relatively inexpensive and more accessible within the healthcare system, but provide limited fields of view and applications, and are thus more relevant to imaging specific organs, for instance after local injection of genome editors. Fluorescence and bioluminescence imaging can be used for imaging small and large animals (pre-clinical) up to the scale of the whole body. MPI (magnetic particle imaging) small animal scanners exist, and clinical MPI scanners are being developed with or without hand-held devices¹⁴⁸.

Table 3.

Available molecular imaging modalities for tracking editors and edited cells

Device	Detection mode	Direct labelling			Reporter gene			Cell quantitation	Resolution	Refs
		Example	Sensitivity	Clinical use (potential)	Example	Sensitivity	Clinical use (potential)
BLI	Hot spot	n/a	n/a	No (No)	Firefly luciferase/D-luciferin	+++	No (No)	−	+	142,156,188,220,221
MPI	Hot spot	Iron oxide nano-particles	++	No (Yes)	No	n/a	No (Yes)	+	+	176,177,222,223
CT	Contrast	Gold particles	+	No (Yes)	No	n/a	No (No)	+	+++	183,189,224
US	Contrast	Micro-bubbles	+	Yes	ARGs	+	No (Yes)	−	++	225
PAI	Contrast	ICG	+	No (Yes)	OATP1/ICG	+	No (Yes)	−	++	226
SPECT	Hot spot	[99mTc]Tc-HMPAO	+++	Yes	NIS/[99mTc]TcO4−	++	Yes	+	++	227
PET	Hot spot	89Zr	+++	Yes	NIS/[18F]TFB	+++	Yes	+	+	140,180,188,228
MRI	Contrast (1H)	T1: Gd T2: SPIO	+/+++	Yes	OATP1/Gd-EOB-DTPA	+	Yes	−/+	+++	156,169,172,188,229,230
	Hot spot (19F)	19F-PFC	+	Yes	No	n/a	No (Yes)	+	+

¹H, hydrogen; ^18/19F, fluorine; ^99mTc-HMPAO, technetium-99m hexamethyl propylenamine oxine; ARG, acoustic reporter gene; CT, computed tomography, FDG, fluorodeoxyglucose; Gd, gadolinium; Gd-EOB-DTPA, gadolinium ethoxybenzyl diethylenetriamine pentaacetic acid; ICG, indocyanine green; BLI, bioluminescence imaging; MPI, magnetic particle imaging; MRI, magnetic resonance imaging; n/a, not applicable; NIS, sodium iodide symporter; OATP, organic anion transporting polypeptide; PET, positron emission tomography; PFC, perfluorocarbon; SPECT, single photon emission computed tomography; SPIO, superparamagnetic iron oxide; TFB, tetrafluoroborate; T1/T2, tissue relaxation time 1 (T1)- or 2 (T2)-weighted MRI; US, ultrasound; PAI, photoacoustic imaging. Adapted from Ref.¹³⁹.

When using tracers, as with MPI, SPECT or PET, absolute quantification is possible, as the detected tracer directly corresponds to the total quantity of labelled cells. With contrast agents, as in MRI and ultrasound, this is less straightforward. With optical imaging, which includes both bioluminescence and fluorescence imaging, the signal is attenuated in deeper tissues, with the exact depth of localization often not known, although some systems can perform 3D optical imaging with somewhat improved localization. MRI and CT are the modalities that can provide overall anatomical structures. Hence, tracer-based modalities such as PET require a combination with MRI or CT to provide the essential anatomical context in detail. Each modality is discussed in greater detail below.

Bioluminescence imaging

Bioluminescence imaging is a versatile, sensitive and relatively inexpensive modality that is routinely used for high-throughput imaging in preclinical models. Numerous orthogonal luciferase/luciferin combinations exist that also enable dual-bioluminescence imaging to monitor multiple cell types or biological events in individual animals¹⁴⁹. Bioluminescence imaging has been used frequently to track the effects of genome editing in various cell types in vivo. For example, the effect of knocking out genes with CRISPR–Cas9 editing on cancer growth and metastasis in mice can be tracked with bioluminescence imaging^150,151. Similarly, bioluminescence imaging has been applied to monitor tumour formation following in vivo delivery of genome editors¹⁵² and to track the efficacy of CRISPR-based gene-editing cancer therapies^153,154.

Bioluminescence imaging can be used alone, or alongside other clinically relatable reporter genes for multi-modal tracking of cells that have been CRISPR-edited at safe-harbour loci to improve safety^155,156. In addition, bioluminescence imaging has been used to track pathogens edited with CRISPR^157,158 or to detect editing events after delivery of CRISPR-based systems¹⁵⁹. Finally, luciferase reporter transgenic mouse models have been developed to evaluate efficacy and safety of genome editors and delivery systems for various genome editing therapeutics^160,161 or designed for tracking of specific cell types¹⁶². Bioluminescence imaging is compatible with both firefly luciferase, which has had extensive applications, as well as red fluorescence with nanoparticles^163,164. Recently, a fluorescent orange calcium sensor compatible with bioluminescence imaging has been developed, which can be used to monitor calcium transients in internal organs such as the liver or heart¹⁶⁵. Bioluminescence imaging after fetal administration of an HIV-1-derived dual fusion lentiviral vector using an intraperitoneal approach in early gestation showed persistence for over 12 years¹⁶⁶. These studies indicate that bioluminescence imaging is highly effective for monitoring long-term gene expression in rhesus monkeys and assessing cell trafficking when combined with PET.

MRI and MPI

Both direct and indirect labelling methods exist for MRI to allow for short-term and long-term cell tracking, respectively. For cells that have been edited ex vivo, and similar to PET using radiolabelling techniques, direct labelling can be applied before injection into an animal or human. This includes the use of superparamagnetic nanoparticles to magnetize labelled cells causing them to distort the magnetic field, and providing “negative contrast”¹⁶⁷. An alternative direct labelling method is to use ¹⁹F as a tracer that can be detected with ¹⁹F MRI directly, as demonstrated for imaging CAR T-cells^168,169. An added benefit of using fluorine as a label is that it can be used to measure local oxygen levels, which is relevant for tumour therapy¹⁷⁰.

MRI reporter genes have been harnessed to track edited cells and delivery vectors. The OATP1 transporter, which selectively uptakes the Gd³⁺-based positive contrast agent Gd-EOB-DTPA, has been used to track CRISPR-edited metastatic cancer cells in vivo¹⁵⁶. Lysine rich protein (LRP), which is a genetically engineered chemical exchange saturation transfer (CEST) MRI reporter¹⁷¹, has been used to track AAV gene therapy vectors in vivo¹⁷². This approach led to the discovery that CEST MRI can be used to detect the large number of surface lysine, serine, and threonine residues on therapeutic AAV vectors without the need for a reporter gene¹⁷³. CEST MRI can be used to detect cellular responses, such as tumor cell apoptosis in response to viral therapy, and this approach is enhanced by applying deep learning tools¹⁷⁴. Similarly, new MRI approaches have focused on tracking unlabelled stem cells^145,175, which may accelerate clinical translation and could be used to track genome-edited stem/progenitor cells in the future.

MPI can be applied for in vivo (stem) cell tracking¹⁷⁶. In MPI, the same types of SPIO nanoparticles used as contrast agents for MRI are used as tracer agents to provide a ‘hot spot’ signal with true cell quantification and without tissue background signal¹⁷⁷, enabling in vivo cytometry on a time scale of minutes. Similar to MRI, the biodistribution of CART-cells has recently been assessed with MPI¹⁷⁸, and MPI was used to track the intra-arterial delivery of ALSpatient-derived, genome-corrected iPSCs in transgenic animal models of ALS¹⁷⁹.

Nuclear imaging

Nuclear imaging tools such as PET and SPECT offer high sensitivity and clinical relevance. The amount of radionuclide-labelled drug required for these imaging tools is well below pharmacological levels, thus translation into patients can be more readily accomplished compared to other clinical imaging modalities (for example, MRI). PET can provide 3D quantitative images deep inside the body of radiotracers in the subnanomolar concentration range^180–182. In addition, new transformative total-body PET scanners are now commercially available for human use with an extended axial field-of-view that allows the entire body (all tissues and organs) to be imaged simultaneously¹⁴⁷.

An AAV vector construct (SaCas9/PCSK9/HSV-sr39TK) has been shown to be highly reproducible for identifying edited somatic cells in the liver after administration. In proof-of-concept studies, when fetal rhesus monkeys were administered this vector prenatally using an intrahepatic, ultrasound-guided approach, imaging in utero or infants postnatally consistently showed that editing in the liver was sustained across gestation and postnatally, using methods similar to those established previously using other vectors^164,183. Methods have also recently been developed to label the genome editor LbCas12a (Cpf1) with the PET radionuclide ⁸⁹Zr without loss of function, to track its distribution following packaging in LNP¹⁸⁴.

Prior studies have shown that stem or progenitor cells were effectively radiolabelled with radioactive copper (⁶⁴Cu) for PET, and indicated that each cell type has a different radiolabelling efficiency and must be tested to optimize conditions^141,143. Further studies improved contrast-to-noise imaging for cell trafficking when using ⁸⁹Zr radioimmunoconjugates to identify engrafted cells post-transplant¹⁴⁰. In other studies, ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) PET imaging was used to monitor glucose metabolism in osteosarcoma tumours in mice that were edited to knock out P2rx7 ¹⁸¹. CRISPR has been used to knock out Cxcr4 prior to establishing leukaemia xenografts in mice and CXCR-targeted ⁶⁸Ga-Pentixafor PET imaging was used to visualize differences in naive and edited tumours¹⁸⁵.

Two groups have applied PET reporter gene imaging to visualize edited iPSCs that were then differentiated into cardiomyocytes. When rhesus macaque-derived iPSCs were edited to express rhesus sodium/iodide symporter (NIS) from the AAVS1 safe harbour locus, and derived cardiomyocytes were injected into the myocardium of immunodeficient mice following injury, PET imaging with NIS-targeted ¹⁸F-TFB successfully visualized edited cells for up to 10 weeks¹⁸⁶. Similarly, multimodal reporter imaging with bioluminescence imaging and ¹⁸F-FHBG PET was used to monitor the survival of cardiomyocytes derived from iPSCs edited to express multiple reporter genes for fluorescence, bioluminescence imaging, and PET at the AAVS1 locus. Imagingafter injection into infarcted myocardium of rats was performed, as well as ¹⁸F-FDG PET to monitor myocardial glucose metabolism¹⁸⁷. Simultaneous PET/functional MRI was includedto analyse the consequences of AAV-mediated VMAT2 knock-down in the substantia nigra pars compacta in adult rats¹⁸². Use of the NIS reporter gene, as well as bioluminescence imaging and MRI reporters, is currently being explored to track edited CAR T-cells, similar to approaches that have been developed for tracking tumour cells¹⁸⁸.

CT and Ultrasound

CT has been used to evaluate a gene editing strategy for lung diseases. Following intratracheal aerosol of Cy5-labeled peptide cargo in young rhesus monkeys, CT was used to assess delivery into the lung lobes immediately after administration, which ultimately showed ~5% editing¹⁸⁹. The same approach was used to deliver an AAV5 vector harbouring Streptococcus pyogenes Cas9 ( Cas9), resulting in approximately ~8% editing¹⁸³. Ultrasound has been used primarily to guide the injection of viral or nonviral vectors into specific anatomical sites such as the liver and kidney, and when using a prenatal administration approach.

V. Conclusions and Outlook

Genome editing therapies are increasingly advancing into clinical studies, and systems for predicting and monitoring the biological effects of these therapies are needed to guide these translational efforts. The development of k-abe within months of its intended patient’s birth would seem to herald a new age for on-demand genome editing therapies to treat a variety of hereditary enzymatic disorders. This singular treatment, however, was only made possible through around-the-clock efforts by a large, highly motivated, and collaborative team of academia and industry investigators, across many different institutions¹⁷. It is now appropriate to consider how this development workflow can be made more lean and scalable, to maximize benefit for patient populations. Generating a mouse model to demonstrate efficacy, for instance, could conceivably be circumvented through use of donor-derived human microphysiological systems, possibly in combination with other new approach methodologies (NAMs) such as in silico prediction. Such techniques will become more powerful over time, enhanced by machine learning approaches that extrapolate and convert between orthogonal experimental systems and real-world clinical outcomes.

The human-focused toolkit encompassing the micro- to macro-scale workflow described herein fits naturally within this framework. The other key ingredient is cooperative and collaborative efforts unified around the unmet need, which will serve to accelerate the process of therapy development and connect the dots between these tools. If datasets associated with translational efforts are widely shared, it will empower the NAMs framework, reduce risk, and lower barriers for subsequent studies. An advantage of consortium efforts such as the SCGE Biological Effects Initiative is that preliminary data and technical difficulties can be openly discussed and publicized through sharing mechanisms such as the consortium’s Toolkit (BOX 2). Discussions of the working group have revealed significant challenges for technology and therapy development.

Box 2. The SCGE Toolkit.

The SCGE Toolkit (https://scge.mcw.edu/toolkit/) catalogs, integrates, and curates data generated by the main initiatives of the SCGE Consortium²⁹, and provides access to protocols, models, and approaches to capture both the intended and unexpected outcomes of gene editors and their delivery systems. Toolkit users can identify the detailed gene editing and delivery system reagents used in these preclinical studies, which include human cell and organoid models, along with protocols for quantitative and qualitative assessment of editing and other biological outcomes. The site provides relevant example data and images and enables dynamic interaction with a variety of small- and large-scale genome editing datasets. For instance, a user interested in editing the kidneys may query how different AAV serotypes transduce human kidney organoids, and discover that AAV2/2 efficiently transduces proximal tubules (LTL⁺), but also induces substantial DNA damage in these structures. Entries are categorized by project, each containing one or more experiments, which can be rendered interactively or downloaded as raw data. To make consortium data available for public use, including unpublished data, the Toolkit has defined relevant metadata related to the component model systems, editors, guides, vectors and delivery systems in each experiment to standardize submitted data for display and comparison.

Among such challenges, inefficiencies in genome editing delivery and editing remains a substantial barrier, not only to the translation of candidate therapeutics, but also for the development of preclinical surveillance systems. Editing efficiency remains low across biological systems in part owing to low uptake of the genome editor by cells²⁸. Cas9, for example, is ~160 kDa in molecular weight, which is a considerable barrier for transmembrane delivery regardless of the vector used. For viral transduction systems, production of Cas9 (160 kDa) may be less efficient compared to smaller cargos such as GFP (27 kDa) that are commonly used to benchmark transduction, which can lead to erroneous over-estimates of Cas9 transduction. In microphysiological systems, even if genome editors can be delivered at detectable levels, the resultant gene editing events have remained relatively low. This likely reflects the inherent limitations of genome editing technologies, which are far less than 100 % efficient in native biological systems¹. As cell cycle and metabolism are lineage-specific, DNA surveillance and repair processes in human primary or iPSC-derived lineages may differ from those of immortalized cell lines, resulting in lower editing efficiency^79,190. Highly proliferative cells may be more susceptible to DNA repair outcomes, including unwanted deletions and chromosomal rearrangements^191,192. A rule of thumb is to conduct studies in systems that resemble the on-target organ lineage as closely as possible²⁸.

Substantial variability between different batches of cultures, viral and non-viral vectors, and different editing tools, contributes to sub-optimal reproducibility between experiments, and can confound the interpretation of the results. Hence, rigorous quality control criteria are needed, including baseline characteristics that can be used for normalization across different experiments^193,194. In heterocellular systems, it is difficult to know which cell types are being edited specifically, and which are undergoing off-target events. Genetically-encoded fluorescence-based reporters of genome editing activity (for example, ‘traffic light’ reporters) can be combined with immunofluorescence analysis to reveal cell types that exhibit editing activity^195,196. However, even when such tools are available, they provide measurements of editing events only at the reporter locus, which must be relied upon as a surrogate for the actual human target locus. In addition, the genomic target in animals may be different from the human target sequence, requiring humanizing transgenic animals at the target site. Similarly, no current technology for nominating off-target mutations can distinguish effects on different cell types within a mixed cell population. Moreover, there remains no ready assay in either microphysiological systems or animal models to assess the chronic risk of tumorigenesis resulting from such mutations, which may take years to develop in patients in vivo^62–65.

In many respects, surmounting these are robustness challenges with preclinical systems (for example, organoids) would bode well for clinical translation of a genome editing therapeutic. Conversely, if a therapeutic approach is unsuccessful in meeting these challenges in a preclinical setting, it will likely require further optimization before advancing to the clinic. For some genome editing treatments, for instance those that seek to reduce levels of a specific protein in the blood, clinical studies may have ready access to biomarkers of target engagement⁷. But for many clinical studies, whether the genome editor has the desired effect on its molecular target is more difficult to discern, because the target organs are inaccessible in human subjects and therefore lack appropriate biomarkers. In these cases, the ability to demonstrate efficacy and mechanism of action in preclinical systems, and the development of in vivo imaging tools that bridge the gap into the clinic, will be particularly important.

Methods to measure editing outcomes, microphysiological systems, and in vivo non-invasive imaging tools are complementary technologies that provide assessments of on-target and off-target effects of somatic cell genome editing at different biological scales. Opportunities exist for the combinatorial use of the different types of technologies. Measurements of editing outcomes can be performed alongside phenotyping in microphysiological systems or in vivo imaging applications, to provide a useful benchmark for observed phenotypic effects. Similarly, in vivo imaging tools can be applied not only to animal models, but also to microphysiological systems, at either the cellular level with fluorescence microscopy or at the bulk population level using bioluminescence imaging, to provide important insights into the efficiency of different viral and non-viral delivery systems, to evaluate the kinetics of transgene expression from these, and to enable indirect visualization of cells that have been actively edited with different types of editors using activatable optical reporter genes¹⁶⁵. This approach provides a pathway from micro- to whole-body imaging along the entire spectrum of genome editing development, from cultured cells to preclinical animal models to patients. The continued development of tools that assess the biological effects of genome therapeutics remains essential to realize the exciting potential of this innovative field.

Glossary

Bystander edit

An unintended on-target mutation, typically associated with base editing, which may be synonymous or non-synonymous

Editing efficiency

The ability of a genome editor to successfully undertake desired edits at the desired genomic location. In the context of this Review, ‘efficiency’ is used interchangeably with ‘effectiveness’, whereas ‘efficacy’ typically connotes a phenotypic impact of measurable size

In vivo imaging

A suite of non-invasive imaging modalities enabling live tracking of administered molecules and cells in living organisms, such as MRI or PET

Intended on-target edit

The corrective genome editing event, at the locus specified by the guide RNA, for which the editor has been designed, and which is expected to confer therapeutic benefit

Unintended on-target edit

An editing event at the locus specified by the guide RNA that introduces mutations that are not corrective and/or planned

Liver-on-a-chip

A microfluidic culture device that mimics features of hepatic structure and function in vitro

Microphysiological system

A complex culture system capable of modeling tissue-like structure or function in vitro or ex vivo, such as an organoid, organ-on-chip, or explant

Multi-cistronic vectors

Vectors contain IRES and/or 2A peptide sequences to allow for co-expression of multiple genes from a single transcript

Nomination

A process whereby off-target sites are putatively identified

Nephron

Each human kidney contains ~1 million intricately structured nephron subunits, which maintain homeostasis through blood filtration and urine production

Off-target edit

An editing event at a locus in the genome other than the locus for which the guide RNA was designed

Organoid

A multicellular structure grown in vitro, including multiple cell types organized in a pattern that resembles a bodily tissue

Prime editing

A method that utilizes reverse transcriptase to copy sequences into genomic DNA

Tubuloids

Expandable organoid cultures grown from adult kidney tissues or urine