Yang Liu

← Primer

HSC · Healing Source Code

编辑之后留下来的是谁——p53 如何在暗处筛选被改写的细胞 Who Is Left After the Edit — How p53 Quietly Selects the Cells We Rewrite

2026-07-25 · p53, 基因编辑安全性, 造血干细胞, 克隆动力学, DNA损伤应答

每一份基因编辑产品的报告上都有一个漂亮的数字:编辑效率百分之多少。这个数字是从那些还活着的细胞身上读出来的。可培养瓶并不是一个中立的容器——从核酸酶剪断 DNA 的那一刻起,细胞就开始被一套古老的机制分门别类:哪些停下、哪些死去、哪些照常分裂。执行这场分类的,是被称作”基因组守卫者”的 p53。于是有了一个让人不安的推论:一场以 p53 为裁判的淘汰赛,最容易通过的选手,恰恰是裁判管不住的那一个。

一次剪断,不只是一次剪断

用 CRISPR-Cas9 改写基因组,本质上是先制造一次事故——一个 double-strand break——再指望细胞把它修好。可细胞并不知道这次断裂是”治疗”;在它看来,DNA 断了就是 DNA 断了,与辐射、与化疗药造成的损伤同属一类。ATM 激酶被激活,p53 被稳定下来,p21 被转录,细胞周期停在检查点上;如果损伤看起来无法收拾,程序性死亡就会启动。这套 DNA damage response 存在的全部意义,就是不让带着未修复断裂的细胞继续分裂。

造血干细胞对这道防线的反应尤其重。2010 年,Milyavsky 等人在 Cell Stem Cell 上比较人的造血干细胞与下游祖细胞对 γ 射线的反应,发现干细胞修复双链断裂更慢、γH2AX 焦点消散得更迟,却更容易走向 p53 与 ASPP1 依赖的凋亡。同一篇论文还有一个后来反复被引用的观察:让 p53 失活确实能减少辐射后的凋亡、保住体内重建能力,但只有过表达 Bcl-2 的干细胞在自我更新上真正获益——p53 在这里除了执行凋亡,还独立地维持着自我更新。这一点后面会变得很要紧。

2018 年 6 月的两篇论文

真正把”编辑本身构成筛选”摆上台面的,是 2018 年 6 月 11 日同日上线于 Nat Med 的两篇论文。

Haapaniemi 等人在永生化的人视网膜色素上皮细胞里做了一件朴素的事:看 Cas9 编辑之后细胞群体发生了什么。他们报告,编辑诱发了 p53 介导的 DNA damage response 与 cell-cycle arrest,其结果是对 p53 通路完好的细胞产生了负向选择;抑制 p53 既能消除这一反应,又能提高同源重组修复的效率。文章结尾的建议今天读来仍然锋利:开发基于 CRISPR-Cas9 的细胞疗法时,p53 的功能状态应当被监测。

同一天,Novartis 的 Ihry 等人在人多能干细胞里补上了另一半图景。当他们把 indel 效率推到平均 80% 以上,一个此前被低效率掩盖的现象浮了出来:Cas9 造成的 double-strand break 对多能干细胞有明显毒性,能杀死其中的大多数,而这种毒性依赖 TP53——P53 完好的细胞被精确改造的效率因此被严重压低。他们同样提醒,多能干细胞本就可能获得 P53 突变,用 CRISPR 改造过的这类细胞做替代治疗应当谨慎。

两篇合起来是一个闭环:编辑对 p53 完好的细胞是有代价的,对 p53 失能的细胞则小得多。任何代价不均等的处理施加在异质群体上,都是一次选择。

筛选不是比喻

两年后,Enache 等人在 Nat Genet 上把这个推论做成了可以计数的事实。他们比较 165 对人类癌细胞系与其稳定表达 Cas9 的衍生系,发现仅在 TP53 野生型的细胞系里,引入 Cas9 本身就上调了 p53 通路,γH2AX 焦点也随之增多。随后他们对其中 42 对做了深度靶向外显子测序:2 对在 Cas9 系里新出现了非沉默的 TP53 突变,另有 2 对里既有的 TP53 突变显著扩张;在”引入 Cas9 后倾向于新增突变”的基因排序里,TP53 位列前 4%。最干净的证据来自竞争实验:把 TP53 缺失的细胞与同源野生型按 1:8 混合,缺失一方本就会缓慢占优,而携带 Cas9 时占优得更快;换成 ARID1A 或 FBXW7 缺失,这种加速就消失了。

但这些实验的地基必须说清楚:用的是永生化癌细胞系,Cas9 长期稳定表达,群体里往往本就存在带 TP53 突变的亚克隆。这与临床级工艺把高保真 Cas9 蛋白以 ribonucleoprotein 形式电转进原代细胞,蛋白只短暂存在便被清除,是很不一样的处境。

到了造血干细胞里,故事变了形

那么在真正要用来治病的细胞里,情况如何?2019 年,San Raffaele 的 Schiroli 等人用优化过的锌指核酸酶和 Cas9,在人的造血干/祖细胞里制造一个或多个 double-strand break,再以单细胞分辨率追踪 DDR 焦点、细胞周期与转录反应。结论是:即便只制造一个断裂,p53 通路激活也是所有亚群——包括最原始的那一群——中占主导的反应;而当断裂负荷升高、或用 AAV6 递送同源修复模板时,激活会叠加,压制被编辑细胞的增殖、产量与 engraftment。好在 DDR 负荷较低时这种损害是可逆的,也可以被短暂抑制 p53 所抵消——他们用的 GSE56,正是一个截短的显性负性 p53 片段,来自 Milyavsky 那篇论文。

次年,同一支团队在 Nat Biotechnol 上用条形码克隆追踪把代价量化了。Ferrari 等人发现,编辑激活 p53 之后,人鼠嵌合体内造血干细胞的克隆库明显收窄,但成功植入的被编辑克隆仍保有多系分化与自我更新能力;短暂抑制 p53 可以让移植物重新变回多克隆。

这两篇论文把问题的形状改变了。在原代造血干细胞里,p53 施加的压力主要不表现为”选出 TP53 突变体”,而表现为减员:一部分克隆干脆没能挤进移植物。这是一种功能性的筛选,而不是(至少不主要是)遗传性的筛选;它的后果不写在编辑效率里,而写在克隆多样性里。

有没有人直接去找过 TP53 突变

有。2022 年,Cromer 等人在 Nat Commun 上做了一次针对性的搜寻:三位供者的原代造血干/祖细胞,用高保真 Cas9 蛋白分别靶向 AAVS1、HBB 与 ZFPM2,在第 4 天与第 10 天取基因组 DNA,以临床肿瘤诊断级别的超深度测序覆盖 523 个癌症相关基因的外显子,并在 AAVS1 条件下补做全外显子与全基因组测序——其中全基因组测序只在一位供者上做。结果是没有检测到致瘤变异的引入或富集。他们同时给出了对细胞系研究的解读:那些结果依赖于起始群体中已经存在的 p53 突变亚克隆,而健康供者的原代细胞里本不该有。

这是目前最直接的阴性证据,边界也同样清楚:三位供者、ex vivo 十天、有限的检测下限,既没覆盖移植之后在人体内数年尺度上发生的事,也没覆盖本就带着克隆造血的患者。

压力是真的,只是它作用在别处

与其说造血干细胞逃过了 p53 的选择,不如说这里的压力换了一种施加方式,而且高度依赖工艺。2022 年,Ferrari 等人查明了 AAV 为何格外碍事:病毒基因组及其片段在细胞里滞留得远比预想的久,MRN 复合物被招募到 AAV 的反向末端重复序列上,把 p53 依赖的 DDR 持续拉高;改用整合缺陷型慢病毒载体递送模板,DNA 负荷更低、DDR 更短。2025 年,Conti 等人报告 CRISPR-Cas9/AAV6 编辑会诱发一种由 p53 与 IL-1/NF-κB 共同驱动的持续类衰老反应,在长期移植实验中限制移植物大小与克隆多样性;同年,Araki 等人在人鼠异种移植模型里发现,移植后第 1–14 天连续给予 G-CSF,会通过增殖压力放大被编辑细胞早期的 p53 反应,反而妨碍植入,而把给药推迟到第 5 天就能减轻。p53 反应的强度不是一个固定的物理常数,而是被递送方式、模板载体乃至移植后用药共同调节的变量。

于是”短暂抑制 p53”看上去像一个顺理成章的解法,而它恰恰是这条线索上最需要小心的一步。Milyavsky 那篇论文早已提示,p53 在人的造血干细胞里并不只管凋亡,还独立地维持着自我更新;把守卫按住的同时,你也按住了它的另一份工作。Schiroli 等人在短期检测里没有看到核型异常或突变负荷升高,但那是体外、两到四周尺度的观察,而且他们自己就写明了这组阴性结果的限度:核型分析灵敏度有限,深度测序也只覆盖了靶向的那部分基因组——“没检测到”与”没发生”,在基因毒性这件事上从来不是同一句话。

不制造断裂,就能绕过去吗

部分可以。2021 年,Newby 等人在 Nature 上用腺嘌呤 base editor 把镰刀型贫血的 HBB^S^ 等位基因转写成良性的 Makassar 变异,在患者来源的造血干/祖细胞里达到约 80% 的转换,并明确记录到 base editing 避开了 Cas9 核酸酶处理后所见的 p53 激活与较大片段缺失。

但”避开 double-strand break”不等于”隐形”。2022 年,Li 等人在 Nat Commun 上报告,在人多能干细胞里,prime editing 与胞嘧啶 base editing 的效率同样受 p53 限制——共递送显性负性 p53 片段就能显著提高效率;单链切口一样会被感知。2026 年 Lei 等人在 Cell Stem Cell 上给出的对照则说明风险是换了形状而不是消失:在一个小鼠 Unc13d 模型里,超活性胞嘧啶 base editor 的脱靶谱比 CRISPR-Cas9 更宽、落点更分散,只是易位的总频率比 Cas9 低一个数量级;而 Cas9 在靶点周围造成的大片段缺失与倒位则明显更多。少了断裂,不自动等于少了基因毒性,它只是把风险换了一副面孔。

临床上看到了什么,又看不到什么

到目前为止,核酸酶编辑产品的已发表随访里没有出现这个机制所预言的事故。2024 年,Frangoul 等人在 N Engl J Med 上报告 exa-cel 治疗重度镰刀型贫血的三期结果:44 名接受输注的患者,中位随访 19.3 个月,全部实现中性粒细胞与血小板植入,未发生癌症。

但”没看到”与”不存在”之间的距离,恰好等于随访时间:髓系肿瘤是以年计的事件,而这类研究的随访还以月计。人体尺度上真正让人无法轻松的证据来自另一个方向。2020 年,Bolton 等人在 Nat Genet 上分析癌症治疗如何重塑克隆造血的适应度版图,发现放疗、铂类与拓扑异构酶 II 抑制剂会优先选择 DNA damage response 基因——TP53、PPM1D、CHEK2——的突变克隆,连续采样直接看到了这些克隆压过其他克隆的过程。这不是编辑,但它证明了一件事:人的造血系统面对 DDR 方向的选择压力时,确实会用富集 p53 通路缺陷克隆的方式作答。而在基因治疗流程里施加这类压力的,除了核酸酶,还有清髓 conditioning 本身。

临床上有一处相邻的观察。美国医学遗传学与基因组学学会 2024 年的一份治疗学通报在讨论 Casgevy 与 Lyfgenia 时记录到,已有少数患者在接受 Lyfgenia——一款慢病毒基因添加产品,而非编辑产品——之后出现骨髓增生异常综合征与急性髓系白血病,深入调查排除了插入突变致病;通报推测,这一风险与该人群中被加速的克隆造血有关,而这正是它提到”已有人建议在基因治疗前筛查白血病前期克隆”的理由。至于已经接受治疗的人,这正是 Bolton 那类连续采样能回答、而编辑效率回答不了的问题:读数在移植后外周血中 TP53 等克隆造血基因随时间的消长里。

一个仍然开着的问题

把这条线索完整摆出来,会看到一个罕见的形态:机制上极有说服力、在细胞系里已被明确证实、在原代造血干细胞里被反复观察到其上游反应,却至今没有在编辑产品上落成一次可指认的事故。它既不能被当成危言耸听搁置,也不能被当成已发生的灾难叙述。它真正改变的,是我们该如何读那个漂亮的数字——编辑效率是幸存者报出来的成绩,它不告诉你有多少克隆在报数之前就出局了,更不告诉你留下来的那些是因为修得好,还是因为它们对损伤根本不敏感。要回答这一层,只能去看克隆的身份与数量随时间怎么变。

参考文献

  1. Milyavsky M, Gan OI, Trottier M, et al. A distinctive DNA damage response in human hematopoietic stem cells reveals an apoptosis-independent role for p53 in self-renewal. Cell Stem Cell. 2010;7(2):186-197. DOI
  2. Haapaniemi E, Botla S, Persson J, Schmierer B, Taipale J. CRISPR-Cas9 genome editing induces a p53-mediated DNA damage response. Nat Med. 2018;24(7):927-930. DOI
  3. Ihry RJ, Worringer KA, Salick MR, et al. p53 inhibits CRISPR-Cas9 engineering in human pluripotent stem cells. Nat Med. 2018;24(7):939-946. DOI
  4. Enache OM, Rendo V, Abdusamad M, et al. Cas9 activates the p53 pathway and selects for p53-inactivating mutations. Nat Genet. 2020;52(7):662-668. DOI
  5. Schiroli G, Conti A, Ferrari S, et al. Precise gene editing preserves hematopoietic stem cell function following transient p53-mediated DNA damage response. Cell Stem Cell. 2019;24(4):551-565.e8. DOI
  6. Ferrari S, Jacob A, Beretta S, et al. Efficient gene editing of human long-term hematopoietic stem cells validated by clonal tracking. Nat Biotechnol. 2020;38(11):1298-1308. DOI
  7. Cromer MK, Barsan VV, Jaeger E, et al. Ultra-deep sequencing validates safety of CRISPR/Cas9 genome editing in human hematopoietic stem and progenitor cells. Nat Commun. 2022;13(1):4724. DOI
  8. Ferrari S, Jacob A, Cesana D, et al. Choice of template delivery mitigates the genotoxic risk and adverse impact of editing in human hematopoietic stem cells. Cell Stem Cell. 2022;29(10):1428-1444.e9. DOI
  9. Conti A, Giannetti K, Midena F, et al. Senescence and inflammation are unintended adverse consequences of CRISPR-Cas9/AAV6-mediated gene editing in hematopoietic stem cells. Cell Rep Med. 2025;6(6):102157. DOI
  10. Araki D, Chen V, Redekar N, et al. Post-transplant G-CSF impedes engraftment of gene-edited human hematopoietic stem cells by exacerbating p53-mediated DNA damage response. Cell Stem Cell. 2025;32(1):53-70.e8. DOI
  11. Newby GA, Yen JS, Woodard KJ, et al. Base editing of haematopoietic stem cells rescues sickle cell disease in mice. Nature. 2021;595(7866):295-302. DOI
  12. Li M, Zhong A, Wu Y, et al. Transient inhibition of p53 enhances prime editing and cytosine base-editing efficiencies in human pluripotent stem cells. Nat Commun. 2022;13(1):6354. DOI
  13. Lei L, Kaufmann MM, Lao J, et al. Genotoxicity profiling reveals distinct platform- and cell type-specific effects in therapeutic gene editing for genetic hyperinflammation. Cell Stem Cell. 2026;33(6):930-944.e5. DOI
  14. Frangoul H, Locatelli F, Sharma A, et al. Exagamglogene autotemcel for severe sickle cell disease. N Engl J Med. 2024;390(18):1649-1662. DOI
  15. Bolton KL, Ptashkin RN, Gao T, et al. Cancer therapy shapes the fitness landscape of clonal hematopoiesis. Nat Genet. 2020;52(11):1219-1226. DOI
  16. Lesmana H, Kim SY, Corado AM, et al. Casgevy (exagamglogene autotemcel) and Lyfgenia (lovotibeglogene autotemcel) for individuals 12 years and older with sickle cell disease (SCD) and recurrent vaso-occlusive crises (VOC): A therapeutics bulletin of the American College of Medical Genetics and Genomics (ACMG). Genet Med Open. 2024;2:101875. DOI

Every report on a gene-editing product carries one handsome number: the editing efficiency, in percent. That number is read off the cells that are still alive. But a culture flask is not a neutral container — from the moment a nuclease cuts the DNA, the cells begin to be sorted by an ancient machinery: which ones halt, which ones die, which ones go on dividing as usual. What carries out that sorting is p53, the so-called guardian of the genome. Which leads to an unsettling inference: in an elimination round refereed by p53, the contestant most likely to get through is precisely the one the referee cannot control.

One cut is never just one cut

Rewriting a genome with CRISPR-Cas9 means, in essence, first staging an accident — a double-strand break — and then counting on the cell to repair it. But the cell has no idea that this particular break is “treatment”; as far as it is concerned, broken DNA is broken DNA, of a piece with the damage inflicted by radiation or by chemotherapy drugs. ATM kinase is activated, p53 is stabilized, p21 is transcribed, the cell cycle stops at a checkpoint; and if the damage looks beyond repair, programmed death begins. The entire point of this DNA damage response is to keep cells carrying unrepaired breaks from dividing further.

Hematopoietic stem cells lean on that defense particularly hard. In 2010, Milyavsky and colleagues reported in Cell Stem Cell a comparison of human hematopoietic stem cells and their downstream progenitors after γ-irradiation: the stem cells rejoined double-strand breaks more slowly, their γH2AX foci persisted longer, and yet they were more prone to p53- and ASPP1-dependent apoptosis. The same paper contains an observation that has been cited over and over since: inactivating p53 did reduce apoptosis after irradiation and preserved in vivo repopulating function, but only stem cells overexpressing Bcl-2 truly gained in self-renewal — meaning that p53 here, beyond executing apoptosis, independently sustains self-renewal. That point becomes important later.

Two papers in June 2018

What really put “editing itself constitutes selection” on the table were two papers posted the same day, 11 June 2018, in Nat Med.

Haapaniemi and colleagues did something plain in immortalized human retinal pigment epithelial cells: they looked at what happened to the cell population after Cas9 editing. Editing, they reported, induced a p53-mediated DNA damage response and cell-cycle arrest, the consequence of which was a selection against cells with a functional p53 pathway; inhibiting p53 both abolished that response and raised the rate of homologous recombination. The recommendation at the end of the paper still reads sharply today: p53 function should be monitored when developing cell-based therapies that use CRISPR-Cas9.

That same day, Ihry and colleagues at Novartis supplied the other half of the picture in human pluripotent stem cells. Once they pushed indel efficiency above 80% on average, a phenomenon previously masked by low efficiency surfaced: the double-strand breaks made by Cas9 are toxic to pluripotent stem cells and kill most of them, and that toxicity is TP53-dependent — so the efficiency with which P53-intact cells can be precisely engineered is severely depressed. They likewise cautioned that pluripotent stem cells can acquire P53 mutations on their own, and that cell replacement therapy using CRISPR-engineered cells of this kind should proceed with caution.

Taken together the two papers close a loop: editing carries a cost for cells with intact p53, and a much smaller one for cells in which p53 has failed. Any treatment whose cost falls unevenly across a heterogeneous population is an act of selection.

Selection is not a metaphor

Two years later, Enache and colleagues turned that inference into something countable in Nat Genet. Comparing 165 pairs of human cancer cell lines with their Cas9-expressing derivatives, they found that introducing Cas9 upregulated the p53 pathway — and increased γH2AX foci — specifically in TP53 wild-type lines. They then ran deep targeted exon sequencing on 42 of those pairs: non-silent TP53 mutations newly emerged in the Cas9 line of 2 pairs, and pre-existing TP53 mutations expanded significantly in 2 others; in the ranking of genes by their tendency to acquire new mutations upon Cas9 introduction, TP53 was in the top 4%. The cleanest evidence came from a competition experiment: TP53-null cells mixed with their isogenic wild-type counterparts at 1:8 already drift toward dominance slowly, but did so faster when carrying Cas9; swap in ARID1A- or FBXW7-null cells and that acceleration disappears.

But the foundations of these experiments have to be stated plainly: immortalized cancer cell lines, Cas9 stably expressed over the long term, populations that often already harbor TP53-mutant subclones to begin with. That is a very different situation from a clinical-grade process in which high-fidelity Cas9 protein is electroporated into primary cells as a ribonucleoprotein, and the protein is present only transiently before being cleared.

Inside hematopoietic stem cells, the story changes shape

So what happens in the cells actually meant to treat disease? In 2019, Schiroli and colleagues at San Raffaele used optimized zinc-finger and CRISPR-Cas9 nucleases to induce one or several double-strand breaks in human hematopoietic stem and progenitor cells, then tracked DDR foci, cell-cycle progression, and transcriptional responses at up to single-cell resolution. The conclusion: even a single break made p53 pathway activation the predominant response across all subpopulations — including the most primitive one; and when the break load rose, or when AAV6 was used to deliver a homologous repair template, activation became cumulative, constraining the proliferation, yield, and engraftment of edited cells. Fortunately the impairment was reversible when the DDR burden was low, and could be offset by transient p53 inhibition — the GSE56 they used being a truncated dominant-negative p53 fragment that came from Milyavsky’s paper.

The following year, the same group quantified the cost in Nat Biotechnol using barcoded clonal tracking. Ferrari and colleagues found that once editing activates p53, the clonal repertoire of hematopoietic stem cells in hematochimeric mice shrinks substantially, while the edited clones that do engraft retain multilineage and self-renewing capacity; transient p53 inhibition restored a polyclonal graft.

These two papers changed the shape of the problem. In primary hematopoietic stem cells, the pressure p53 applies shows up mainly not as “selecting for TP53 mutants” but as attrition: a fraction of clones simply never make it into the graft. This is a functional selection rather than — at least not primarily — a genetic one; its consequences are not written in the editing efficiency but in the clonal diversity.

Has anyone gone looking for TP53 mutations directly

Yes. In 2022, Cromer and colleagues reported a targeted search in Nat Commun: primary hematopoietic stem and progenitor cells from three donors, edited with high-fidelity Cas9 protein at AAVS1, HBB, and ZFPM2, with genomic DNA taken at day 4 and day 10, then ultra-deep sequencing at clinical tumor-diagnostic depth across the exons of 523 cancer-relevant genes, plus whole-exome and whole-genome sequencing added for the AAVS1 condition — the whole-genome sequencing on a single donor only. No introduction or enrichment of tumorigenic variants was detected. They also offered a reading of the cell-line studies: those findings depend on p53 mutations already present in the initial pool of cells, which would not be expected in primary cells derived from healthy donors.

This is the most direct negative evidence available, and its boundaries are just as clear: three donors, ten days ex vivo, a finite limit of detection — covering neither what happens over years inside a human body after transplant, nor patients who already carry clonal hematopoiesis.

The pressure is real; it simply acts elsewhere

Rather than say that hematopoietic stem cells escape p53’s selection, it is more accurate to say the pressure here is applied differently, and depends heavily on the process. In 2022, Ferrari and colleagues worked out why AAV is such a particular obstacle: viral genomes and their fragments persist in the cell far longer than expected, the MRN complex is recruited onto the AAV inverted terminal repeats, and p53-dependent DDR is held high; switching to an integrase-defective lentiviral vector to deliver the template gives a lower DNA load and a shorter DDR. In 2025, Conti and colleagues reported that CRISPR-Cas9/AAV6 editing induces a durable senescence-like response driven jointly by p53 and IL-1/NF-κB, restricting graft size and clonal diversity in long-term transplantation assays; the same year, Araki and colleagues found in murine xenograft models that administering G-CSF continuously from day 1 to 14 post-transplant amplifies, through proliferative stress, the early p53 response of edited cells and thereby impedes engraftment, while delaying the start of dosing to day 5 attenuates it. The strength of the p53 response is not a fixed physical constant but a variable tuned jointly by the delivery method, the template vector, and even post-transplant medication.

Which makes “transient p53 inhibition” look like the natural fix — and it is exactly the step on this thread that demands the most care. Milyavsky’s paper had already indicated that p53 in human hematopoietic stem cells does not only handle apoptosis but independently sustains self-renewal; hold down the guardian and you hold down its other job as well. Schiroli and colleagues did not see karyotype abnormalities or an increased mutational burden in their short-term assays, but that was in vitro, over a two-to-four-week window, and they themselves wrote out the limits of that negative result: the sensitivity of their karyotype analysis was limited, and the deep sequencing covered only the targeted portion of the genome — “not detected” and “did not happen” have never been the same sentence where genotoxicity is concerned.

Can you get around it by not making a break

Partly. In 2021, Newby and colleagues reported in Nature that an adenine base editor converted the sickle cell HBB^S^ allele into the benign Makassar variant, achieving about 80% conversion in patient-derived hematopoietic stem and progenitor cells, and explicitly recorded that base editing avoided the p53 activation and larger deletions that have been observed following Cas9 nuclease treatment.

But avoiding the double-strand break is not the same as being invisible. In 2022, Li and colleagues reported in Nat Commun that in human pluripotent stem cells, the efficiencies of prime editing and cytosine base editing are likewise limited by p53 — co-delivering a dominant-negative p53 fragment raised them significantly; a single-strand nick is sensed too. The comparison Lei and colleagues published in Cell Stem Cell in 2026 shows that the risk changes shape rather than vanishing: in a mouse Unc13d model, the off-target spectrum of a hyperactive cytosine base editor was broader and more scattered than that of CRISPR-Cas9, though the total translocation frequency was an order of magnitude lower than in Cas9-edited counterparts; while Cas9 produced markedly more large deletions and inversions around the target site. Fewer breaks does not automatically mean less genotoxicity; it merely gives the risk a different face.

What the clinic has seen, and what it cannot

So far, the published follow-up of nuclease-edited products has not produced the accident this mechanism predicts. In 2024, Frangoul and colleagues reported in N Engl J Med the phase 3 results of exa-cel in severe sickle cell disease: of the 44 patients who received an infusion, at a median follow-up of 19.3 months, neutrophils and platelets engrafted in every one, and no cancers occurred.

But the distance between “not seen” and “not there” is exactly the length of the follow-up: myeloid neoplasms are events measured in years, while follow-up in studies like these is still measured in months. The evidence at human scale that genuinely refuses to be brushed off comes from another direction. In 2020, Bolton and colleagues analyzed in Nat Genet how cancer therapy reshapes the fitness landscape of clonal hematopoiesis, finding that radiation, platinum agents, and topoisomerase II inhibitors preferentially select for mutant clones in DNA damage response genes — TP53, PPM1D, CHEK2 — with sequential sampling showing directly how those clones outcompete the others. This is not editing, but it proves one thing: when the human hematopoietic system faces selective pressure along the DDR axis, it does answer by enriching for clones deficient in the p53 pathway. And in a gene therapy protocol, what applies that kind of pressure is not only the nuclease but the myeloablative conditioning itself.

There is one adjacent observation in the clinic. A 2024 therapeutics bulletin from the American College of Medical Genetics and Genomics, discussing Casgevy and Lyfgenia, records that a few individuals have developed myelodysplastic syndrome and acute myeloid leukemia after Lyfgenia — a lentiviral gene addition product rather than an editing product — although extensive investigations ruled out insertional oncogenesis driving this process; the bulletin postulates that the risk is related to accelerated clonal hematopoiesis in this population, which is precisely why it notes that prescreening individuals with SCD for preleukemic progenitors before gene therapy has been recommended. As for those who have already been treated, this is precisely the kind of question that sequential sampling of the Bolton sort can answer and editing efficiency cannot: the readout lies in how TP53 and other clonal hematopoiesis genes wax and wane in peripheral blood over time after transplant.

A question still open

Laid out in full, this thread has a rare shape: mechanistically highly persuasive, firmly demonstrated in cell lines, its upstream response observed again and again in primary hematopoietic stem cells — and yet it has never yet materialized as a single identifiable accident with an editing product. It can neither be set aside as scaremongering nor narrated as a disaster that has already happened. What it really changes is how we ought to read that handsome number: editing efficiency is a score reported by the survivors, and it tells you neither how many clones dropped out before the count was taken, nor whether those that remained did so because they repaired well or because they were simply insensitive to the damage. Answering that can only come from watching the identity and the number of the clones change over time.

References

  1. Milyavsky M, Gan OI, Trottier M, et al. A distinctive DNA damage response in human hematopoietic stem cells reveals an apoptosis-independent role for p53 in self-renewal. Cell Stem Cell. 2010;7(2):186-197. DOI
  2. Haapaniemi E, Botla S, Persson J, Schmierer B, Taipale J. CRISPR-Cas9 genome editing induces a p53-mediated DNA damage response. Nat Med. 2018;24(7):927-930. DOI
  3. Ihry RJ, Worringer KA, Salick MR, et al. p53 inhibits CRISPR-Cas9 engineering in human pluripotent stem cells. Nat Med. 2018;24(7):939-946. DOI
  4. Enache OM, Rendo V, Abdusamad M, et al. Cas9 activates the p53 pathway and selects for p53-inactivating mutations. Nat Genet. 2020;52(7):662-668. DOI
  5. Schiroli G, Conti A, Ferrari S, et al. Precise gene editing preserves hematopoietic stem cell function following transient p53-mediated DNA damage response. Cell Stem Cell. 2019;24(4):551-565.e8. DOI
  6. Ferrari S, Jacob A, Beretta S, et al. Efficient gene editing of human long-term hematopoietic stem cells validated by clonal tracking. Nat Biotechnol. 2020;38(11):1298-1308. DOI
  7. Cromer MK, Barsan VV, Jaeger E, et al. Ultra-deep sequencing validates safety of CRISPR/Cas9 genome editing in human hematopoietic stem and progenitor cells. Nat Commun. 2022;13(1):4724. DOI
  8. Ferrari S, Jacob A, Cesana D, et al. Choice of template delivery mitigates the genotoxic risk and adverse impact of editing in human hematopoietic stem cells. Cell Stem Cell. 2022;29(10):1428-1444.e9. DOI
  9. Conti A, Giannetti K, Midena F, et al. Senescence and inflammation are unintended adverse consequences of CRISPR-Cas9/AAV6-mediated gene editing in hematopoietic stem cells. Cell Rep Med. 2025;6(6):102157. DOI
  10. Araki D, Chen V, Redekar N, et al. Post-transplant G-CSF impedes engraftment of gene-edited human hematopoietic stem cells by exacerbating p53-mediated DNA damage response. Cell Stem Cell. 2025;32(1):53-70.e8. DOI
  11. Newby GA, Yen JS, Woodard KJ, et al. Base editing of haematopoietic stem cells rescues sickle cell disease in mice. Nature. 2021;595(7866):295-302. DOI
  12. Li M, Zhong A, Wu Y, et al. Transient inhibition of p53 enhances prime editing and cytosine base-editing efficiencies in human pluripotent stem cells. Nat Commun. 2022;13(1):6354. DOI
  13. Lei L, Kaufmann MM, Lao J, et al. Genotoxicity profiling reveals distinct platform- and cell type-specific effects in therapeutic gene editing for genetic hyperinflammation. Cell Stem Cell. 2026;33(6):930-944.e5. DOI
  14. Frangoul H, Locatelli F, Sharma A, et al. Exagamglogene autotemcel for severe sickle cell disease. N Engl J Med. 2024;390(18):1649-1662. DOI
  15. Bolton KL, Ptashkin RN, Gao T, et al. Cancer therapy shapes the fitness landscape of clonal hematopoiesis. Nat Genet. 2020;52(11):1219-1226. DOI
  16. Lesmana H, Kim SY, Corado AM, et al. Casgevy (exagamglogene autotemcel) and Lyfgenia (lovotibeglogene autotemcel) for individuals 12 years and older with sickle cell disease (SCD) and recurrent vaso-occlusive crises (VOC): A therapeutics bulletin of the American College of Medical Genetics and Genomics (ACMG). Genet Med Open. 2024;2:101875. DOI