Yang Liu

← Primer

HSC · Healing Source Code

LNP 的肝脏天赋——成就体内编辑的那套机制,到了骨髓却成了要克服的出身 The LNP's gift for the liver—the very machinery that made in vivo editing work is the birthright it must now overcome in the marrow

2026-07-25 · LNP, 递送, 造血干细胞, 体内编辑, ApoE

2020 年冬天开始,很多人的上臂里被推进了同一样东西:一小团纳米尺度的脂质颗粒,里面裹着一段 mRNA。人们记住的是里面那段 mRNA,几乎没人注意外面那层壳。可那层壳并不是为疫苗设计的,它诞生得更早,目标也更窄——它是被一群化学家花了十几年,一步一步调出来去找肝脏的。今天,当基因编辑领域想把编辑器直接送进人体、送进骨髓深处那一小群造血干细胞时,才发现最难对付的不是这层壳的缺陷,恰恰是它当年被精心训练出来的那项天赋。

一个为肝脏调出来的壳

把带负电的核酸送进细胞,最古老的办法是用带正电的脂质把它包起来。这条路在培养皿里很好用,进了动物体内却很快撞墙:永久带正电的脂质会黏住血液里几乎所有东西,毒性大、清除快。真正的转折是把”带电”改成”看情况带电”——ionizable lipid。它的胺基头部有一个被精心调过的 pKa:在配制颗粒的酸性缓冲液里带正电,足以把核酸抱紧;进入 pH 7.4 的血液就恢复中性,不再到处招惹;等颗粒被吞进逐渐酸化的内体,它又重新带上正电——一般认为,正是这一步去撬开内体的膜。

这套逻辑何时从直觉变成可计算的工程,有明确坐标。2010 年,Semple 等人在 Nature Biotechnology 上报告,他们以 SNALP 配方里的 DLinDMA 为起点做理性设计,筛出的 DLin-KC2-DMA 在啮齿类里 0.01 mg/kg 的 siRNA 剂量就有活性,非人灵长类里是 0.1 mg/kg。两年后,Jayaraman 等人在 Angewandte Chemie 上把这件事彻底量化:五十多个头部结构不同的脂质里,活性与 pKa 的相关性紧到刺眼,最优窗口落在 6.2–6.5 这个窄区间;而在其中一组同系物里,活性峰值出现在 pKa 6.44,对应的分子叫 DLin-MC3-DMA。

那么为什么是肝脏?这常被误会成”工程师把它设计成了肝靶向”。2010 年 Akinc 等人在 Molecular Therapy 上给出的机制更有意思:ionizable lipid 做的脂质纳米颗粒(LNP)一进血液,就被内源的 apolipoprotein E(ApoE)吸附到表面,而 ApoE 的本职工作,正是把脂蛋白交给肝细胞表面的 LDL 受体;ApoE 缺失的小鼠里,这类颗粒效力大幅下降。LNP 不是被人为瞄准了肝脏,它只是伪装成一颗脂蛋白,搭上了身体每天都在跑的那班车。这是一种极省力的靶向——也正因为省力,后来极难摆脱。

从沉默一个基因,到改写一个基因

第一款把这套化学送上药架的药,不是疫苗。2018 年 8 月,FDA 批准了 Alnylam 的 patisiran(商品名 Onpattro),用于遗传性转甲状腺素蛋白淀粉样变的多发性神经病,支持它的三期数据由 Adams 等人发表在 New England Journal of Medicine 上。翻开说明书,配方几乎就是前面那段化学史的成品清单:DLin-MC3-DMA、DSPC、胆固醇,加上一点 PEG 化脂质。同一份说明书里还写着另外两件事——每三周静脉输注 0.3 mg/kg,以及输注前必须预先给激素、对乙酰氨基酚和抗组胺药。这两行字后面会再回响一次。

让这层壳家喻户晓的,是把货物从 siRNA 换成 mRNA 的那一步。2020 年底,Polack 等人在 New England Journal of Medicine 报告了 BNT162b2 的三期结果,同类产品随后在全球范围内大规模使用;生产、放大与监管路径被一次性走通,“给人注射一管脂质包裹的 mRNA”从学术设想变成了工业常规。

货物再换一次,就到了基因编辑。把 Cas 蛋白或 base editor 写成 mRNA、和 sgRNA 装进同一颗 LNP,化学上几乎不需要新发明,却顺带来了一个安全红利:一般认为 mRNA 在体内很快被降解,编辑器是一阵短促的脉冲,而不像整合型病毒载体那样长期驻留。2021 年,Musunuru 等人在 Nature 上报告,单次输注载有 base editor 的 LNP,可在食蟹猴肝脏里近乎完全地敲低 PCSK9,血中 PCSK9 下降约 90%、LDL 胆固醇下降约 60%,单次给药后至少八个月保持稳定。同年 6 月,Gillmore 等人在 New England Journal of Medicine 报告了 NTLA-2001 的一期剂量爬坡:六名遗传性 ATTR 淀粉样变患者接受 0.1 或 0.3 mg/kg 单次静脉输注,第 28 天血清 TTR 平均下降 52% 与 87%(NCT04601051);这条管线如今已推进到三期。

到这里,体内基因编辑看上去像是一件被解决了的事。但把这几项成就并排看,它们共享同一个隐含前提:靶细胞是肝细胞。

少数几个百分点

即便在肝脏这个最顺的场景里,LNP 也远不是一台高效的机器。2013 年,Gilleron 等人在 Nature Biotechnology 上用电镜直接数了挂在 siRNA 上的胶体金颗粒:被吞进细胞的 siRNA,只有 1–2% 逃出内体进入细胞质,而且只在一个兼具早期与晚期内体特征的短暂时间窗里有机会。2015 年,Wittrup 等人在同一本刊上、在培养细胞里用活细胞成像得到了相似量级的数字:被内吞的含 LNP 囊泡里只有约 7% 发生释放,折算下来约 3.5% 的 siRNA 真正到达细胞质;释放发生在内吞后约 5–15 分钟,而细胞质里的 galectin 会立刻认出这个被戳破的内体,把它送去自噬。

也就是说,肝脏方案的成功,建立在九成以上货物白白损失之上。它还能奏效,是因为肝细胞数量巨大、被 ApoE 优先送达、剂量又可以在安全范围内往上加。一旦换成一群数量稀少、藏得很深、也并不欢迎陌生颗粒进门的细胞,这几个百分点就不再是可忽略的损耗,而是成败本身。

想去别的地方

要让 LNP 离开肝脏,得先切断它和 ApoE 的那门亲事。2020 年,Cheng 等人在 Nature Nanotechnology 上提出的 SORT 是最有影响的思路:在原有四组分配方之外掺进一个”补充脂质”,靠它的电荷性质改写整颗粒子的表面化学,把递送重心从肝脏挪到肺或脾。次年,Dilliard 等人在 PNAS 上补齐了机制——补充脂质促使表面的 PEG 化脂质脱落,让不同的血浆蛋白吸附上来,再由这些蛋白对接各自组织里高表达的受体。仍然是借内源蛋白搭车,只是换了一班车。

骨髓比肺和脾更难。它不是一个可以整体点亮的器官,而是由血窦、基质与各类前体细胞构成的复杂空间,真正要找的造血干细胞在其中只占极小一部分。2020 年,Krohn-Grimberghe 等人在 Nature Biomedical Engineering 上让脂质-高分子颗粒在小鼠骨髓的血管内皮细胞里沉默基因,借此调节干/祖细胞与白细胞的释放——漂亮的”进入骨髓”,但改写的是龛(niche)。到达骨髓和到达造血干细胞,是两个问题。

敲 HSC 的门

要点名找造血干细胞,最现成的门牌号是 CD117,也就是干细胞因子受体 c-Kit。2023 年 3 月在线发表、4 月见刊的一项工作里,Shi 等人在 Nano Letters 上报告,把抗 CD117 抗体挂在 LNP 表面,单次静脉注射就能在啮齿类体内把 siRNA 或 mRNA 送进造血干/祖细胞;在 Ai14 报告小鼠里,单次 1 mg/kg 的 Cre mRNA 让约 90% 的 HSPC 与长期造血干细胞转成 tdTomato 阳性,且这些细胞保留了干性与功能。同年 7 月,Breda 等人在 Science 上报告了同一思路更完整的一套工作:CD117/LNP 装载 Cre mRNA,5 μg 静脉注射四个月后,长期造血干细胞(LT-HSC)中带标记的比例为 55%。

这个 55% 值得停下来看两眼,因为它旁边还有一个数字:同样剂量的对照 IgG/LNP-Cre 做到了 19%。抗体带来的是约三倍的富集,而不是从零到有——未靶向的颗粒本身就有相当一部分到得了骨髓。同一篇论文的活体成像也很坦率:靶向与对照颗粒在肝脏产生的信号相当,只有股骨的信号是靶向组独有的。ApoE 那班车,一直都在开。

同一颗颗粒换个货物还能干别的:装上促凋亡的 PUMA mRNA,选择性清掉宿主的造血干细胞,做成不依赖放化疗的 conditioning——0.05 mg/kg 时六天后骨髓中 LSK 与 LT-HSC 频率分别下降 71% 和 58%,而 0.15 mg/kg 以上出现肝酶升高、肺肝淤血直至死亡(这条线留到讲免清髓 conditioning 时再细说)。

至于人的细胞,这篇论文的这一部分需要读得仔细。他们把靶向换成抗人 CD117,用 adenine base editor 把镰刀型贫血的 HBB E6V 突变改成非致病的 E6A(G-Makassar)变体,在四份患者来源标本上做到最高约 88% 的编辑效率,分化后的红系细胞里 HBB^G 最高占 β 类珠蛋白的 91.7%,低氧下几乎不再镰变。数据很漂亮——但它是在培养皿里做的。论文自己的小标题写得清清楚楚:in vitro。

报告基因与真实等位基因之间的落差

Cre-loxP 报告系统有一个容易被忽略的性质:任何一次成功的重组都会被永久记录并放大成荧光,所以它测的是”这个细胞有没有被够到过”,而不是”有多少条等位基因被改写了”。Shi 等人自己也把 Ai14 小鼠称作基因编辑的”替代模型”。当货物真的换成编辑器、再去测序数等位基因,数字会掉一个量级——但这个落差该记在谁头上,需要说清楚:五成到九成出自抗体靶向的 LNP,而下面这组个位数出自不挂抗体的配方靶向 LNP,两组之间同时换了读数方式和平台,所以落差里既有报告基因的水分,也有颗粒本身的差别。至于抗体靶向那一路,本文引到的这两篇里,LNP 的体内读数用的都是报告基因;Breda 那篇里真正数到等位基因的编辑效率,是在培养皿里得到的。

2024 年,Lian 等人在 Nature Nanotechnology 上报告了一类不挂抗体的骨髓归巢 LNP,靠配方本身(例如掺入 20 mol% 的棕榈酸酰肼)获得骨髓倾向,能转染骨髓里至少十四种细胞。在携带人类镰状 β 珠蛋白基因的 Townes 小鼠体内,每周一次、共两次静脉给药,针对 HBG1/HBG2 启动子上 BCL11A 结合位点的 Cas9 方案在 HSPC 中做到 5.2% 的 indel;换成 ABE8e_NRCH 做 Makassar 转换,是 2.43%。这些是个位数,不是五成。更耐人寻味的是,质谱显示九种骨髓归巢配方里有七种,表面吸附最多的蛋白仍然是 ApoE,ApoE 敲除小鼠里信号大幅减弱——那位老朋友并没有被甩掉,只是换了个方式起作用。

离人更近一步的证据来自 2025 年 8 月在线发表、2026 年 3 月见刊的一项工作:Xu 等人在 Nature Biomedical Engineering 上报告,用不依赖抗体的靶向 LNP 递送 ABE8e 与 sgRNA 的 mRNA,可在移植了输血依赖型 β-地中海贫血患者造血干细胞的免疫缺陷 NCG-X 小鼠体内,对 HBG1/2 启动子完成 base editing,并改善衍生红系细胞的珠蛋白链比例。这是目前最接近”人类造血干细胞在活体内被 LNP 编辑”的公开证据——但承载这些细胞的,仍然是一只小鼠。

还有一条路把肝脏从障碍变成入口:造血干细胞并非一生都住在骨髓里,胎儿期它们住在肝脏。2024 年,Palanki 等人在 PNAS 上利用这个时间差,做出靶向 CD45 的 ionizable LNP,经子宫内单次静脉注射,在胎鼠造血干细胞里实现了概念验证性的编辑。LNP 的组织趋向性从来不只是颗粒的性质,还取决于靶细胞此刻站在哪里。

还没付的账

即便有一天效率追上来了,还有两笔账挂在那里。

第一笔是重复给药。相对 AAV,LNP 一直被认为可以反复使用——它不携带衣壳蛋白,不会诱发那种一次就封死后路的中和抗体。但它带 PEG。2022 年,Ju 等人在 ACS Nano 上检测了 130 名成人的血浆:人群本底就普遍存在低水平抗 PEG 抗体,接种 mRNA-1273 后抗 PEG IgG 平均升高 13.1 倍、IgM 升高 68.5 倍,BNT162b2 组的升幅小得多(1.78 倍与 2.64 倍);抗体升高与全身反应原性相关,也与 PEG 化颗粒被血中白细胞捕获的增加相关。Onpattro 说明书要求每次输注前预先用药,针对的是输注相关反应——它的机制与抗 PEG 抗体是否相关尚无定论,但同样提示:这类颗粒反复进入人体,免疫系统不会毫无反应。如果编辑造血干细胞需要多次给药——从目前的效率看很可能需要——这笔账迟早要算。

第二笔是阈值。治好一个人的镰刀型贫血,不是编辑到一点就够,而要有足够大比例的造血干细胞被改写,循环中的红细胞才能整体越过发病门槛;这个比例该是多少,是一个独立的问题(留到讲疗效阈值时再细说)。目前公开登记的血红蛋白病基因编辑临床试验,仍然全部是 ex vivo 的:采集、体外编辑、清髓、回输。把 LNP 送进人体去编辑造血干细胞,还没有走到人身上。

同一个天赋,两种命运

回头看,几乎所有让这层壳成功的性质,都是围绕肝脏磨出来的:pKa 调在 6.2–6.5,是为了在血液里安静、在内体里发作;吸附 ApoE,是为了不花力气就被送到肝细胞门口。这些性质今天原封不动地跟着它进入骨髓——ApoE 依旧在,肝脏依旧是最先被点亮的那个器官,而内体逃逸这一关,也没有因为换了器官就变得容易。

所以这场工程的实质,不是给 LNP 增加一项新能力,而是要求它做一件与出身相反的事:去一个它天生不该去的地方,找一群数量稀少、并不活跃、也没打算接纳它的细胞。抗体、配方、时间差,都是朝同一个方向使劲。从疫苗针筒到骨髓深处,看上去只有几厘米。

参考文献

  1. Semple SC, Akinc A, Chen J, et al. Rational design of cationic lipids for siRNA delivery. Nat Biotechnol. 2010;28(2):172-176. DOI
  2. Jayaraman M, Ansell SM, Mui BL, et al. Maximizing the potency of siRNA lipid nanoparticles for hepatic gene silencing in vivo. Angew Chem Int Ed Engl. 2012;51(34):8529-8533. DOI
  3. Akinc A, Querbes W, De S, et al. Targeted delivery of RNAi therapeutics with endogenous and exogenous ligand-based mechanisms. Mol Ther. 2010;18(7):1357-1364. DOI
  4. Adams D, Gonzalez-Duarte A, O’Riordan WD, et al. Patisiran, an RNAi therapeutic, for hereditary transthyretin amyloidosis. N Engl J Med. 2018;379(1):11-21. DOI
  5. U.S. Food and Drug Administration. ONPATTRO (patisiran) lipid complex injection — Prescribing Information. NDA 210922, Alnylam Pharmaceuticals; approved 10 August 2018. Drugs@FDA
  6. Polack FP, Thomas SJ, Kitchin N, et al. Safety and efficacy of the BNT162b2 mRNA Covid-19 vaccine. N Engl J Med. 2020;383(27):2603-2615. DOI
  7. Musunuru K, Chadwick AC, Mizoguchi T, et al. In vivo CRISPR base editing of PCSK9 durably lowers cholesterol in primates. Nature. 2021;593(7859):429-434. DOI
  8. Gillmore JD, Gane E, Taubel J, et al. CRISPR-Cas9 in vivo gene editing for transthyretin amyloidosis. N Engl J Med. 2021;385(6):493-502. DOI · NCT04601051
  9. Gilleron J, Querbes W, Zeigerer A, et al. Image-based analysis of lipid nanoparticle-mediated siRNA delivery, intracellular trafficking and endosomal escape. Nat Biotechnol. 2013;31(7):638-646. DOI
  10. Wittrup A, Ai A, Liu X, et al. Visualizing lipid-formulated siRNA release from endosomes and target gene knockdown. Nat Biotechnol. 2015;33(8):870-876. DOI
  11. Cheng Q, Wei T, Farbiak L, et al. Selective organ targeting (SORT) nanoparticles for tissue-specific mRNA delivery and CRISPR-Cas gene editing. Nat Nanotechnol. 2020;15(4):313-320. DOI
  12. Dilliard SA, Cheng Q, Siegwart DJ. On the mechanism of tissue-specific mRNA delivery by selective organ targeting nanoparticles. Proc Natl Acad Sci U S A. 2021;118(52):e2109256118. DOI
  13. Krohn-Grimberghe M, Mitchell MJ, Schloss MJ, et al. Nanoparticle-encapsulated siRNAs for gene silencing in the haematopoietic stem-cell niche. Nat Biomed Eng. 2020;4(11):1076-1089. DOI
  14. Shi D, Toyonaga S, Anderson DG. In vivo RNA delivery to hematopoietic stem and progenitor cells via targeted lipid nanoparticles. Nano Lett. 2023;23(7):2938-2944. DOI
  15. Breda L, Papp TE, Triebwasser MP, et al. In vivo hematopoietic stem cell modification by mRNA delivery. Science. 2023;381(6656):436-443. DOI(Erratum: Science. 2025;388(6751):eadz0744)
  16. Lian X, Chatterjee S, Sun Y, et al. Bone-marrow-homing lipid nanoparticles for genome editing in diseased and malignant haematopoietic stem cells. Nat Nanotechnol. 2024;19(9):1409-1417. DOI
  17. Xu S, Liang D, Wang Q, et al. In vivo genome editing of human haematopoietic stem cells for treatment of blood disorders using mRNA delivery. Nat Biomed Eng. 2026;10(3):473-489. DOI
  18. Palanki R, Riley JS, Bose SK, et al. In utero delivery of targeted ionizable lipid nanoparticles facilitates in vivo gene editing of hematopoietic stem cells. Proc Natl Acad Sci U S A. 2024;121(32):e2400783121. DOI
  19. Ju Y, Lee WS, Pilkington EH, et al. Anti-PEG antibodies boosted in humans by SARS-CoV-2 lipid nanoparticle mRNA vaccine. ACS Nano. 2022;16(8):11769-11780. DOI

Starting in the winter of 2020, the same thing was pushed into a great many upper arms: a nanoscale clump of lipid particles with a stretch of mRNA wrapped inside. What people remember is the mRNA within; almost no one noticed the shell without. Yet that shell was not designed for a vaccine. It was born earlier, with a narrower aim—a group of chemists spent more than a decade tuning it, step by step, to find the liver. Today, when the gene editing field wants to send editors directly into the body, into that small population of hematopoietic stem cells deep in the bone marrow, it turns out the hardest thing to deal with is not the shell’s defects, but precisely the gift it was so carefully trained to have.

A shell tuned for the liver

The oldest way to get negatively charged nucleic acids into cells is to wrap them in positively charged lipids. That works well in the dish, but hits a wall in animals: permanently cationic lipids stick to nearly everything in the blood—toxic, rapidly cleared. The real turning point was changing “charged” to “charged when the situation calls for it”—the ionizable lipid. Its amine head group carries a carefully tuned pKa: positively charged in the acidic buffer where the particle is formulated, enough to hold the nucleic acid tight; neutral again in blood at pH 7.4, no longer provoking everything it meets; and once the particle has been swallowed into a progressively acidifying endosome, positively charged once more—it is generally thought that this is the step that pries open the endosomal membrane.

There are clear coordinates for when this logic went from intuition to computable engineering. In 2010, Semple and colleagues reported in Nature Biotechnology that, taking DLinDMA from the SNALP formulation as a starting point for rational design, the screened-out DLin-KC2-DMA was active at siRNA doses as low as 0.01 mg/kg in rodents and 0.1 mg/kg in nonhuman primates. Two years later, Jayaraman and colleagues in Angewandte Chemie quantified the whole thing: across more than fifty lipids with different head groups, the correlation between activity and pKa was startlingly tight, with the optimal window falling in the narrow band of 6.2–6.5; and within one series of close analogues, peak potency came at a pKa of 6.44, the corresponding molecule being DLin-MC3-DMA.

So why the liver? This is often misread as “the engineers designed it to target the liver.” The mechanism Akinc and colleagues gave in Molecular Therapy in 2010 is more interesting: as soon as a lipid nanoparticle (LNP) made with ionizable lipid enters the blood, endogenous apolipoprotein E (ApoE) adsorbs onto its surface—and ApoE’s day job is precisely to hand lipoproteins to the LDL receptor on the surface of hepatocytes; in mice lacking ApoE, the potency of such particles drops sharply. The LNP was not aimed at the liver by human design. It merely disguised itself as a lipoprotein and boarded a bus the body runs every day. This is targeting at minimal effort—and precisely because it costs so little, it later proved extremely hard to shake off.

From silencing a gene to rewriting one

The first drug to carry this chemistry onto a pharmacy shelf was not a vaccine. In August 2018, the FDA approved Alnylam’s patisiran (trade name Onpattro) for the polyneuropathy of hereditary transthyretin-mediated amyloidosis; the phase 3 data supporting it were published by Adams and colleagues in the New England Journal of Medicine. Open the label and the formulation reads almost like the finished parts list of that chemical history: DLin-MC3-DMA, DSPC, cholesterol, plus a little PEGylated lipid. The same label states two other things—0.3 mg/kg by intravenous infusion every three weeks, and mandatory premedication with a corticosteroid, acetaminophen and antihistamines before each infusion. Those two lines will echo again later.

What made this shell a household object was the step of swapping the cargo from siRNA to mRNA. At the end of 2020, Polack and colleagues reported the phase 3 results of BNT162b2 in the New England Journal of Medicine, and products of the same class were subsequently used at large scale worldwide; manufacturing, scale-up and the regulatory path were all walked through at once, and “injecting a tube of lipid-wrapped mRNA into a person” went from an academic proposition to industrial routine.

Swap the cargo once more and you arrive at gene editing. Writing a Cas protein or a base editor as mRNA and packing it with an sgRNA into the same LNP requires almost no new chemistry, yet brings a safety dividend along the way: mRNA is generally thought to be degraded quickly in the body, so the editor is a brief pulse rather than a long-term resident like an integrating viral vector. In 2021, Musunuru and colleagues reported in Nature that a single infusion of LNPs carrying a base editor achieved near-complete knockdown of PCSK9 in the liver of cynomolgus monkeys, with blood PCSK9 down roughly 90% and LDL cholesterol down about 60%, stable for at least eight months after that single dose. In June of the same year, Gillmore and colleagues reported in the New England Journal of Medicine the phase 1 dose escalation of NTLA-2001: six patients with hereditary ATTR amyloidosis received a single intravenous infusion at 0.1 or 0.3 mg/kg, and by day 28 mean serum TTR had fallen by 52% and 87% respectively (NCT04601051); that pipeline has since advanced to phase 3.

At this point in vivo gene editing looks like a solved problem. But set these achievements side by side and they share one implicit premise: the target cell is a hepatocyte.

A few percent

Even in the liver, the smoothest setting of all, the LNP is far from an efficient machine. In 2013, Gilleron and colleagues in Nature Biotechnology directly counted colloidal-gold particles conjugated to siRNAs by electron microscopy: of the siRNA taken into the cell, only 1–2% escaped the endosome into the cytosol, and only during a brief window when the particle sat in a compartment sharing early and late endosomal characteristics. In 2015, Wittrup and colleagues, in the same journal and in cultured cells, obtained numbers of similar magnitude by live-cell imaging: only about 7% of internalized LNP-containing vesicles released their contents, working out to roughly 3.5% of the siRNA actually reaching the cytosol; release occurred about 5–15 minutes after endocytosis, and cytosolic galectin immediately recognized the punctured endosome and consigned it to autophagy.

In other words, the success of the liver approach rests on more than ninety percent of the cargo being lost outright. It still works because hepatocytes are enormously numerous, because ApoE delivers them preferentially, and because the dose can be pushed up within a safe range. Change the target to a population that is scarce, deeply hidden, and not exactly welcoming to strange particles, and those few percent stop being negligible losses and become the whole question of success or failure.

Wanting to go elsewhere

To get an LNP away from the liver, you first have to break off its marriage to ApoE. The most influential idea came in 2020, when Cheng and colleagues proposed SORT in Nature Nanotechnology: adding a “supplemental” lipid on top of the original four-component formulation, using its charge properties to rewrite the surface chemistry of the whole particle and shift the center of delivery from the liver to the lung or the spleen. The following year, Dilliard and colleagues filled in the mechanism in PNAS—the supplemental lipid promotes desorption of the PEGylated lipid from the surface, allowing different plasma proteins to adsorb, and those proteins then engage cognate receptors highly expressed in specific tissues. Still hitching a ride on endogenous proteins; just a different bus.

The bone marrow is harder than lung or spleen. It is not an organ you can light up as a whole, but a complex space built from sinusoids, stroma and all manner of precursor cells, in which the hematopoietic stem cells you are actually after make up only a tiny fraction. In 2020, Krohn-Grimberghe and colleagues reported in Nature Biomedical Engineering that lipid-polymer nanoparticles could silence genes in bone marrow endothelial cells in mice, and thereby modulate the release of stem and progenitor cells and of leukocytes—a beautiful case of “getting into the bone marrow,” but what was rewritten was the niche. Reaching the marrow and reaching hematopoietic stem cells are two different problems.

Knocking on the HSC’s door

To call hematopoietic stem cells by name, the most readily available house number is CD117, the stem cell factor receptor c-Kit. In work published online in March 2023 and appearing in print that April, Shi and colleagues reported in Nano Letters that hanging an anti-CD117 antibody on the LNP surface allowed a single intravenous injection to deliver siRNA or mRNA into hematopoietic stem and progenitor cells in rodents; in Ai14 reporter mice, a single 1 mg/kg dose of Cre mRNA turned about 90% of HSPCs and long-term hematopoietic stem cells tdTomato-positive, and those cells retained their stemness and functionality. That same July, Breda and colleagues reported in Science a fuller body of work along the same lines: with CD117/LNP carrying Cre mRNA, four months after a 5 μg intravenous injection, the fraction of marked long-term hematopoietic stem cells (LT-HSCs) was 55%.

That 55% is worth pausing over, because there is another number beside it: control IgG/LNP-Cre at the same dose reached 19%. What the antibody buys is roughly a threefold enrichment, not a jump from zero—untargeted particles already get a substantial share into the marrow on their own. The live imaging in the same paper is equally candid: targeted and control particles produced comparable signal in the liver, and only the femoral signal was unique to the targeted group. That ApoE bus has been running all along.

The same particle with a different cargo can do something else: load it with pro-apoptotic PUMA mRNA to selectively clear the host’s own hematopoietic stem cells, making a conditioning regimen independent of radiation or chemotherapy—at 0.05 mg/kg, the frequencies of LSK and LT-HSC cells in the marrow fell by 71% and 58% respectively six days later, while 0.15 mg/kg or more produced elevated liver enzymes, venous congestion of the lungs and liver, and death (that thread is left for the piece on myeloablation-free conditioning).

As for human cells, this part of the paper needs to be read carefully. They switched the targeting to anti-human CD117 and used an adenine base editor to change the sickle cell HBB E6V mutation into the non-pathogenic E6A (G-Makassar) variant, achieving editing rates as high as about 88% in four patient-derived specimens, with HBB^G reaching up to 91.7% of beta-like globin in the differentiated erythroid cells and almost no sickling under hypoxia. The data are beautiful—but they were generated in a dish. The paper’s own subheading says it plainly: in vitro.

The gap between a reporter gene and a real allele

The Cre-loxP reporter system has a property that is easy to overlook: any successful recombination is permanently recorded and amplified into fluorescence, so what it measures is “has this cell ever been reached,” not “how many alleles were rewritten.” Shi and colleagues themselves call the Ai14 mouse a surrogate model for gene editing. When the cargo really is swapped for an editor and alleles are counted by sequencing, the numbers drop by an order of magnitude—but who that gap should be charged to needs saying clearly: the fifty-to-ninety percent figures come from antibody-targeted LNPs, while the single digits below come from antibody-free, formulation-targeted LNPs, so two variables—the readout and the platform—changed at once, and the gap contains both the inflation of a reporter gene and a difference between the particles themselves. As for the antibody-targeted line, in the two papers cited here the in vivo readouts are all reporter genes; the allele-level editing efficiencies actually counted in the Breda paper were obtained in a dish.

In 2024, Lian and colleagues reported in Nature Nanotechnology a class of antibody-free, bone-marrow-homing LNPs that acquire their marrow preference from the formulation itself (for example by incorporating 20 mol% palmitic hydrazide), able to transfect at least fourteen cell types in the bone marrow. In Townes mice carrying the human sickle β-globin gene, with two weekly intravenous doses, a Cas9 approach directed at the BCL11A binding motif in the HBG1/HBG2 promoter produced indels in 5.2% of alleles in HSPCs; switching to ABE8e_NRCH for the Makassar conversion gave 2.43%. These are single digits, not fifty percent. More intriguing still, mass spectrometry showed that in seven of the nine bone-marrow-homing formulations the most abundantly adsorbed surface protein was still ApoE, and the signal fell markedly in ApoE knockout mice—the old friend had not been shaken off, only put to work a different way.

A step closer to humans comes from work published online in August 2025 and appearing in print in March 2026: Xu and colleagues reported in Nature Biomedical Engineering that antibody-free targeted LNPs delivering ABE8e and sgRNA as mRNA can carry out base editing at the HBG1/2 promoter in immunodeficient NCG-X mice engrafted with hematopoietic stem cells from patients with transfusion-dependent β-thalassaemia, improving the globin chain balance in the derived erythroid cells. This is the closest public evidence yet for “human hematopoietic stem cells edited by an LNP in a living body”—but what carries those cells is still a mouse.

There is one more route that turns the liver from obstacle into entrance: hematopoietic stem cells do not live in the bone marrow all their lives; during fetal development they reside in the liver. In 2024, Palanki and colleagues in PNAS exploited that window of time, building CD45-targeting ionizable LNPs that achieved proof-of-concept editing in fetal mouse hematopoietic stem cells after a single in utero intravenous injection. An LNP’s tissue tropism has never been only a property of the particle; it also depends on where the target cell happens to be standing.

Bills not yet paid

Even if efficiency one day catches up, two bills are still outstanding.

The first is repeat dosing. Relative to AAV, LNPs have always been thought reusable—they carry no capsid protein and do not provoke the kind of neutralizing antibody that closes the door after a single use. But they do carry PEG. In 2022, Ju and colleagues tested plasma from 130 adults in ACS Nano: low-level anti-PEG antibodies were already common in the population before vaccination, and after mRNA-1273, anti-PEG IgG rose a mean of 13.1-fold and IgM 68.5-fold, with much smaller increases in the BNT162b2 group (1.78-fold and 2.64-fold); the antibody rise correlated with increased systemic reactogenicity, and also with increased association of PEGylated particles with leukocytes in blood. The Onpattro label requires premedication before every infusion, aimed at infusion-related reactions—whether their mechanism has anything to do with anti-PEG antibodies is unsettled, but it points the same way: when particles of this kind enter a human body repeatedly, the immune system does not stay silent. If editing hematopoietic stem cells requires multiple doses—which, judging by current efficiencies, it very likely will—this bill will eventually come due.

The second is the threshold. Curing someone’s sickle cell disease is not a matter of editing a little; a large enough fraction of hematopoietic stem cells must be rewritten for the circulating red cells as a whole to clear the threshold of disease—and what that fraction should be is a separate question (left for the piece on therapeutic thresholds). The gene editing trials for hemoglobinopathies currently on public registries are still, without exception, ex vivo: collect, edit outside the body, condition, reinfuse. Sending an LNP into a human body to edit hematopoietic stem cells has not yet reached people.

One gift, two fates

Looking back, almost every property that made this shell a success was ground down around the liver: a pKa set at 6.2–6.5, so it stays quiet in the blood and acts inside the endosome; ApoE adsorption, so it is carried to the hepatocyte’s door for free. Those properties follow it unchanged into the bone marrow today—ApoE is still there, the liver is still the first organ to light up, and the endosomal escape step has not become any easier just because the organ changed.

So the substance of this engineering effort is not to add a new capability to the LNP, but to require it to do something contrary to its origins: to go where it was never meant to go, and to find a population of cells that is scarce, not very active, and not planning to let it in. Antibodies, formulations, developmental timing—all push in the same direction. From a vaccine syringe to the depths of the bone marrow looks like only a few centimeters.

References

  1. Semple SC, Akinc A, Chen J, et al. Rational design of cationic lipids for siRNA delivery. Nat Biotechnol. 2010;28(2):172-176. DOI
  2. Jayaraman M, Ansell SM, Mui BL, et al. Maximizing the potency of siRNA lipid nanoparticles for hepatic gene silencing in vivo. Angew Chem Int Ed Engl. 2012;51(34):8529-8533. DOI
  3. Akinc A, Querbes W, De S, et al. Targeted delivery of RNAi therapeutics with endogenous and exogenous ligand-based mechanisms. Mol Ther. 2010;18(7):1357-1364. DOI
  4. Adams D, Gonzalez-Duarte A, O’Riordan WD, et al. Patisiran, an RNAi therapeutic, for hereditary transthyretin amyloidosis. N Engl J Med. 2018;379(1):11-21. DOI
  5. U.S. Food and Drug Administration. ONPATTRO (patisiran) lipid complex injection — Prescribing Information. NDA 210922, Alnylam Pharmaceuticals; approved 10 August 2018. Drugs@FDA
  6. Polack FP, Thomas SJ, Kitchin N, et al. Safety and efficacy of the BNT162b2 mRNA Covid-19 vaccine. N Engl J Med. 2020;383(27):2603-2615. DOI
  7. Musunuru K, Chadwick AC, Mizoguchi T, et al. In vivo CRISPR base editing of PCSK9 durably lowers cholesterol in primates. Nature. 2021;593(7859):429-434. DOI
  8. Gillmore JD, Gane E, Taubel J, et al. CRISPR-Cas9 in vivo gene editing for transthyretin amyloidosis. N Engl J Med. 2021;385(6):493-502. DOI · NCT04601051
  9. Gilleron J, Querbes W, Zeigerer A, et al. Image-based analysis of lipid nanoparticle-mediated siRNA delivery, intracellular trafficking and endosomal escape. Nat Biotechnol. 2013;31(7):638-646. DOI
  10. Wittrup A, Ai A, Liu X, et al. Visualizing lipid-formulated siRNA release from endosomes and target gene knockdown. Nat Biotechnol. 2015;33(8):870-876. DOI
  11. Cheng Q, Wei T, Farbiak L, et al. Selective organ targeting (SORT) nanoparticles for tissue-specific mRNA delivery and CRISPR-Cas gene editing. Nat Nanotechnol. 2020;15(4):313-320. DOI
  12. Dilliard SA, Cheng Q, Siegwart DJ. On the mechanism of tissue-specific mRNA delivery by selective organ targeting nanoparticles. Proc Natl Acad Sci U S A. 2021;118(52):e2109256118. DOI
  13. Krohn-Grimberghe M, Mitchell MJ, Schloss MJ, et al. Nanoparticle-encapsulated siRNAs for gene silencing in the haematopoietic stem-cell niche. Nat Biomed Eng. 2020;4(11):1076-1089. DOI
  14. Shi D, Toyonaga S, Anderson DG. In vivo RNA delivery to hematopoietic stem and progenitor cells via targeted lipid nanoparticles. Nano Lett. 2023;23(7):2938-2944. DOI
  15. Breda L, Papp TE, Triebwasser MP, et al. In vivo hematopoietic stem cell modification by mRNA delivery. Science. 2023;381(6656):436-443. DOI (Erratum: Science. 2025;388(6751):eadz0744)
  16. Lian X, Chatterjee S, Sun Y, et al. Bone-marrow-homing lipid nanoparticles for genome editing in diseased and malignant haematopoietic stem cells. Nat Nanotechnol. 2024;19(9):1409-1417. DOI
  17. Xu S, Liang D, Wang Q, et al. In vivo genome editing of human haematopoietic stem cells for treatment of blood disorders using mRNA delivery. Nat Biomed Eng. 2026;10(3):473-489. DOI
  18. Palanki R, Riley JS, Bose SK, et al. In utero delivery of targeted ionizable lipid nanoparticles facilitates in vivo gene editing of hematopoietic stem cells. Proc Natl Acad Sci U S A. 2024;121(32):e2400783121. DOI
  19. Ju Y, Lee WS, Pilkington EH, et al. Anti-PEG antibodies boosted in humans by SARS-CoV-2 lipid nanoparticle mRNA vaccine. ACS Nano. 2022;16(8):11769-11780. DOI