Yang Liu

← Primer

HSC · Healing Source Code

一个人有多少造血干细胞——条形码数克隆，突变钟数代际，两把尺子差了几个数量级 How many blood stem cells does a person have — barcodes count clones, mutations count generations, and the two rulers disagree by orders of magnitude

2026-09-27 · 造血干细胞, 克隆追踪, 体细胞突变, 谱系树, 有效群体大小, 基因治疗

移植实验能回答的问题其实很窄。把有限数量的细胞输给一只被清髓的动物，如果血液长回来了，你就知道输进去的那群细胞里至少有一个具备重建能力——这告诉你「能不能」，不告诉你「有多少」和「多久一次」。而后两个问题恰恰是理解造血最要紧的：一个成年人身上到底有多少个此刻真在干活的造血干细胞（hematopoietic stem cell，HSC），它们平均隔多久分裂一次。

这两个数在人身上没法用移植去问。你不能把一个健康人的骨髓清空，再输回可数的细胞去看谁长出来了。于是整个领域绕了一条道：不去扰动系统，而去读细胞自己留下的记录。记录有两种——一种是我们替它写的，一种是它自己在复制中写下的。

我们替细胞写的记录

第一条路的想法很朴素：给每个细胞发一个身份编号，编号随分裂原样传给后代，此后在血里任何一支成熟细胞上查到这个编号，就知道它的祖先是谁。最早的编号不是刻意设计的，而是基因治疗的副产品——逆转录病毒载体整合进基因组的位置本身是随机的，这个整合位点（integration site）就是一枚天然的条形码。后来的载体干脆在序列里预留一段随机寡核苷酸，读起来比整合位点更定量。

2014 年，Wu 等在《Cell Stem Cell》上用这套条形码做了非人灵长类的自体移植追踪，在恒河猴身上把单个克隆对各个成熟细胞谱系的贡献分开来读，并据此指出自然杀伤细胞的谱系来源与其他淋巴细胞不同。三年后，Koelle 等在《Blood》上把时间轴拉长：4 只接受了条形码标记细胞移植的恒河猴，随访最长达到移植后 49 个月，单核细胞、粒细胞、B 细胞与 T 细胞的长期输出来自一个多克隆的干祖细胞池，且各个克隆的贡献比例相当稳定——重建初期之后，几乎看不到「克隆接力」那种此起彼伏的更替。这是条形码方法最漂亮的一次交付：它把「造血是不是由少数克隆轮班维持」这个纯思辨的问题，变成了一条可以逐月读数的曲线。

但条形码方法有一个绕不过去的前提：为了把编号写进细胞，你得先把细胞取出来、转导、再移回去，而这个过程本身就是一次移植。小鼠模型里两项工作正面处理了这个问题。2014 年，Sun 等在《Nature》上用转座子在体内给细胞打标签，不做移植，结果发现未受扰动的稳态造血是由成千上万个克隆陆续被征召来维持的，每个克隆对成熟细胞的贡献都很小；在成年期的大部分时间里，主要的执行者是长寿的祖细胞，而不是经典定义下的 HSC。次年，Busch 等在《Nature》上用可诱导的遗传标记标定最原始的一群 HSC，再用极限稀释和建模追踪标记在下游各层的推进，估出成年小鼠体内约有五千个「在产出」的 HSC，而且标记细胞与其后代之间要达到平衡所需的时间长得出人意料——长到会超出一只小鼠的寿命。

这两项结果合在一起说明了一件重要的事：移植后看到的克隆行为和不受扰动的稳态造血不是同一件事。移植是一次强制的动员，它把系统推到一个平时不在的状态上去测量。这也是为什么在人身上，我们更想要一把不需要动手的尺子。

细胞自己写下的记录

第二条路把尺子换成了细胞自己的基因组。每一次 DNA 复制都会留下少量未被修复的碱基替换，这些突变绝大多数落在没有功能的区域，不带来任何表型，但它们会被子代忠实继承。于是任意两个造血细胞之间共享的突变数，就编码了它们上一次共有祖先出现在多久以前；把足够多的单个细胞的突变谱摆在一起，可以像重建物种演化史一样，重建一个人体内造血的系统发生树。

这台钟走得很慢，但走得相当匀。2018 年，Osorio 等在《Cell Reports》上从单个人类造血细胞建系扩增后做全基因组测序，测得造血干祖细胞平均每年积累约 14 个碱基替换，且绝大多数突变发生在出生之后、来自内源性的突变过程。同年，Lee-Six 等在《Nature》上把这套办法用到一位 59 岁健康男性身上：140 个由单细胞建系而来的造血干祖细胞克隆，共识别出 129,582 个自发的全基因组体细胞突变，重建出的谱系树给出一个估计——任一时刻真正在产出白细胞的造血干细胞数量在 50,000 到 200,000 之间；而且这个群体在生命早期稳步扩张，到青春期左右达到一个相对稳定的平台。2022 年，Mitchell 等在《Nature》上把样本从一个人扩到 10 个人、年龄从 0 岁跨到 81 岁，测了 3,579 个单细胞来源克隆的基因组，测得的突变速率是每年约 17 个，给出的群体规模是 20,000 到 200,000 个大致均匀贡献的干细胞与多能祖细胞。

几百，一万一，还是二十万

问题在于，早于这些谱系树的估计，给出的是完全不同量级的数。

2011 年，Catlin 等在《Blood》上用了一个完全不需要测序的信号：女性血细胞里父源与母源 X 染色体失活比例会随年龄缓慢漂移，漂移的幅度取决于干细胞池有多小、更新有多慢。把这个随年龄变化的比例套进随机模型，他们推出人类 HSC 平均每 40 周复制一次（范围 25 到 50 周），对应的稳态干细胞数量在模型中取 11,000 左右。更早的一支工作走的是跨物种标度：2006 年 Dingli 与 Pacheco 在《PLoS One》上用网织红细胞计数作为「在产出」干细胞池的指标，推出活跃干细胞数量随体重的四分之三次幂标度，落到 70 公斤的人身上，活跃池只有几百个细胞，对应的复制间隔约为每 60 周一次。

几百、一万一、二十万——三个数字之间差着三个数量级。把它们并列摆出来，比挑一个当共识要有信息量得多，因为差异几乎全部来自「你量的是哪个量」。

谱系树给出的不是人头数，而是群体遗传学意义上的有效群体大小（effective population size）：它描述的是能解释所观察到的突变多样性的那个群体规模。有效群体大小对群体结构、对是否存在选择、对个别克隆的隐性扩张都很敏感，通常与真实细胞数不相等。X 染色体漂移量到的是另一回事——它读的是随机性的强度：池子越小、更新越慢，漂移就越剧烈；这种方法对「有多少个细胞在贡献方差」灵敏，对总数并不灵敏。而标度法里的「活跃池」从定义上就只是一个小的工作子集，背后还挂着一个大得多的静息储备——那个储备为什么被主动维持、维持的代价是什么，是另一篇的事。

所以正确的读法不是「有一项研究错了」，而是：稳态造血由一个数量级在万到十万的干细胞群体维持，其中在任一时刻真正出手的可能只是其中一小部分，而单个细胞的更新极慢——以年，而不是以天或周为单位。分辨率还不够把这三条收敛成一个数字，这是这一章诚实的边界。

这台钟同时是一件读数工具

值得注意的是，学会数克隆之后，能被看见的东西不止「有多少」。同样的分辨率一旦到手，克隆多样性本身就成了一个可测量的指标：Mitchell 等的那套数据显示，造血的克隆多样性在大约 65 岁之前一直很高，此后急剧下降，源头是少数克隆在成年期持续的指数式扩张——在 75 岁以上的每一位受试者身上，30% 到 60% 的造血由 12 到 18 个独立克隆承担。这条线通向克隆造血与衰老，本连载另有一篇专讲，这里只记下它在方法论上的位置：它是数克隆这件事的自然延伸。

更直接的延伸是基因治疗。当我们把改造过的细胞输回一个人体内，我们真正想知道的是有多少个克隆真的落了脚、它们的贡献比例随时间怎么变、多样性会不会收窄到只剩几支。这正是整合位点作为条形码的用处所在。2018 年，Scala 等在《Nature Medicine》上追踪了 6 位接受慢病毒基因治疗的 Wiskott-Aldrich 综合征患者，随访最长 5 年，共读出 148,093 个独立克隆，并据此把移植后早期重建与随后的稳态阶段区分开来：原始的多能祖细胞与 HSC 在这两个阶段承担的角色并不相同，一部分体外激活过的 HSC 在移植后会经历一段明确的延迟才开始出手，而偏淋巴系的早期祖细胞似乎能长期存活、不必依赖 HSC 的持续供应。

这些结论都不是从移植是否成活里读出来的，而是从克隆编号的名单里读出来的。换句话说：基因治疗之所以能看见克隆多样性的收窄、能分辨是哪一层细胞在支撑第三年的血象，是因为领域先花了十几年学会怎么数克隆、怎么读那台慢钟。测量工具本身，最后成了疗法的仪表盘。

参考文献

  1. Wu C, Li B, Lu R, Koelle SJ, Yang Y, Jares A, et al. Clonal tracking of rhesus macaque hematopoiesis highlights a distinct lineage origin for natural killer cells. Cell Stem Cell. 2014;14(4):486–499. DOI
  2. Koelle SJ, Espinoza DA, Wu C, Xu J, Lu R, Li B, Donahue RE, Dunbar CE. Quantitative stability of hematopoietic stem and progenitor cell clonal output in rhesus macaques receiving transplants. Blood. 2017;129(11):1448–1457. DOI
  3. Sun J, Ramos A, Chapman B, Johnnidis JB, Le L, Ho YJ, Klein A, Hofmann O, Camargo FD. Clonal dynamics of native haematopoiesis. Nature. 2014;514(7522):322–327. DOI
  4. Busch K, Klapproth K, Barile M, Flossdorf M, Holland-Letz T, Schlenner SM, Reth M, Höfer T, Rodewald HR. Fundamental properties of unperturbed haematopoiesis from stem cells in vivo. Nature. 2015;518(7540):542–546. DOI
  5. Osorio FG, Rosendahl Huber A, Oka R, Verheul M, Patel SH, Hasaart KAL, et al. Somatic mutations reveal lineage relationships and age-related mutagenesis in human hematopoiesis. Cell Reports. 2018;25(9):2308–2316.e4. DOI
  6. Lee-Six H, Øbro NF, Shepherd MS, Grossmann S, Dawson K, Belmonte M, et al. Population dynamics of normal human blood inferred from somatic mutations. Nature. 2018;561(7724):473–478. DOI
  7. Mitchell E, Spencer Chapman M, Williams N, Dawson KJ, Mende N, Calderbank EF, et al. Clonal dynamics of haematopoiesis across the human lifespan. Nature. 2022;606(7913):343–350. DOI
  8. Catlin SN, Busque L, Gale RE, Guttorp P, Abkowitz JL. The replication rate of human hematopoietic stem cells in vivo. Blood. 2011;117(17):4460–4466. DOI
  9. Dingli D, Pacheco JM. Allometric scaling of the active hematopoietic stem cell pool across mammals. PLoS ONE. 2006;1(1):e2. DOI
  10. Abkowitz JL, Catlin SN, McCallie MT, Guttorp P. Evidence that the number of hematopoietic stem cells per animal is conserved in mammals. Blood. 2002;100(7):2665–2667. DOI
  11. Scala S, Basso-Ricci L, Dionisio F, Pellin D, Giannelli S, Salerio FA, et al. Dynamics of genetically engineered hematopoietic stem and progenitor cells after autologous transplantation in humans. Nature Medicine. 2018;24(11):1683–1690. DOI

Transplantation answers a narrower question than it looks. Give a limited number of cells to a conditioned animal, watch blood come back, and you have learned that at least one cell in that graft could rebuild the system. That is a statement about capacity. It says nothing about how many such cells a body holds, or how often any one of them divides — and those are the two numbers that decide how you think about hematopoiesis.

Neither can be asked of a human by transplantation. You cannot empty a healthy person’s marrow and put back a countable number of cells to see whose descendants appear. So the field took a detour: instead of perturbing the system, read the records the cells already carry. There are two kinds of record — the ones we write for them, and the ones they write themselves every time they copy their genome.

The records we write

The first idea is blunt. Give every cell a serial number, let the number pass unchanged to its descendants, and from then on any mature blood cell you sequence tells you which ancestor it came from. The earliest serial numbers were a byproduct rather than a design: retroviral vectors integrate at effectively random genomic positions, and that integration site is a natural barcode. Later vectors carried a stretch of random oligonucleotide on purpose, which reads out far more quantitatively than an integration site does.

In 2014, Wu and colleagues, writing in Cell Stem Cell, used such barcodes in an autologous transplant setting in rhesus macaques, resolving how individual clones contributed to each mature lineage — and argued on that basis that natural killer cells have a lineage origin distinct from other lymphocytes. Three years later, Koelle and colleagues, in Blood, stretched the time axis: four macaques transplanted with barcoded cells, followed for up to 49 months, showed long-term output of monocytes, granulocytes, B cells and T cells arising from a polyclonal pool, with individual clonal contributions remaining strikingly stable. After the initial reconstitution phase there was little sign of clonal succession — no relay of clones taking turns. This is barcoding’s finest delivery: it turned a purely speculative question about whether blood is maintained by a rotating handful of clones into a curve you can read month by month.

But the method carries an inescapable precondition. To write the serial number into a cell, you must take the cell out, transduce it, and put it back — and that is itself a transplant. Two mouse studies confronted the problem head on. In 2014, Sun and colleagues, in Nature, used transposon tagging to label cells in situ, with no transplant, and found that unperturbed steady-state blood production is maintained by the successive recruitment of thousands of clones, each contributing minimally; for most of adulthood the main workhorses were long-lived progenitors rather than classically defined HSCs. The following year, Busch and colleagues, also in Nature, applied inducible genetic labelling to the most primitive HSC compartment and tracked the label’s progression through downstream stages by limiting dilution and modelling. They estimated that at least 30% — roughly 5,000 — HSCs are productive in an adult mouse, and found that the time needed for labelled HSCs and their progeny to reach equilibrium was surprisingly long: long enough to exceed a mouse’s lifespan.

Together these results make a point worth holding onto. Clonal behaviour after transplantation and clonal behaviour in the undisturbed steady state are not the same phenomenon. Transplantation is a forced mobilization; it pushes the system into a state it does not normally occupy, and then measures it there. Which is precisely why, in humans, one would rather have a ruler that requires no intervention at all.

The records cells write themselves

The second route swaps the ruler for the cell’s own genome. Every round of DNA replication leaves behind a small number of unrepaired base substitutions. The overwhelming majority land in functionally inert sequence and produce no phenotype at all — but they are inherited faithfully. The number of mutations two blood cells share therefore encodes how long ago their last common ancestor existed. Assemble enough single-cell mutation profiles and you can reconstruct a phylogeny of one person’s hematopoiesis, exactly as one reconstructs the evolutionary history of species.

The clock is slow, but it ticks reasonably evenly. In 2018, Osorio and colleagues, in Cell Reports, clonally expanded single human blood progenitors and sequenced their genomes, measuring an accumulation of roughly 14 base substitutions per year, with most mutations arising after birth from endogenous mutational processes. The same year, Lee-Six and colleagues, in Nature, applied the approach to a single healthy 59-year-old man: 140 single-cell-derived haematopoietic stem and progenitor colonies, 129,582 spontaneous genome-wide somatic mutations, and a phylogeny yielding an estimate of 50,000 to 200,000 stem cells actively making white blood cells at any one time — a population that grew steadily in early life and reached a stable plateau by adolescence. In 2022, Mitchell and colleagues, in Nature, widened the frame from one person to ten, spanning ages 0 to 81, sequencing 3,579 single-cell-derived colonies. They measured a mean of 17 mutations per year and inferred a stable population of 20,000 to 200,000 HSCs and multipotent progenitors contributing roughly evenly to blood production.

A few hundred, eleven thousand, or two hundred thousand

The trouble is that estimates predating the phylogenies land in a different order of magnitude entirely.

In 2011, Catlin and colleagues, in Blood, exploited a signal requiring no sequencing at all. In women, the ratio of maternal to paternal X-chromosome phenotypes in blood drifts slowly with age, and the magnitude of that drift depends on how small the stem cell pool is and how slowly it turns over. Fitting the age-dependent ratio to a stochastic model, they inferred that human HSCs replicate on average once every 40 weeks, with a range of 25 to 50 weeks — with the steady-state stem cell number in their model taken at around 11,000. An earlier line of work went by cross-species scaling instead: in 2006, Dingli and Pacheco, in PLOS ONE, used reticulocyte counts as a marker of the active stem cell pool and derived a three-quarter-power scaling of active stem cell number with body mass. Extrapolated to a 70 kg human, the active pool comes out at only a few hundred cells, with a corresponding replication interval of roughly once every 60 weeks.

A few hundred, eleven thousand, two hundred thousand — three orders of magnitude apart. Laying them side by side is far more informative than picking one and calling it consensus, because almost all of the discrepancy comes from which quantity each method actually measures.

A phylogeny does not return a headcount. It returns an effective population size in the population-genetic sense: the population size that would explain the observed mutational diversity. Effective population size is sensitive to population structure, to the presence of selection, and to cryptic expansions of individual clones, and it is generally not equal to the true number of cells. X-chromosome drift measures something else again — the strength of stochasticity. The smaller the pool and the slower the turnover, the more violently the ratio drifts; the method is exquisitely sensitive to how many cells contribute variance and rather insensitive to the total. And in the scaling approach, the “active pool” is by construction only a small working subset, backed by a far larger quiescent reserve — why that reserve is actively maintained, and what maintaining it costs, belongs to another instalment.

So the right reading is not that one study got it wrong. It is this: steady-state hematopoiesis is sustained by a stem cell population somewhere in the tens of thousands to low hundreds of thousands, of which only a fraction may be actively contributing at any given moment, and the individual cells turn over extremely slowly — on a timescale of years, not days or weeks. The resolution is not yet good enough to collapse those three estimates into a single number, and that is the honest boundary of this chapter.

The clock doubles as an instrument

What is worth noticing is that once you can count clones, more than “how many” becomes visible. With that resolution in hand, clonal diversity itself becomes a measurable quantity. Mitchell and colleagues’ dataset showed high clonal diversity up to roughly age 65, after which it fell precipitously, driven by a handful of clones expanding exponentially through adulthood: in each subject over 75, some 30% to 60% of hematopoiesis was accounted for by 12 to 18 independent clones. That thread runs on into clonal hematopoiesis and aging, which has its own instalment in this series; the point here is only methodological. Measuring diversity is the natural extension of learning to count clones.

The more immediate extension is gene therapy. When engineered cells go back into a person, what one actually wants to know is how many clones took hold, how their relative contributions shift over time, and whether diversity will narrow to a few dominant branches. This is exactly where integration sites as barcodes earn their keep. In 2018, Scala and colleagues, in Nature Medicine, tracked six patients with Wiskott-Aldrich syndrome treated by lentiviral gene therapy for up to five years, reading out 148,093 individual clones, and used them to separate the early post-transplant reconstitution phase from the subsequent steady state. Primitive multipotent progenitors and HSCs played distinct roles across those two phases; a fraction of the HSCs activated in vitro proved resilient and underwent a defined delayed activation period after transplant; and early lymphoid-biased progenitors appeared capable of long-term survival, such that they might be maintained independently of continuous production from HSCs.

None of those conclusions can be read off whether the graft took. They are read off a list of clonal serial numbers. Which is to say: gene therapy can see clonal diversity narrowing, and can tell which compartment is holding up the blood counts in year three, only because the field spent well over a decade learning how to count clones and how to read that very slow clock. The measuring instrument became the dashboard for the therapy.

References

  1. Wu C, Li B, Lu R, Koelle SJ, Yang Y, Jares A, et al. Clonal tracking of rhesus macaque hematopoiesis highlights a distinct lineage origin for natural killer cells. Cell Stem Cell. 2014;14(4):486–499. DOI
  2. Koelle SJ, Espinoza DA, Wu C, Xu J, Lu R, Li B, Donahue RE, Dunbar CE. Quantitative stability of hematopoietic stem and progenitor cell clonal output in rhesus macaques receiving transplants. Blood. 2017;129(11):1448–1457. DOI
  3. Sun J, Ramos A, Chapman B, Johnnidis JB, Le L, Ho YJ, Klein A, Hofmann O, Camargo FD. Clonal dynamics of native haematopoiesis. Nature. 2014;514(7522):322–327. DOI
  4. Busch K, Klapproth K, Barile M, Flossdorf M, Holland-Letz T, Schlenner SM, Reth M, Höfer T, Rodewald HR. Fundamental properties of unperturbed haematopoiesis from stem cells in vivo. Nature. 2015;518(7540):542–546. DOI
  5. Osorio FG, Rosendahl Huber A, Oka R, Verheul M, Patel SH, Hasaart KAL, et al. Somatic mutations reveal lineage relationships and age-related mutagenesis in human hematopoiesis. Cell Reports. 2018;25(9):2308–2316.e4. DOI
  6. Lee-Six H, Øbro NF, Shepherd MS, Grossmann S, Dawson K, Belmonte M, et al. Population dynamics of normal human blood inferred from somatic mutations. Nature. 2018;561(7724):473–478. DOI
  7. Mitchell E, Spencer Chapman M, Williams N, Dawson KJ, Mende N, Calderbank EF, et al. Clonal dynamics of haematopoiesis across the human lifespan. Nature. 2022;606(7913):343–350. DOI
  8. Catlin SN, Busque L, Gale RE, Guttorp P, Abkowitz JL. The replication rate of human hematopoietic stem cells in vivo. Blood. 2011;117(17):4460–4466. DOI
  9. Dingli D, Pacheco JM. Allometric scaling of the active hematopoietic stem cell pool across mammals. PLOS ONE. 2006;1(1):e2. DOI
  10. Abkowitz JL, Catlin SN, McCallie MT, Guttorp P. Evidence that the number of hematopoietic stem cells per animal is conserved in mammals. Blood. 2002;100(7):2665–2667. DOI
  11. Scala S, Basso-Ricci L, Dionisio F, Pellin D, Giannelli S, Salerio FA, et al. Dynamics of genetically engineered hematopoietic stem and progenitor cells after autologous transplantation in humans. Nature Medicine. 2018;24(11):1683–1690. DOI