Population dynamics of normal human blood inferred from somatic mutations

a, Sorting of stem and progenitor cells. Human bone marrow (BM) and peripheral blood (PB) mononuclear cells (time point 1) were stained with anti-CD34, anti-CD38, anti-CD45RA, anti-CD90, anti-CD10 and anti-CD135 antibodies. After exclusion of debris and doublets, gatings on CD34, CD38 and CD90 were used to separate CD34⁺CD38⁻CD90⁺CD45RA⁻ HSCs. The CD34⁺CD38⁺ compartment was gated for CD10⁻ cells before gating on CD135 (also known as FLT3) and CD45RA to separate progenitor compartments: CD135⁺CD45RA⁻ common myeloid progenitor (CMPs), CD135⁺CD45RA⁺ granulocyte–macrophage progenitors (GMPs) and CD135⁻CD45RA⁻ megakaryocyte–erythrocyte progenitors (MEPs). b, Sorting of B and T lymphocytes. Peripheral blood mononuclear cells (time point, after 4 months) were stained with anti-CD4, anti-CD8 and anti-CD19 antibodies. After exclusion of debris and doublets, the CD4⁺CD8⁺CD19⁻ gate was used to isolate T cells, while the CD4⁻CD8⁻CD19⁺ gate was used to isolate B cells. n = 20,000 events.

Example histograms of the variant allele fraction (VAF—the proportion of sequencing reads that report the mutation) of mutations in single colonies. a, The VAF of all mutations on autosomes in a typical clonal colony. Because there are two copies of each autosome, and each mutation occurs on only one of them, in a clonal sample the VAF of autosomal mutations is binomially distributed with a mean of 0.5. b, The VAF of all mutations on the X chromosome in the same clonal colony. Because the subject is male, there is only one copy of the X chromosome, and so true mutations here must have a VAF of 1. Occasionally, lower VAFs are seen when a mutation is not detected on a read, when a read from another locus is aberrantly mapped to the locus in question and thus lowering the apparent coverage, or when a mutation is acquired in vitro. c, d, The VAF of autosomal and X chromosome mutations, respectively, in a typical colony seeded by more than one cell. As not all the reads come from the same cell, and most mutations are private to a given cell, a lower proportion of DNA molecules carry the mutation in a polyclonal colony than in a clonal colony, resulting in a leftward shift of the peak of the VAF histogram. These histograms suggest that the number of mutations acquired by the colonies after a few weeks of in vitro expansion is a small fraction of those acquired in vivo over 60 years of life.

a, A histogram of substitution (left) and indel (right) burden per colony. b, The location around the genome of substitutions from all clones combined is shown as a circos plot. The outermost ring of the circos plot depicts the karyotypic ideogram. Moving inwards, base substitutions are shown as rainfall plots in which the height of the dot in the substitution ring is proportional to the log₁₀ of the distance to the next mutation and with the colour of the dot illustrating the base change, as shown in the key. c, A comparison of the substitution burden between stem cells and progenitor cells. There were not significantly more mutations in progenitors than stem cells (P = 0.14, Wilcoxon rank-sum test).

a, The trinucleotide context of substitutions for all colonies combined. Substitutions can be classed according to the base change (referred to by the pyrimidine of the mutated base pair), and the bases 5′ and 3′ of the mutated one, into 96 categories. The counts in each of these categories are shown. b, Comparison with pooled acute myeloid leukaemia genomes, excluding genomes with >1,500 mutations, and publicly available data on normal tissues that have been whole-genome sequenced so far. The ordering of bars is the same as in a, and the same figure as in a is provided again at the same resolution of the other data for ease of comparison. Please note that these samples have been sequenced on different platforms using different systems, which is likely to result in small differences. Normal liver, normal colon and normal small intestine data were obtained from whole-genome sequencing of single-cell-derived organoids¹⁹, whereas normal neurons were derived from single cells that had undergone whole-genome amplification³⁷. c, Example trinucleotide substitution plots for a selection of individual colonies derived from either stem cells (which have the prefix BMH) or progenitor cells (which have the prefix BMP). The ordering of bars is the same as in a.

a, The phylogeny of cells as presented in Figs. 2, 4, 6, but with the addition of P values next to every node, derived by bootstrapping the substitution matrix 1,000 times, building a tree using SCITE for each replicate, and counting the proportion of the bootstrapped trees that support each node. b–f, Phylogenies constructed using different datasets and methods. In each case the phylogeny was constructed using 100 bootstraps of the data, and the P value for each node shown underneath it. Branches are coloured by whether a branch ancestral to exactly the same descendants is also present in the SCITE tree, and are drawn with a thicker line if the branch is recovered in ≥70% of bootstrap replicates. b, Substitution and indel datasets combined, building the tree by maximum parsimony. c, Substitution, indel and neighbour-joining datasets combined, building the tree by neighbour joining. d, Substitutions, tree build by maximum parsimony. e, Indels, tree built by maximum parsimony. f, Short tandem repeats, tree built by neighbour joining.

a, The phylogeny showing different stem and progenitor cell types. b, The phylogeny is shown as in a, but with the labels underneath coloured according to which cell types are being compared. The first row of labels has stem cells from bone marrow in red, progenitor cells from bone marrow in grey and stem cells from peripheral blood in black. The second row of labels has stem cells in red and bone marrow progenitors in black. The third row of labels has MEPs in red, CMPs in black, GMPs in blue and stem cells in grey. c–e, Analysis of molecular variance is used to test for clustering on the phylogeny for stem cells derived from peripheral blood versus bone marrow cells (c), stem cells versus progenitors (d) and different progenitor types (e). In each panel, a histogram of the null distribution of the statistic used to detect clustering is shown. Distributions were obtained by randomly permuting which cells were assigned to which category. Comparisons are only between cell types not shown in grey in b. The observed value of the statistic is shown as a red vertical line.

a, The joint prior distribution for stem cell numbers (HSCs) and the generation time for the first approximate Bayesian computation (ABC). b, The location in sample space of the 10% of simulations that produced summary statistics (using only the ltt summary statistics; Supplementary Methods and Supplementary Information) most similar to the observed summary statistics. c, The joint prior distribution for the second ABC, in the area of sample space indicated to be plausible by the first set of simulations. d, The joint posterior distribution of the best 500 simulations from the second ABC, as shown in Fig. 5 for ease of reference. ‘n’, ‘o’ and ‘p’ on the plot indicate the position in sample space from which panels n–p were drawn. e–i, Cross-validation of the model to choose the number of accepted simulations and the weighting applied to the ltt summary statistics (Supplementary Methods and Supplementary Information). j, For illustrative purposes, five simulations were sampled for each of three population sizes along the plausible diagonal of sample space indicated in b. One set of summary statistics are shown for these simulations in k. k, The red line indicates a simulation coming from the area of sample space indicated by a red point in j; and similarly for blue and green lines. The black dotted line indicates the observed values for these summary statistics. These summary statistics provide a count—for the different numbers of samples (x axis)—of how many of the 3,952 mutations that we considered (y axis) are in this many samples with two or more reads, using error model 1 (which simulates errors according to the error rate in control DNA (Supplementary Methods)). The same summary statistics were calculated for different mutant read number cut-offs. l, For each of the 1,000 simulations that produce summary statistics that were the most similar to the observed data, the Euclidean distance from the observed data (y axis) is plotted against the number of stem cells in that simulation (x axis). This information is used by the neural network regression step to define the most likely value for the number of stem cells. The most similar values are seen at around 100,000 stem cells, which was the location of the median of the posterior distribution from neural network regression. m, The observed phylogeny, with branch points indicated by asterisks. n–p, Phylogenies drawn from simulations that occur at the points in sample space indicated in d. n, A relatively plausible simulation, since the pattern of branch points is not dissimilar from the pattern of the observed phylogeny (m). Simulations with smaller stem cell populations and faster stem cell turnover rates resulted in phylogenies in which the stem cells were very closely related to each other (o), whereas those with larger populations and slower turnover result in phylogenies in which the stem cells only share an embryonic common ancestor, and no branches are seen through the tree (p).

a, Correlations between the VAFs of all sequenced samples, shown on a log scale. Note that samples that were sequenced to a lower depth cannot have VAFs as small as samples sequenced to higher depths. b, Targeted sequencing information with no error correction. The data are shown as in Fig. 4 for all analysed samples, but focusing on only the first 350 mutations of molecular time. To allow a better comparison between samples that were sequenced at different depths, a higher detection threshold and different detection threshold are used relative to Fig. 4. c, Targeted sequencing information after using cord blood controls for sequencing error correction with the Bayesian generalized Poisson mixed-effects model. The colour scale is the same as in b. The data for the granulocytes at the nine-month time point are the same as in Fig. 4 (provided again for ease of comparison), but plotted with a different colour scale.

a, The phylogeny with targeted sequencing information in different blood fractions overlaid as in Fig. 6, shown again here for ease of reference. The colouring of mutations reflects in which peripheral blood cell fractions they could be detected, as indicated by the colour key. Arrows indicate adult clones with multilineage output, with letters corresponding to panels b–f. B, B lymphocytes; G, granulocytes; G low VAF, granulocytes, allele fraction too low to be detected in lymphocytes; T, T lymphocytes. b–f, VAFs of all mutations on branches (indicated by arrows in a) with mutations beyond molecular time 100 that are detectable in granulocytes and B lymphocytes but not in T lymphocytes.