Cell-type-resolved mosaicism reveals clonal dynamics of the human forebrain

Abstract

Debate remains around anatomic origins of specific brain cell subtypes and lineage relationships within the human forebrain ^1–7. Thus, direct observation in the mature human brain is critical for a complete understanding of its structural organization and cellular origins. Here, we utilize brain mosaic variation within specific cell types as distinct indicators for clonal dynamics, denoted as cell-type-specific Mosaic Variant Barcode Analysis. From four hemispheres and two different human neurotypical donors, we identified 287 and 780 mosaic variants (MVs), respectively, that were used to deconvolve clonal dynamics. Clonal spread and allelic fractions within the brain reveal that local hippocampal excitatory neurons are more lineage-restricted than resident neocortical excitatory neurons or resident basal ganglia GABAergic inhibitory neurons. Furthermore, simultaneous genome-transcriptome analysis at both a cell-type-specific and single-cell level suggests a dorsal neocortical origin for a subgroup of DLX1⁺ inhibitory neurons that disperse radially from an origin shared with excitatory neurons. Finally, the distribution of MVs across 17 locations within one parietal lobe reveals that restriction of clonal spread in the anterior-posterior axis precedes restriction in the dorsal-ventral axis for both excitatory and inhibitory neurons. Thus, cell-type resolved somatic mosaicism can uncover lineage relationships governing the development of the human forebrain.

Forebrain development is under the control of morphogens and transcription factors that mediate patterning of three-dimensional structures including the hippocampus, cortex, and basal ganglia ^8–12. Although clonal dynamics in the mouse forebrain have been investigated using single-cell viral barcoding ^13,14, studies in humans are limited to clonal relationships at broad spatial levels ^15–18. Given that the human brain comprises diverse cell types originating from various sources that intermingle, and eventually reside in proximal locations, direct mapping of clonal dynamics within distinct human forebrain cell types is essential.

While most forebrain cells are thought to originate from radial glia that line the telencephalic lateral ventricles ^19–21, observations in rodents instead suggest ventral telencephalic progenitors as a source of GABAergic cortical inhibitory neurons ¹, supported by subsequent investigations in non-human primate, and human fetal tissue ^{3,5,13,22–24}. However, conflicting findings suggest a recently evolved, potentially primate-specific dorsal telencephalic source of inhibitory neurons based on marker staining or single-cell lineage tracing in cultured human fetal brain tissues ^{2,4,6,25–27}. Yet none of these studies directly observed lineage relationships of inhibitory neurons within the fully developed human brain, leaving this longstanding debate unresolved.

Postzygotic mutations transmit faithfully to daughter cells that distribute in mosaic patterns, referred to as mosaic variants (MVs). The human brain, like other tissues, acquires MVs in part due to rapid expansions of initial founder cell pools ^16,17, resulting in clonal lineages sharing MVs that can vary in bulk allele fraction (AF, portion of alternative allele among total alleles) depending upon cell mixing. As neurogenesis predominantly occurs during brain development, the distribution of MVs in adult using Mosaic Variant Barcode Analysis (MVBA) can reveal clonal dynamics and lineage relationships that likely originated during embryogenesis ^16–18.

Such approaches face challenges when dealing with small cell populations, such as cortical inhibitory neurons, due to technical limitations in obtaining adequate sample quantity from postmortem human tissues, and in deriving high-quality MV call sets. To overcome these challenges, we developed a methanol-fixed nuclei sorting (MFNS) protocol, which was added to prior protocols for cell-type-specific mosaic variant barcode analysis in bulk, sorted nuclei, and single nuclei, allowing DNA isolation from high-quality intact nuclei suitable for library preparation. Additionally, we utilized a recently developed single-cell multi-omics approach to generate DNA genotypes and RNA transcriptomes from the same cell ²⁸, deconvolving lineages and mapping cell-type-specific clonal dynamics within the human brain.

Identification of brain MVs

Deep sequencing (300X) of a single biopsy detects dozens of clonal MVs including single nucleotide variants (SNVs) and small insertions or deletions (indels) ^29,30. To examine genomic relationships across distinct cell types in human brains, we further improved a prior protocol, requiring at least 50,000 nuclei ¹⁷, for lower cell number input (>200 nuclei) and greater cell type diversity through MFNS, termed cell-type-specific mosaic variant barcode analysis (cMVBA), comprising three phases (Fig. 1a and Methods). In the ‘Tissue collection’ phase, we revisited a subset of samples collected from a previously published donor ID01 ¹⁷. We also newly biopsied (8 mm punch) each lobe of the neocortex (CTX), basal ganglia (BG), hippocampus (HIP), thalamus (TH), and cerebellum (CB) from ID01 or from an ascertained new donor (ID05) (Extended Data Fig. 1, Supplementary Data 1). We additionally collected non-brain tissue including heart, liver, both kidneys, adrenal, and skin to define MV distribution across the body. In the ‘MV discovery’ phase, a subset of bulk tissues (10 and 32 tissues from ID01 and ID05, respectively, Supplementary Data 1) underwent 300X whole genome sequencing (WGS), followed by state-of-the-art MV calling and filtering based on established methods (Methods). In the ‘Validation and quantification’ phase, we first prepared DNA samples extracted from bulk tissue and sorted nuclear populations, or amplified DNA from single nuclei (Supplementary Data 1). NEUN⁺, DLX1⁺, TBR1⁺, NEUN⁻/LHX2⁺, OLIG2⁺, NEUN⁺/DARPP32⁺, PU.1⁺ nuclear pools represented pan-neurons, GABAergic inhibitory neurons, excitatory neurons, astrocytes, oligodendrocytes, medium spiny neurons, and microglia, respectively ^17,31. To isolate intact DNA from underrepresented cell types such as cortical inhibitory neurons, we implemented MFNS by screening several antibodies targeting DLX1 (pan-inhibitory neuronal marker), TBR1 (excitatory neuronal marker), and COUPTFII (CGE-derived inhibitory neuronal marker; encoded by NR2F2) (Methods), and confirmed that each antibody labels a particular cell type within the biopsy (Extended Data Fig. 2a–c).

Figure 1. Comprehensive cMVBA identifies cell-type-resolved and region-specific MVs.

(a) cMVBA workflow overview consists of three phases: 1. Tissue collection: Cadaveric organs accessed for tissue punches from organs listed (red circles and black arrows: punch locations in organs and frontal lobe) for MV detection. 2. A subset of the bulk tissue punches undergo 300x WGS followed by best-practice MV calling pipelines to generate a list of MV candidates. 3. DNA from each punch bulk tissue, methanol fixed nuclear sorted (MFNS) samples, or individual nuclei are subjected to validation and quantification of MV candidates via massive parallel amplicon sequencing (MPAS). AFs of the validated MVs from MPAS are used to determine clonal dynamics of different neuronal cell types and reconstruct features of brain development. (b) UMAP plot from snRNA-seq with MFNS sorted or unsorted cortical nuclear pools (n = 3322). Sorted nuclear groups labeled with distinct colors. (c) Cortical cell types based on marker expression and differentially expressed genes in each cluster. (d) Cell type proportion within each sorted cortical nuclear population. (e-f) Number of MVs categorized by location detected in each donor ID01 or ID05 (Supplementary Data 4). ‘Brain-only’ MVs (i.e., detected only in brain tissue, but not other organs) including subtypes in red. ‘C+D only’: Brain-only MVs exclusively detected in both COUPTFII⁺ and DLX1⁺ populations but not the other cell types. Ast, Astrocyte; ExN, Excitatory neurons; InN, inhibitory neurons; OL, oligodendrocytes; OPC, oligodendrocyte precursor cells; U, undefined.

We further confirmed the cell-type composition of DLX1⁺, TBR1⁺, and COUPTFII⁺ nuclear pools compared with unsorted and DAPI⁺ nuclear pools with single-nucleus RNA-seq (Fig. 1b). Cell types of each cluster were identified based on marker expression patterns (Fig. 1c, Extended Data Fig. 2d and e). DLX1⁺ and COUPTFII⁺ pools contained mostly inhibitory neurons (>75%), and almost 100% of TBR1⁺ nuclear pools were confirmed as excitatory neurons (Fig. 1d). Of note, COUPTFII⁺ nuclei were mostly in a subcluster of inhibitory neuronal clusters, whereas DLX1⁺ nuclei covered all inhibitory neuronal clusters, suggesting that nuclei sorted by COUPTFII antibody reflect a subset (~38%) of DLX1⁺ nuclei highly expressing NR2F2 (Extended Data Fig. 2f). Cell type identities of the DLX1⁺ and TBR1⁺ nuclei were further confirmed as inhibitory and excitatory neurons, respectively, by comparing DNA methylation patterns from reference methylomes of inhibitory and excitatory neurons across marker gene regions ³² (Extended Data Fig. 2g–j).

Using a total of 321 samples from ID01 and 147 samples (bulk or sorted nuclei) from ID05, we conducted ultra-deep massive parallel amplicon sequencing (MPAS) of each candidate MV (ave. coverage ~10,000X). This step served two functions: 1] providing orthogonal validation for the MV within the sample. 2] providing accurate assessment of AF of each detected variant, allowing for downstream analyses (Extended Data Fig. 3a,b, Supplementary Data 2 and 3). A total of 287 and 780 MV candidates detected in WGS were thus positively validated and quantified in ID01 (Extended Data Fig. 3c) and ID05 (Extended Data Fig. 3d), respectively (Methods), subsequently used to annotate MVs according to brain region and cell type (Fig. 1e and f, Extended Data Fig. 4a and b, Supplementary Data 4). Notably, we captured 2.7 times more MVs than ID01 due to the improvement of the employed tissue biopsy method. We next performed mutational signature analysis using 368 brain-specific somatic single nucleotide variants (sSNVs) detected in ID01 or ID05 (Extended Data Fig. 4c). As expected, clock-like mutations such as signatures 1 and 5 were major components of the mutational spectrum, reflecting developmental origins. The AF distribution of organ-shared MVs showed absence of a peak at 25% AF, implying asymmetric clonal branching during early embryonic development consistent with previous observations ^33,34 (Extended Data Fig. 4d and e). The proportion of ‘Brain-only’ MVs showed a similar distribution in ID01 and ID05, accounting for 29.6% (85/287) and 37.8% (295/780) of total MVs, respectively. A total of 7 and 29 MVs were exclusively found in DLX1⁺ or COUPTFII⁺ but not in TBR1⁺ neurons (C+D only) in ID01 and ID05, respectively. We observed similar trends of MV hemispheric restriction and microglia distribution as we and others recently reported ^17,35,36 (Extended Data Fig. 5). This suggests that the cMVBA pipeline reports anatomic- and cell-type-specific MVs that can be used to profile clonal dynamics and reconstruct lineage relationships of specific cell types.

Genetic similarity of forebrain parts

The telencephalon derives from the most rostral part of the neural tube, subsequently committing to CTX, BG, and HIP, composing major structures of the adult forebrain. Since we observed significantly fewer Brain-only MVs shared between HIP and other telencephalic structures such as CTX and BG (Fig. 1e, f), we hypothesized the HIP founder cells are restricted in lineage compared to other brain regions (Fig. 2a). Just as clustering based on haplotypes allele frequency patterns can assess genetic structures among populations ^37,38, clustering of biopsied samples in an individual of MV AFs can be used to infer clonal relationships. We thus performed hierarchical clustering using AFs in bulk samples from ID01. We found HIP samples strongly clustered away from CTX and BG samples, suggesting that HIP progenitors are clonally more distinct from CTX or BG cells (Fig. 2b).

Figure 2. Human hippocampal lineage diverges from the cortex and basal ganglia.

(a) Model of clonal dynamics in forebrain anlage. Cells restricted to anlage A, which acquire a new MV (purple) are rarely present in anlage B. Later, B diverges into B’ and B”, which share more clones (yellow) than with anlage A. This analysis was applied to the geographies of the basal ganglia (BG), cortex (CTX) and hippocampus (HIP). (b-c) Heatmaps with 30 bulk samples from left hemispheric CTX, BG, and HIP (y-axis, b) or 12 selected sorted cell types (y-axis, c), compared with MVs identified in at least two samples (146 MVs, x-axis in (b) or 131 MVs, x-axis in (c)), depicted in Fig. 2d. Dendrograms at right shows greater HIP lineage separation (purple, arrow) compared with CTX or BG (green and yellow) either using bulk tissue (b) sorted nuclei (c), suggesting HIP earlier lineage restriction. (d) Counts of shared MVs across CTX, BG or HIP within sorted nuclear pools showing many more shared MVs between CTX and BG compared with HIP in both donors (permutation P<0.001). (e-f) Contour plots of informative 113 and 131 MVs from (b) and (c) (blue) and two kernel density estimation plots (grey). Axes show absolute normalized AF difference for each MV averaged across all samples from the respective tissues (CTX, HIP, BG). Black line: identity line, black dots: individual MVs, large red dot: averaged across all MVs, suggesting AF differences are smaller between CTX and BG than HIP. Abbreviations: Cau, caudate; DG, dentate gyrus; HIP and Hip, hippocampal tissue, where Hip refers to hippocampal subregion not specified; I, insular cortex; O, occipital cortex; P, parietal cortex; PF, prefrontal cortex; Put, putamen; T, temporal cortex; mO, medial occipital cortex; GP, globus pallidus; sqrt-t (AF), square-root transformed allele fraction.

Forebrain structures contain heterogeneous cell types, not only derived from local progenitors but also cells migrating from distant brain regions ^1,39,40. To exclude the possibility that migrating cells contributed to these findings, we repeated hierarchical clustering, this time restricting analysis to only locally originating cell types, i.e. excitatory (TBR1⁺ nuclei) in HIP or CTX and inhibitory neurons (DLX1⁺ nuclei) in BG. Hierarchical clustering with Manhattan distances of AFs in sorted nuclei samples from ID05 (Fig. 2c) and ID01 (Extended Data Fig. 6a) replicated the genomic similarity of HIP-TBR1⁺ clones from CTX-TBR1⁺ and BG-DLX1⁺ clones, confirming that cortical excitatory neurons show AF patterns more like inhibitory neurons in BG than excitatory neurons in HIP. Furthermore, CTX-TBR1⁺ nuclei shared many more MVs with BG-DLX1⁺ (40 and 43 for ID01 and ID05, respectively) than with HIP-TBR1⁺ nuclei (5 and 1 for ID01 and ID05, respectively) (Fig. 2d). The AF variation was greater between CTX and HIP than between CTX and BG, exhibiting average vector of data points above the identity line (Fig. 2e, f and Extended Data Fig. 6b). Taken together, this analysis suggests that the clonality of HIP progenitors is unlikely a result of migration of cells from other forebrain structures, and instead the results of locally proliferative cells within the HIP anlage.

Clonal dynamics of inhibitory neurons

While in vitro analysis of neuronal progenitors from dorsal human brain cortical tissue shows the potential to develop into inhibitory neurons ⁶, direct evidence for a dorsal origin of cortical inhibitory neurons in the mature human brain is lacking. cMVBA allowed for comparison of genomic similarity between different classes of cortical inhibitory neurons and other cell types. Hierarchical clustering was carried out for AFs measured in DLX1⁺, TBR1^+, and COUPTFII⁺ nuclear pools isolated from widespread sampling from cortical areas in both ID01 and ID05 (Fig. 3a). Intriguingly, in four different cortical hemispheres, most of COUPTFII⁺ nuclei (i.e. caudal ganglionic eminence (CGE)-derived inhibitory neurons that distribute across cortical areas ⁶) were exclusively clustered together, whereas most DLX1⁺ and TBR1⁺ nuclear populations in the same punch were clustered together (Fig. 3b, c). Bootstrap analysis further statistically validated that many of these dendrogram clusters appeared very unlikely to have arisen by chance (Extended Data Fig. 7a and b). Even after simulating removal ~25% of the excitatory neuronal component from DLX1⁺ nuclear pools (Fig. 1d) using computational deconvolution (Methods), most of DLX1⁺ and TBR1⁺ nuclear pools at the same cortical lobe remained clustered together apart from COUPTFII⁺ nuclear pools (Extended Data Fig. 7c and d). Conversely, we simulated TBR1⁺ nuclear contamination into COUPTFII⁺ nuclear pools at a similar level (25%) and compared with DLX1⁺ nuclear pools (Extended Data Fig. 7e and f). The “contaminated” COUPTFII⁺ nuclear pools continued to group together with the original COUPTFII samples, rather than clustering with TBR1⁺ nuclear pools in the same lobe, unlike the deconvolved DLX1⁺ nuclei did (Extended Data Fig. 7c and d), indicating that deconvolved DLX1⁺ nuclei may have a fraction of dorsal clones sharing MVs with local TBR1⁺ neurons.

Figure 3. Clonal dynamics of cortical excitatory and inhibitory neurons.

(a) cMVBA workflow uses MFNS nuclei for MPAS assessment of AFs in cortical punches. (b-c) Heatmaps of sorted nuclei based on AFs of 146 informative shared MVs from ID01 (b) and 186 from ID05 (c) (y-axis) compared with color-coded hemisphere, cell type or region. PF, prefrontal cortex; F, frontal cortex; P, parietal cortex; T, temporal cortex; O, occipital cortex; mO, medial occipital cortex; I, insular cortex; CC, cingulate cortex; EC, Entorhinal cortex. sqrt-t (AF), square-root transformed allelic fraction. Dendrograms indicate that subcortically derived COUTFII samples cluster together (teal), whereas DLX1 and TBR1 samples derived from the same region cluster together. (d-g) Lolliplots comparing regions (x-axis) with sqrt-t AF (y-axis) for representative MVs. Height of individual lollipop: AF; color: cell type; dashed line: threshold. Next to each lolliplot is the ‘geoclone’ representation of sqrt-t AF shaded intensity (pink) from tissue where detected. Gray boxes: not sampled. (h-i) Standard deviation (SD) of sqrt-t AFs for 146 and 186 MVs in the three different cell types in donor ID01 (h) and ID05 (i), respectively. Each dot: single MV measured in 34 vs. 45 punches across the neocortex in ID01 and ID05, respectively. One-way ANOVA with Tukey’s multiple comparison test with adjusted p-values. PF, prefrontal cortex; F, frontal cortex; P, parietal cortex; O, occipital cortex; T, temporal cortex; I, insular cortex; CC, cingulate cortex; mOC, medial occipital cortex; L, left; R, right.

Next, we measured AF correlations between MVs within each cell type (Supplementary Data 5). As expected, MV clustering confirmed hemispheric restrictions of all three cell types (DLX1⁺, TBR1⁺ and COUTPFII⁺ nuclei). Additionally, MVs in TBR1⁺ nuclei showed many small subclusters enriched in particular lobes, in contrast to widespread distributions of COUPTFII⁺ nuclei. This result was replicated in four independent hemispheres from both donors. A similar pattern was also observed in the global DLX1⁺ nuclear pools, although with less intensity. The data suggests wide distribution of COUPTFII⁺ ventrally-derived cortical inhibitory neuronal clones through tangential migration to the dorsal telencephalon as reported in mice ⁴¹, while focal distribution of TBR1⁺ and at least some DLX1⁺ neurons distribute radially from a dorsal telencephalic, likely radial glial, source. Collectively, if we exclude the very unlike scenario that cortical-derived excitatory and inhibitory neurons originate from ventrally derived progenitor cells that migrate to specific locations together, the result suggests a significant portion of dorsal clones have the potential to differentiate into both cortical excitatory and inhibitory neurons.

Next, we investigated distributions of individual MVs across cell types and cortical areas by mapping each MV AF onto a ‘Lolliplot’ and geographic map, called a ‘Geoclone’ (Supplementary Data 6). As an example, MV 1-64512024-C-T was inhibitory neuron-specific, showing notably uniform enrichment across cortical areas in COUPTFII ⁺ and DLX1⁺ but not TBR1⁺ nuclear pools (Fig. 3d). This suggests a subset of nuclei in DLX1⁺ pools distribute in patterns similar to COUPTFII⁺ nuclei. MV 10-116196503-C-T was present in all three cell types, but with less AF variation in COUPTFII⁺ nuclei than the other two cell types (Fig. 3e), suggesting a wide distribution across cortical regions and tangential migration from a ventral origin. Furthermore, MV 16-69679204-G-C was locally enriched in both DLX1 ⁺ and TBR1⁺ nuclei of the same cortical lobes but absent in COUPTFII⁺ nuclei, implying a subset of DLX1⁺ cells share locally proliferating dorsal telencephalic origins with TBR1⁺ cells (Fig. 3f). Interestingly, we found some MVs (i.e. MV 2-71776656-C-A) that were distinctively enriched in DLX1⁺ nuclei distributed to both prefrontal lobes of ID01 (Fig. 3g), suggesting some inhibitory neurons or their progenitors can populate both hemispheres. However, prefrontal samples were not available in ID05, and more data will be required to confirm these results.

To examine whether COUPTFII⁺ clones are more uniformly distributed than DLX1 or TBR1⁺ clones, standard deviations of AFs of shared MVs in different cortical areas were calculated for the three neuronal populations (Fig. 3h, i). COUPTFII⁺ nuclei standard deviations were significantly lower than the other two cell types in both ID01 and ID05, suggesting that the distribution of CGE-derived inhibitory neuronal clones is wider than excitatory neurons, whereas cortical pan-inhibitory neuronal clones showed a patchy distribution in a fashion similar to excitatory neuronal clones.

We next estimated the proportion of dorsally derived cells within DLX1⁺ nuclei, using the least squares method, to identify the proportion of dorsal and ventral origins that best approximates the AFs of DLX1⁺ nuclei (Extended Data Fig. 7g). This estimation assumes that TBR1⁺ and COUPTFII⁺ nuclei exclusively originate from dorsal and ventral origins, respectively (Methods). The mean proportion of dorsally derived clones across lobes and individuals was 58.875%. This implies more than half of cortical inhibitory neurons may derive from dorsal progenitors.

Dorsal origins of inhibitory neurons

The prior data could not exclude the possibility that the observed genomic similarity between TBR1⁺ and DLX1⁺ sorted nuclear pools derived in part from a rare cell type or from pool contaminants from the sorting protocol. We thus conducted single-cell simultaneous DNA + RNA sequencing (ResolveOME, see Methods), incorporating primary template-directed amplification (PTA) coupled with single-nuclear MPAS (snMPAS) with full-transcript snRNA-seq, in individual NEUN⁺ nuclei from right frontal and temporal cortex in ID05. This allowed for both a priori identification of both cell types and MVs in the same cell ⁴² (Fig. 4a, Supplementary Data 2 and 7). After basic quality control, UMAP clustering with a reference dataset ^43,44 distinguished between cortical excitatory and inhibitory neurons, along with a few minor non-neuronal cell types (Extended Data Fig. 8a–c, Supplementary Data 10). The detection frequency of MVs in single-cell genotyping was positively correlated with AFs in sorted populations, as expected (Extended Data Fig. 8d–f). Informative MVs were detected from a total of 85 excitatory and 33 inhibitory neurons, allowing direct observation of single-cell level MV distribution (Fig. 4b) with 5.863% and 1.052% of false-negative and false-positive rates, respectively (Methods).

Figure 4. snMPAS incorporating snRNA-seq supports the existence of dorsally derived cortical inhibitory neurons in humans.

(a) Frontal and temporal small punches of the right hemisphere of ID05 NEUN⁺ 191 single-nuclei DNA were subjected to primary template-directed amplification (PTA) followed by single nuclear MPAS (snMPAS) genotyping with concurrent snRNA-seq, termed ResolveOME, for analysis. (b) A double-ranked plot divided in half based on brain region (frontal: left, temporal: right). 68 MVs identified in anywhere between 2 to 14 cells (No. of detection). A total of 118 (85 excitatory and 33 inhibitory) neurons plotted. MVs were further classified according to whether they were detected in both frontal and temporal lobes (F-T Shared) vs. a single lobe F-only (purple) or T-only (blue). Dark purple or blue: detected in 3 or more nuclei, light purple or blue: detected in only 2 nuclei. ExN & InN shared: MVs shared between ExNs and InNs exclusively within F-only or T-only MVs. x-axis: cell type, i.e. ExN (green): excitatory neurons; InN1 (orange): All inhibitory neurons except for InN2. InN2 (red): Inhibitory neurons carrying at least one ExN & InN shared MV. For example, MV 13-69308268-A-G (arrow) was detected in 2 nuclei (*), an ExN and an InN. (c) Distribution MV 13-69308268-A-G across cortical areas in ID05. Left: lolliplot and Right: geoclone showing sqrt-t AFs of each sorted nuclear pool in different cortical locations, reproducing single-nucleus data. Dashed line: threshold. (d) Pseudo-bulk analysis after aggregation based on individual MVs (x-axis) and cell types and regions (y-axis). (e) Model for the shared origin of local ExN and InN cortical neurons. Dorsal clones (triangle and square) can produce both TBR1⁺ excitatory neurons and DLX1⁺ inhibitory neurons. Ventrally derived DLX1⁺ inhibitory neurons (circle) tangentially migrate and are more likely to disperse across the cortex.

These genotypes allowed for the assessment of shared MVs in individual excitatory or inhibitory neurons in individual nuclei across various brain regions. We observed numerous excitatory and inhibitory neurons in the same lobe carrying MVs exclusively in the frontal lobe (F-only) or temporal lobe (T-only). As an example, MV 13-69308268-A-G, detected in both excitatory and inhibitory neurons of the right frontal but not temporal lobe (Fig. 4b), was intriguingly enriched in TBR1⁺ and DLX1⁺ nuclear pools with high AFs but below the level of detection in the COUPTFII⁺ nuclear pool (Fig. 4c). Furthermore, overall distribution patterns of this MV in TBR1⁺ and DLX1⁺ nuclei across cortical lobes were very similar to each other, but distinct from that of COUPTFII⁺, supporting this MV is shared between locally born and cortical resident excitatory and inhibitory neurons but underrepresented within cortical inhibitory neurons derived from ventral telencephalic sources (Fig. 4c).

We further generated a phylogeny tree by considering alleles on each of the genomic positions as ‘pseudo-sequence’ for each sample, and a sequence-based phylogenic tree was reconstructed to deconvolve the clonal relationship between single cells (Extended Data Fig. 9 and Methods). Although the overall structure was not very stable from the bootstrap, for the closest branch that has the highest bootstrap values between 2 samples, which in general has higher bootstrap values, we found 3.5 times more ExN-InN pairs in the same lobe (14 pairs) compared to those in different lobes (4 pairs). This result supports evidence that a substantial proportion of human cortical excitatory and inhibitory neurons derive from related cortical progenitor cells.

We also calculated the probability of observing cortical inhibitory neurons carrying seemingly locally enriched MVs by chance using MVs shared in more than two cells in one lobe but none in the other lobe. Among the 33 inhibitory neurons analyzed, we identified 15 with distinct local MVs. These MVs were found exclusively in one lobe and were shared with at least two other local cells, including one excitatory neuron within the same lobe. This observation was highly unlikely to occur by chance (one-tailed permutation test P<0.0001, Extended Data Fig. 8g, Methods), suggesting the existence of locally derived inhibitory neurons from progenitor cells that also produce local excitatory neurons.

Inhibitory neurons carrying locally enriched MVs (InN2; Fig. 4b) showed comparable inhibitory neuronal marker expression (Extended Data Fig. 8h) but a decreased tendency for RNA expression of CGE markers (Extended Data Fig. 8i), or a RELN⁺ neuronal marker (Extended Data Fig. 8j) compared to those not carrying locally enriched MVs (InN1). Instead, InN2 displayed an increased tendency for a parvalbumin-positive (PV⁺) inhibitory neuronal marker (Extended Data Fig. 8k) and some unique transcription patterns (Extended Data Fig. 8l–m). Thus, a significant portion of dorsally-derived inhibitory neurons may contribute to PV⁺ neurons, whereas COUPTFII⁺ neurons contribute to ventrally-derived inhibitory neurons of other classes.

We further observed 21 MVs specific to only local excitatory neurons but not inhibitory neurons in the same lobe (Fig 4b), implying a subset of fate-restricted dorsal telencephalic neural progenitors generate mostly excitatory neurons. This was also supported by pseudo-bulk analysis, by aggregating cells based on each cell type and region, demonstrating local excitatory neuron-specific MVs (Fig. 4d). Interestingly, we could not find evidence of MVs specific to cortical inhibitory neurons within anatomically defined regions, which may be due to their sparse population from total cortical cells, or a limited number of inhibitory neuron-specific MVs. Taken together, single-cell level genotyping incorporating transcriptomics supports the concept that dorsal telencephalic neural progenitor cells may have the potential to generate both excitatory and inhibitory neurons, even amongst progenitor pool predominantly generating excitatory neurons (Fig. 4e).

Anterior-posterior restriction in a lobe

We used this same approach to study clonal dynamics within a single human cerebral lobe. Prior data suggests that clonality between lobes is more restricted than within a lobe ¹⁶. Our prior data suggests a restriction of clonal spread (RCS) along the anterior-posterior axis follows the establishment of the midline RCS, but did not consider cell-type-specific effects ¹⁷. We thus assessed clonal dynamics of DLX1⁺ and TBR1⁺ cells, selecting the parietal lobe for analysis. We performed high-density biopsies from a total of 17 small punches offset by 1cm distances followed by FANS and MPAS genotyping, capable of distinguishing between A-P and D-V RCSs with samples clustered based on AFs of informative MVs (Fig. 5a–b, Supplementary Data 8). Notably, the main clusters (C1 and C2) were formed along the A-P rather than the D-V axis in both cell types in hierarchical clustering (Fig. 5c) and UMAP plots (Extended Data Fig. 10). The RCS dominated along the A-P over the D-V axis in both cell types, represented by MVs 18-33999883-C-T, 4-10646818-G-A, and 3-172635725-G-A (Fig. 5d). Absolute values of normalized AF difference through the A-P axis were larger than those through the D-V axis (Fig. 5e, f). These results suggest that the A-P axis RCS is established prior to D-V in both cell types.

Figure 5. Earlier establishment of the A-P axis compared to D-V restricted clonal spread (RCS) within a cortical lobe.

(a) Workflow for the observation of clonal dynamics in a lobe. A total of 17 punches were radially sampled and subjected to MFNS to assess MVs. The AF of MVs in different sites were mapped onto the geoclones (checkerboard). (b) Heatmap and dendrogram hierarchical cluster of sorted nuclear pools based on sqrt-t AFs of 186 informative MVs from ID05 detected in the right parietal lobe. Sidebars left of the heatmap: cell type and sample location information, according to the convention in Fig. 3. Colors for lobar location taken from (a). Dendrogram highlights two main clusters (C1: blue; C2: red) that when mapped back onto the sampled spatial coordinates (c) are separated in the AP dimension (red and blue circle). (d) Geoclones of three individual MVs from b (box, arrows), showing shades of pink are more different in the A-P axis than in the D-V axis. Gray box: not available. (e-f) Cell-type-resolved contour plots of 94 shared MVs from b (center) with colored kernel density estimation plots (in the periphery) for DLX1 (top) or TBR1 (bottom), showing MVs from both cell types with a greater normalized difference of sqrt-t AFs in A-P than in D-V. Grey: kernel density estimation plot of the orthogonal axis. Arrow head: local peak of the density estimation plot due to MVs with greater difference of AFs in A-P than D-V. Black line: identity line. Dots: individual MVs, large red dot: average across all MVs.

Discussion

Here, using clonal dynamics, we interrogate cellular origins across the neurotypical mature human forebrain. Previous studies relied on the dynamics of single-cell transcriptomic or epigenomic profiles and were not able to reconstruct cellular lineages among human forebrain cell types based on clonality. We found that MVs originating in locally born cellular populations demonstrate stronger lineage restriction within the hippocampus than restriction to either the neocortex or basal ganglia. This is consistent with prior viral barcode tracing in mice showing hippocampal lineage restriction prior to neural tube closure at E9.5, at a time even prior to Prox1 expression ^14,45. We hypothesize that hippocampal lineage restriction may be complete by the time anterior neural plate boundary (aNPB) expresses WNT3A that defines the cortical hem anlage at post-conception week 6 in humans ^46,47. It is also possible that the pallial and subpallial lineage restriction to future subcortical and cortical structures occurs after aNPB definition at the time of dorsal and ventral axis differentiation under the control of BMP signaling ^48,49. Future studies might assess if the early hippocampal lineage restriction we observed here occurs in other vertebrate species, where WNT and BMP-mediated neurulation is well-conserved ^50–52. From the clinical perspective, this restriction may explain why some hemimegalencephaly cases show structural abnormalities in both cortex and basal ganglia ⁵³. More investigations on the distribution of dysplastic cells across the hippocampus, cortex, and basal ganglia can provide insights into brain dysplasias.

While prior studies in mammals suggested most or all cortical inhibitory neurons derive from the ventral telencephalon in mammals ^{3,5,13,22–24}, subsequent studies in non-human primates, human fetal brain, or stem cell culture suggested a potential primate-specific dorsal pallial contribution ^6,27. Our cMVBA data in the mature human brain provides direct evidence for such a dorsal source of cortical inhibitory neurons. Furthermore, both cortical excitatory and inhibitory neurons showed similar clonal MV dynamics within the same lobe, whereas some excitatory neurons had additional local MVs absent in inhibitory neurons at the single-cell level. This implies that early dorsal progenitors can produce both cell types. However, further fate-restriction may occur in some cell populations or at certain times in development that produce exclusively excitatory or inhibitory neurons within the same location. Future studies with additional donors could better generalize our main findings and define the timing of this lineage restriction.

The cell-type-resolved spatial clonal relationship within single cortical lobes was not previously studied in detail due to spatial resolution limits of prior methods. Here, using MFNS, we describe the RCS of excitatory and inhibitory neurons along the A-P prior to the D-V axis within a lobe. This result implies that most cortical neural progenitors maintain potential to produce excitatory and inhibitory neurons even after a lobe is clonally restricted along the A-P axis. From a clinical perspective, examining whether dysplastic cells within a lobe are more likely to spread along the D-V axis than the A-P axis in pediatric focal epilepsy may provide valuable insights for refining surgical strategies.

This study also incorporates several technical advances. We significantly improved experimental and computational MV detection pipelines, which enabled high-quality libraries to be generated from low-abundance cell types. We employed MFNS, which used non-formaldehyde fixative to preserve DNA quality, and modified MPAS, which now incorporates an on-bead first amplification step. The incorporation of MV calling software based on deep-learning identified MVs that conventional MV callers may have overlooked. These protocols increased MV discovery efficiency by an estimated 250-fold, requiring substantially less materials, and providing significantly improved spatial resolution with ultra-high-depth (10,000X) measured AFs across 17 different locations in a single lobe. Also, we incorporate the ResolveOME approach to simultaneously profile single-cell RNA and DNA from archived tissue. Although we estimate a false-positive rate of our MPAS validation method of 5% (Methods), and may be locus-specific, we cannot fully exclude that some discovered MVs could be false. Despite this caution, these approaches could support future investigations into clonal relationships in human tissues at the single-cell level.

Despite our findings, there remain notable disparities when compared to in vitro human inhibitory neuron studies where cortical inhibitory neurons, derived from human dorsal forebrain organoid outer radial glia, exhibited elevated NR2F2 expression ^6,27. In contrast, our study suggests a substantial number of COUPTFII⁺ cortical clones displayed dispersed clonal dynamics, arguing against a purely radial distribution. This discrepancy could reflect differences between in vitro and in vivo experimental conditions and underscores the need for further investigation of cortical inhibitory neuron diversity.

Methods

Donor recruitment.

Organs of ID05 were collected from the UC San Diego Anatomical Material Program (UCSD-21-160). Organs of ID01 were donated from a 70-year-old female, cause of death was ‘global geriatric decline’ with a contributing cause of ‘post-surgical malabsorption’ as documented¹⁷. Organs of ID05 were donated from a 73-year-old female, medical history indicated ‘knee replacement, fractured pelvis, hernia, fractured fibula, hypothyroidism, empyema, pulmonary arterial hypertension, and scleroderma’. Both donors were documented to be of European ancestry. Organs were collected within a 26-hour postmortem interval for both donors (ID01: 24 hrs, ID05: 26 hrs). Prior medical history showed no signs of neurological, psychiatric or cancer diseases for either, and tested negative for infection with HIV, Hepatitis B, or COVID-19.

According to 45 CFR 46.102(e)(1), The use of human anatomical cadaver specimens of ID01 and ID05 are exempt from oversight of the University of California, San Diego Human Research Protections Program (IRB) but are subjected to oversight by the University of California, San Diego Anatomical Materials Review Committee (AMRC). This study was overseen and approved by the AMRC. The approval number is 106135. Donors met AMRC qualifications: (i) Obtain information or biospecimens through intervention or interaction with the individual, and uses, studies, or analyzes the information or biospecimens; or (ii) Obtains, uses, studies, analyzes, or generates identifiable private information or identifiable biospecimens.

Tissue dissection.

All dissection was performed by an anatomical pathologist or neuropathologist. For dissection to capture MVs from ventral telencephalic progenitors of ID01, archived frozen limbic system parts of both hemispheres were sliced at 1 cm thickness, and tissues from Cau, Put, AMG, GP, HIP (CA1-CA3, and DG wherever distinguishable) and thalamus were obtained. For ID05, after the removal of the meninges, ~500 mg of tissues from cerebral cortical regions, Cau, Put, GP, HIP, cerebellum, heart, liver, adrenals, kidneys, and leg skin samples were collected before freezing. Extensive 10-17 sublobar punch biopsies were collected from the right occipital lobe and the right parietal lobe from ID05 with an 8mm skin punch. Sample information summarized in Extended Data Fig. 1 and Supplementary Data 1. The dissection procedure was conducted on ice at room temperature for ID01 and in a cold room for ID05. During dissection, subsamples and the remnants of the large pieces were immediately labeled and snap-frozen on dry ice, and stored at −80°C.

DNA extraction of bulk tissue.

Small cortical biopsies were first cut in half on dry ice. Half of the biopsy was stored as backup and partly used for single nuclei fluorescence-activated nuclei sorting (FANS) for ID01 and ID05. The other half of the cortical biopsy was homogenized with a Pellet Pestle Motor (Kimble, 749540-0000) and resuspended with 450 μL RLT buffer (Qiagen, 40724) in a 1.5 ml microcentrifuge tube (USA Scientific, 1615-5500). The same experimental procedure was carried out on both punches from the cerebellum, heart, liver, and both kidneys. Nuclear preparations were pelleted at 1,000 ×g for 5 min and resuspended with 450 μL RLT buffer in a 1.5 ml microcentrifuge tube. Both homogenates and nuclear preps were then treated with the same protocol: following 1 min vortex, samples were incubated at 70°C for 30 min. 50 μl Bond-Breaker TCEP solution (Thermo Scientific, 77720) and 120 mg stainless steel beads (0.2 mm diameter, Next Advance, SSB02) were added, and cellular/nuclear disruption was performed for 5 min on a DisruptorGenie (Scientific Industries), supernatant was transferred to a DNA Mini Column from an AllPrep DNA/RNA Mini Kit (Qiagen, 80204) and centrifuged at 8500×g for 30 sec, washed with Buffer AW1 (Qiagen, 80204), centrifuged at 8500×g for 30 sec and washed again with Buffer AW2 (Qiagen, 80204), and then centrifuged at full speed for 2 min. DNA was eluted two times with 50 μl of pre-heated (70°C) EB (Qiagen, 80204) through centrifugation at 8,500×g for 1 min as documented previously¹⁷.

Whole-genome library preparation and deep sequencing.

A total of 1.0 μg of extracted DNA was used for PCR-free library construction using the KAPA HyperPrep PCR-Free Library Prep kit (Roche, KK8505). Mechanical shearing using the Covaris microtube system (Covaris, SKU 520053) was performed to generate fragments with a peak size of approximately 400 base pairs (bp). Each fragmented DNA sample went through multiple enzymatic reactions to generate a library in which an Illumina dual index adapter would be ligated to the DNA fragments. Beads-based double size selection was performed to ensure the fragment size of each sample was between 300-600 bp as measured by an Agilent DNA High Sensitivity NGS Fragment Analysis Kit (Agilent, DNF-474-0500). The concentration of ligated fragments in each library was quantified with the KAPA Library Quantification Kits for Illumina platforms (Roche/KAPA Biosystems, KK4824) on a Roche LightCycler 480 Instrument (Roche). Libraries with concentrations of more than 3 nM and fragments with peak size of 400 bp were sequenced on an Illumina NovaSeq 6000 S4 and/or S2 Flow Cell (FC). Each library was sequenced in 6-8 independent pools. For each sequencing run, 24 WGS libraries were normalized to obtain a final concentration of 2 nM using 10 mM Tris-HCl (pH 8 or 8.5; Fisher Scientific, 50-190-8153). 0.5 to 1% PhiX library was spiked into the library pool as a positive control. The normalized libraries in a pool with a total of 311 μl libraries were incubated with 77 μl of 0.2 N Sodium Hydroxyl (NaOH) (VWR, 82023-092) at room temperature for 8 minutes to denature double-stranded DNA. 78 μl of 400 mM Tris-HCl was used to terminate the denaturing process. The denatured library with a final loading concentration of 400 pM in a pool was loaded on the S4 FC using Illumina SBS kits (Illumina, 20012866) with the following setting on the NovaSeq 6000: PE150:S4 FC, dual Index, Read 1:151, Index_Read2:8; Index_Read3:8; Read 4:151. The target for whole genome sequencing with high-quality sequencing raw data was 120 GB or greater with a Q30 >90% per library per sequencing run. In case the first sequencing run generated less than that, additional sequencing was performed by sequencing the same library on a NovaSeq 6000 S2 FC with a 2x101 read length for ID01 as documented¹⁷ and all data were generated at 2x151 read length for ID05. FASTQ files generated with Picard’s (v 2.20.7) SamToFastq command from the DRAGEN platform were used as input for the bioinformatic pipeline for ID01 and bcl2fastq2 (v 2.20) generated FASTQ files from raw sequence files were used for ID05.

Whole-genome sequencing (WGS) data processing.

FASTQ files were then aligned to the human_g1k_v37_decoy genome by BWA’s (v 0.7.17) mem with -K 100000000 -Y parameters. SAM files were compressed to BAM files via SAMtools’s (v 1.7) view command. BAM files were subsequently sorted by SAMBAMBA’s (v 0.7.0) sort command and duplicated reads marked by its markdup command. Reads aligned to the INDEL regions were realigned with GATK’s (v 3.8-1) RealignerTargetCreator and IndelRealigner following the best practice guideline. Base quality scores were recalibrated using GATK’s (v 3.8.1) BaseRecalibrator and PrintReads. Germline heterozygous variants were called by GATK’s (v 3.8.1) HaplotypeCaller. The distribution of library DNA insertion sizes for each sample was summarized by Picard’s (v 2.20.7) CollectInsertSizeMetrics. The depth of coverage of each sample was calculated by BEDTools’s (v2.27.1) coverage command. The code and Snakemake wrapper of the pipeline are freely accessible on GitHub (https://github.com/shishenyxx/Human_Inhibitory_Neurons).

Mosaic SNV/INDEL detection in WGS data.

Mosaic single nucleotide variants/mosaic small (typically below 20 bp) INDELs were called by using a combination of four different computational methods: 1] MosaicHunter (single-mode, v 1.0)⁵⁴ with a posterior mosaic probability >0.05 ^17,55, 2] Single-mode of GATK’s (v 4.0.4) Mutect2⁵⁶ with “PASS” followed by DeepMosaic (v 1.0.1)⁵⁷ 3] Single-mode of GATK’s (v 4.0.4) Mutect2 with “PASS” followed by MosaicForecast (v 8-13-2019)⁵⁸, were implemented for sample-specific or tissue-shared variants; 4] The intersection of variants from the paired-mode of Mutect2 and Strelka2 (v 2.9.2) (set on “pass” for all variant filter criteria)⁵⁹ were collected for sample-specific variants. For the panel of normal samples required for the pipeline of DeepMosaic and MosaicForecast, we employed an in-house panel of similarly (300×) sequenced normal tissues (n = 15 sperm and 11 blood samples from 11 individuals)¹⁷. For ‘tumor’-’normal’ comparisons, required by Mutect2 and Strelka2 pipelines, we employed left-right combined heart tissues as ‘normal’. Variants were excluded if: 1) residing in segmental duplication regions as annotated in the UCSC genome browser (UCSC SegDup) or RepeatMasker regions, 2) residing within a homopolymer or dinucleotide repeat with more than 3 units, or 3) overlapped with annotated germline INDELs. We further removed any variants with a population allele frequency higher than 0.001 in gnomAD (v 2.1.1)⁶⁰. Finally, variants with a lower CI of AF <0.001 were considered noises from reference homozygous and removed. Fractions of mutant alleles (i.e., AF) for variants called in one sample were calculated in all the other samples together with the exact binomial confidence intervals using scripts described below for MPAS analysis. This bioinformatic pipeline yielded a total of 898 candidate MVs for ID01 and 2195 candidate MVs for ID05 (for skin samples because of the clonal nature only 10% of the total calls were randomly selected) that were interrogated with MPAS. Scripts for variant filtering are provided on GitHub (https://github.com/shishenyxx/Human_Inhibitory_Neurons).

Formaldehyde-fixed nuclear preparation for sorting.

For DARPP32/NEUN dual staining with BG nuclei, frozen Cau and Put of ID01 were homogenized in 1% formaldehyde in Dulbecco’s phosphate-buffered saline (DPBS, Corning) using a motorized homogenizer (Fisherbrand PowerGen 125), then incubated on a rocker at room temperature for 10 min, quenched with 0.125 M glycine at room temperature on a rocker for 5 min, then centrifuged at 1,100×g in a swinging bucket centrifuge. The following steps were all performed on ice except where indicated. Homogenates were washed twice with NF1 buffer (10 mM Tris-HCl pH 8.0, 1 mM EDTA, 5mM MgCl₂, 0.1M sucrose, 0.5% Triton X-100 in UltraPure water) and centrifuged at 1,100×g for 5 min at 4°C in a swinging bucket centrifuge. Next, pellets were resuspended in 5 ml NF1 buffer and Dounce homogenized 5x in a 7 ml Wheaton Dounce Tissue Grinder (DWK Life Sciences) using a ‘loose’ pestle. After 30 minutes of incubation on ice, homogenates were Dounce homogenized 20x with a ‘tight’ pestle and filtered through a 70 μm strainer. To remove myelin debris, homogenates were overlaid on a sucrose cushion (1.2M sucrose, 1 M Tris-HCl pH 8.0, 1 mM MgCl₂, 0.1 M DTT) and centrifuged at 3,200×g for 30 min with acceleration and brakes on ‘low’. Pellets of nuclei were washed with NF1 buffer and centrifuged at 1,600×g for 5 min and stored at −80°C, same as documented^17,61.

Nuclear preparation for MFNS or unfixed nuclei sorting.

For the MFNS protocol, approximately 200 mg of freshly frozen tissue stored at −80°C was prepared, and subsequent procedures were conducted using solutions maintained at 4°C. The prepared tissue was homogenized in 300 μl of lysis buffer (composed of 10 mM Tris-HCl, pH 7.4, 10 mM NaCl, 3 mM MgCl2, 0.1% NP-40, and 1 mM Dithiothreitol in nuclease-free water) while kept on ice. Next, additional 9.7 ml of lysis buffer was added to the homogenate, and the mixture was incubated on ice for 5 minutes. The homogenate was then passed through a 70 μm cell strainer (FALCON, 352350) and centrifuged at 1100g for 5 minutes at 4°C. The supernatant was discarded, and the remaining pellet was gently resuspended and washed with 10 ml of sorting buffer (containing 1% bovine serum albumin, 1 mM EDTA, and 10 mM HEPES in 1x HBSS solution). For the density gradient centrifugation step, the resuspended pellet in 25% Iodixanol solution (OptiPrep^™, Millipore D1556) was layered onto a 29% Iodixanol cushion. The centrifugation was carried out at 10,000g with a swinging bucket rotor, employing low acceleration and braking, at 4°C for 40 min, the pellet resuspended in 80% methanol prechilled and stored at −20°C for at least 30 minutes before further use. For samples to perform single-nucleus transcriptome and PTA with ResolveOME, brain tissue from participant ID05 were homogenized in 1 ml of ice-cold NIB (composed of 0.25 M sucrose, 25 mM KCl, 5 mM MgCl2, 10 mM Tris pH 7.5, 100 mM DTT, and 0.1% Triton X-100) and subjected to homogenization, incubated on a rocker for 5 min at 4 °C and then centrifuged at 1,000g employing low acceleration and braking in a swinging-bucket centrifuge, and the pellet was reconstituted in 0.5 ml of sorting buffer and filtered through a 70-um strainer. Nuclei in the flow-through were immediately subjected to staining.

Fluorescence-activated nuclear sorting.

For pellets of defrosted and homogenized brain nuclei were washed twice in sorting buffer and then re-suspended in 0.2 ml sorting buffer and incubated overnight at 4°C. The following antibodies were used: NEUN Alexa Fluor 488 (1:2,500; Millipore Sigma, MAB377), TBR1 unconjugated (1:1,000; Abcam, ab31940), DLX1 (1:200; Atlas Antibodies, HPA045884), COUPTFII (1:400; Novus Biologicals, PP-H7147-00), and DARPP32 (1:400; Abcam, ab40801). The following day, nuclei were washed with staining buffer and in case an unconjugated antibody was used, nuclei were stained subsequently for 30 minutes with goat anti-rabbit Alexa 647 (1:4,000; ThermoFisher Scientific, A21244) for TBR1, DLX1 and DARPP32, or donkey anti-mouse Alexa 647 (1:4000; ThermoFisher Scientific, A32787) for COUPTFII. Stained nuclei were washed one more time with staining buffer and passed through a 70 μm strainer. Immediately before the sort, nuclei were stained with 0.5 μg/ml DAPI. For single-nucleus transcriptome and whole-genome amplification with ResolveOME, nuclei were stained with Propidium Iodide (1:20; Invitrogen, BMS500PI). Nuclei for the cell type of origin were sorted either on a MoFlo Astrio EQ sorter (Beckman Coulter), BD FACSAria II (Becton-Dickinson), or BD InFlux Cytometer (Becton-Dickinson) similar to previous documentation¹⁷. At least 1000 methanol-fixed or >50,000 formaldehyde-fixed sorted nuclei were pooled in each tube. Sorted nuclei were pelleted in staining buffer at 1,600×g for 10 minutes. Nuclei for DNA extraction and bisulfite sequencing were stored at −80°C. FANS data was visualized using FlowJo v10 software (Ashland, Oregon). Following MPAS (see below) sorted populations were deemed to be of sufficient overall quality if at least 95% variants were sequenced above >1,000×.

Low-input DNA extraction from sorted nuclei.

For low-input DNA extraction from sorted nuclei, we further developed an on-bead DNA extraction method: sorted nuclei were centrifuged down for 1 min at 1,000g (4°C) in a 200 μl PCR tube and resuspended in 20 μl lysis buffer that consists of 30mM Tris-HCl (pH 8.0), 0.5%(v/v) Tween-20 (Sigma-Aldrich P1379), 0.5%(v/v) IGEPAL CA-630 (Sigma I8896), 1.25 μg/ml protease (Qiagen, 19155) as final concentration in nuclease-free water (Ambion, AM9937). The mixture was lightly vortexed for 10 sec and centrifuged at 1,500g for 1 min (4°C). The tube was then subjected to 50°C for 12 min and 75°C for 30 min on a thermocycler. To each lysate, 20 μl of Agencourt AMPure XP beads (Beckman Coulter, A63881) was added. The final AMPure beads to sample ratio was 1:1. After pipetting to achieve mixing, the mixture was left at room temperature for 5 min, placed on a magnet for 5 min, and the supernatant removed. The remaining material was washed twice with 150 μl of 80% (v/v) ethanol for 1 min each. The DNA was then suspended in Low-TE solution and kept at −20°C until Bisulfite sequencing or MPAS.

Bisulfite sequencing of sorted nuclei for cell type of origin.

The low-TE resuspension including DNA and beads were processed by Pico Methyl-Seq Library Prep Kit (Zymo Research, D5455) to generate bisulfite sequencing libraries. Samples were sequenced at PE150 on the Illumina NovaSeq 6000 platform.

Bisulfite sequencing data processing and data visualization.

FASTQ files were analyzed with the Bismark bisulfite mapper and methylation marker (v 0.23.1)⁶², and read pairs were treated as singletons according to the developers’ suggestions. The pipeline was run with Snakemake (v 6.12.3), and bedgraph were generated with BEDTools (v 2.30.0). Python (v 3.10) with argparse, textwrap, and numpy packages and R (4.1.3) with lsa, pheatmap, ggfortify packages were used for visualization. Human single-cell methylation data from published literature³² was downloaded, and reads from excitatory neurons and inhibitory neurons were pooled separately and used as positive controls according to the original authors’ cell-type labels. Codes for plotting the cosine similarity and hierarchical clustering of methylation patterns of gold-standard inhibitory neurons, sorted DLX1 positive cells, excitatory neurons, and sorted TBR1 positive cells. Control samples from the heart, cortical microglia, and cortical oligodendrocytes were available on GitHub (https://github.com/shishenyxx/Human_Inhibitory_Neurons).

Single nucleus transcriptome and PTA.

After sorting, a total of 192 single nuclei from each of 2 samples (right temporal and frontal cortex of ID05) were snap-frozen on dry ice. Nuclei underwent the ResolveOME^™ workflow (BioSkryb Genomics Inc.). Briefly, Biotin-dT-primed first strand cDNA was generated. After termination of the reaction and nuclear lysis, whole genome amplification with primary template-directed amplification was performed ⁶³. The mRNA-derived cDNA was affinity purified with streptavidin beads from the combined pool of cDNA and amplified genome. Remaining cDNA were pre-amplified on beads. Independently, amplified cDNA and single-cell genomic DNA from each cell underwent SPRI (Beckman Coulter, B23319) cleanup prior to library preparation. Illumina libraries were prepared using the ResolveOME library preparation kit (BioSkryb Genomics Inc.) with NEXTFLEX Unique Dual Index Barcodes (PerkinElmer Applied Genomics, NOVA-534100). Libraries were sequenced at low-pass (2x 50bp paired end) targeting 2 million reads on a NextSeq (Illumina) instrument. Libraries of interest were identified based on QC sequencing, and were subsequently sequenced at paired-end 150 bp (DNA libraries) and paired-end 100 bp (RNA derived libraries) on a NovaSeq X Plus (Illumina) platform.

Massive parallel amplicon sequencing (MPAS) and single nucleus MPAS (snMPAS).

Two customized AmpliSeq Custom DNA Panel for Illumina (20020495, Illumina, San Diego, CA, USA) were used for MPAS for ID01 (#203019), as well as MPAS for ID05 (#201745), respectively. Designed genomic regions are provided in Supplementary Data 9. A list of 259 MVs used in our previous study and 639 candidate MVs from the MV detection pipeline in ID01 described above was subjected to the AmpliSeq design system. For the first panel, we randomly selected 120 high-confidence heterozygous variants as positive controls. These heterozygous variants presented with estimated AFs between 48-52% for all the 25 sequenced bulk tissues, and with read depths between 270-330×. Of the 120 variants, 45 were private variants and 75 were present in gnomAD at different population allele frequencies. We also randomly selected 30 reference homozygous variants as negative controls. These reference homozygous variants presented with ~0% AF across all sequenced samples, with average depth 270-330×, and gnomAD (v 2.1.1) allele frequency >0.5 to exclude any potential contamination or amplification bias. For the second panel, 2195 candidate MVs detected from ID05 as well as 152 randomly chosen variants detected as heterozygous and 59 as alternative homozygous in ID05 were subjected to the AmpliSeq design system. DNA from extracted tissue, amplified single nuclei, and a duplicate unrelated control sample was diluted to 5 ng/μl in low TE provided in AmpliSeq Library PLUS (384 Reactions) kit (Illumina, 20019103). For sorted nuclei, The low-TE resuspension including DNA and beads stored after ‘Low-input DNA extraction from sorted nuclei’ step mentioned above was prepared. AmpliSeq was carried out following the manufacturer’s protocol (document 1000000036408v07). For amplification in bulk samples, 14 cycles each with 8 minutes were used, for amplification in low-input sorted nuclei on beads, additional cycles and time is added accordingly based on the recommended table from the manufacturer. After amplification and FUPA treatment, libraries were barcoded with AmpliSeq CD Indexes (Illumina, 20031676) and pooled with similar molecular numbers based on measurements made with a Qubit dsDNA High Sensitivity kit (Thermo Fisher Scientific, Q32854) and a plate reader (Eppendorf, PlateReader AF2200). To avoid index hopping, the two MPAS library pools for ID01 and ID05 and the snMPAS library pool for ID05) were sequenced on separate lanes on different NovaSeq 6000 runs. 192 GB of FASTQ data were obtained from the ID01 MPAS libraries, 339 GB of FASTQ data were obtained from the ID05 MPAS libraries, 193 GB of FASTQ data were obtained from the ID05 snMPAS libraries, all aiming for an average of 10000× for each variant.

Data analysis for MPAS and snMPAS.

Raw reads from MPAS and snMPAS were mapped to the human_g1k_v37_decoy genome with BWA’s (v 0.7.17) mem command. BAM files were processed without removing PCR duplicates. Reads near insertion/deletions were re-aligned with GATK’s (v 3.8.1) IndelRealigner and base qualities scores were recalibrated with GATK’s (v 3.8.1) BaseRecalibrator. The final BAM files were parsed by SAMtools’s (v 1.7) mpileup and the 95% confidence intervals (CIs) of the measured AF of all the candidate MVs, together with the homozygous (negative control) and heterozygous (positive control) variants were estimated based on an exact binomial estimation (https://github.com/shishenyxx/Human_Inhibitory_Neurons). Following depth calculation, regions of 639 mosaic candidates, 120 heterozygous variants (positive controls), and 30 homozygous variants (negative controls) were detected and subjected to the next genotyping steps with 259 previously validated MVs ¹⁷ for ID01, and 2195 mosaic candidates, 152 heterozygous variants (positive controls), and 59 homozygous variants for ID05. The genotypes of candidate MVs from MPAS were determined by comparing them to the AF distribution of the reference homozygous and heterozygous variants. The exact binomial lower bounds of all reference homozygous variants with >30 read depth were estimated and the 95% single-tail confidence threshold for the lower bound was calculated to be 0.001677998 (ID01) and 0.002360687 (ID05). The distribution of the exact binomial upper bound of all heterozygous variants was calculated and 0.4706769 (ID01) and 0.4779163 (ID05) were the threshold for the upper bond based on ~5% false discovery rate (FDR) based on the built-in heterozygous and alternative homozygous genomic positions. Mosaic candidates from WGS were considered positive if variants met the following criteria at the same DNA samples: 1) the 95% exact binomial lower bound was >0.001677998 (ID01) or >0.002360687 (ID05), respectively, 2) the lower CI of the unrelated control sample was <0.001677998 (ID01) or <0.002360687 (ID05), 3) the 95% exact binomial upper bound was <0.4706769 (ID01) or <0.4779163 (ID05), respectively, 4) the sequencing depth was >30, and 5) the assessed alternative allele was supported by ≥3 reads. 6) for ID01, variants are detected in heart, GP, Cau, Put, and Thal in WGS. These criteria ensured the FDR for each variant was under 5%. After the MPAS quantification, 328 MVs (28 newly positively validated MVs from brains, 41 from heart and original 259 variants¹⁷) from ID01 and 780 MVs from ID05 were considered positively validated in the sample where the variant was originally detected. The validation rate was 36.5% (328/898) and 35.5% (780/2195) in ID01 and ID05, respectively. 287 out of 328 MVs excluding 41 heart MVs from ID01 and all 780 MVs from ID05 were used for all the analysis presented throughout the manuscript. In snMPAS, mosaic candidates from WGS were considered positive if the lower CI of AF was larger than the upper CI of AF in the normal control sample. To assess precision of snMPAS data, we first calculated the chance of the failure of detection of true-positive variant calls (dropout), using heterozygous variants. Among AF values of the variants known as heterozygous in ID05 bulk and homozygous in control bulk samples, 94.137% (5379/5714) of those were recovered as heterozygous in snMPAS, resulting in an estimated false-negative rate of 5.863%. Next, we calculated the chance of false-positive discovery using homozygous genomic sites. The known homozygous genomic positions in bulk samples recovered as homozygous were 98.948% (7149/ 7225), resulting in an estimated false-positive rate of 1.052%.

Mutational signature analysis.

Mutational signature analysis was performed using a web-based somatic mutation analysis toolkit (Mutalisk ⁶⁴). PCAWG SigProfiler full screening model was used.

Computational deconvolution for DLX1⁺ populations.

Given the consistency of the FANS sorting strategy and the consistent contribution in each cell type carrying the same mutation, we assume that the contamination rate α for TBR1⁺ in DLX1⁺ nuclei is the same across different samples. For variant number n, the vector of observed variant allelic fraction in DLX1 is,

$$\begin{gathered} {AF}_{DLX1-observed}=\left\{\begin{array}{c}{AF}_{{DLX1-observed}_1}\\ {AF}_{{DLX1-observed}_2}\\ \cdots \\ {AF}_{{DLX1-observed}_n}\end{array}\right\} \\ =\left(1-\text{α}\right)\cdot {AF}_{DLX1-theoretical}+\text{α}\cdot {AF}_{TBR1-observed} \\ =\left(1-\text{α}\right)\cdot \left\{\begin{array}{c}{AF}_{{DLX1-theoretical}_1}\\ {AF}_{{DLX1-theoretical}_2}\\ \cdots \\ {AF}_{{DLX1-theoretical}_n}\end{array}\right\}+\text{α}\cdot \left\{\begin{array}{c}{AF}_{{TBR1-observed}_1}\\ {AF}_{{TBR1-observed}_2}\\ \cdots \\ {AF}_{{TBR1-observed}_n}\end{array}\right\} \end{gathered}$$

Thus, the theoretical allelic fraction rates for each DLX1 sorted population in the same lobe is calculated as,

$${AF}_{DLX1-theoretical}=\left\{\begin{array}{c}\frac{{AF}_{{DLX1-observed}_1}-\text{α}·{AF}_{{TBR1-observed}_1}}{1-\text{α}}\\ \frac{{AF}_{{DLX1-observed}_2}-\text{α}·{AF}_{{TBR1-observed}_2}}{1-\text{α}}\\ \cdots \\ \frac{{AF}_{{DLX1-observed}_n}-\text{α}·{AF}_{{TBR1-observed}_n}}{1-\text{α}}\end{array}\right\}$$

The contamination rate α was set as 0.25 for Extended Data Fig. 7c and d, based on the fraction of excitatory neuronal fractions in DLX1⁺ nuclei described in Fig. 1d.

Simulated TBR1⁺ nuclei contamination for COUPTFII⁺ populations.

Given that n is the number of variants, and α is the fraction of excitatory neurons in sorted DLX1⁺ nuclei, the AFs of artificial contaminated COUPTII⁺ nuclei contaminated by TBR1⁺ nuclei (${AF}_{COUPTF\mathit{2}\text{-}mixture}$) were generated by mixing α fraction of TBR1⁺ nuclei AFs (${AF}_{COUPTF\mathit{2}\text{-}observed}$) and 1 – α fraction of COUPTFII⁺ nuclei AFs (${AF}_{TBR\mathit{1}\text{-}observed}$),

$${AF}_{COUPTF2-mixture}=\left\{\begin{array}{c}(1-\text{α})\cdot {AF}_{{COUPTF2-observed}_1}+\text{α}\cdot {AF}_{{TBR1-observed}_1}\\ (1-\text{α})\cdot {AF}_{{COUPTF2-observed}_2}+\text{α}\cdot {AF}_{{TBR1-observed}_2}\\ \cdots \\ (1-\text{α})\cdot {AF}_{{COUPTF2-observed}_n}+\text{α}\cdot {AF}_{{TBR1-observed}_n}\end{array}\right\}$$

The fraction α was set as 0.25 for Extended Data Fig. 7e–f based on the portion of excitatory neurons in DLX1⁺ nuclei described in Fig. 1d.

Estimating the contribution of dorsal and ventral origin for DLX1⁺ inhibitory neurons

After computational decontamination for DLX1⁺ nuclear pools, we obtained ${AF}_{DLX1-theoretical}$, to further estimate the contribution from dorsally- (${AF}_{Dorsal}$) and ventrally (${AF}_{Ventral}$)-derived clones assuming there is no third origin, we introduce $\mathrm{\beta }$ as the proportion of dorsally derived clones, we have

$$\begin{gathered} {AF}_{DLX1-theortical}=\left\{\begin{array}{c}{AF}_{{DLX1-theoretical}_1}\\ {AF}_{{DLX1-theoretical}_2}\\ \cdots \\ {AF}_{{DLX1-theoretical}_n}\end{array}\right\} \\ =\text{β}\cdot {AF}_{Dorsal}+(1-\text{β})\cdot {AF}_{Ventral} \\ =\text{β}\cdot \left\{\begin{array}{c}{AF}_{{Dorsal}_1}\\ {AF}_{{Dorsal}_2}\\ \cdots \\ {AF}_{{Dorsal}_n}\end{array}\right\}+(1-\text{β})\cdot \left\{\begin{array}{c}{AF}_{{Ventral}_1}\\ {AF}_{{Ventral}_2}\\ \cdots \\ {AF}_{{Ventral}_n}\end{array}\right\} \end{gathered}$$

The loss function $L\left(\text{β}\right)$ for the estimation is defined as the residual sum of squares,

$$\begin{gathered} L(\text{β})=\sum _{i=1}^n{({AF}_{{estimated}_i}-{AF}_{{DLX1-theoretical}_i})}^2 \\ =\sum _{i=1}^n{(\text{β}\cdot {AF}_{{Dorsal}_i}+(1-\text{β})\cdot {AF}_{{Ventral}_i}-{AF}_{{DLX1-theoretical}_i})}^2 \end{gathered}$$

For each lobe, the optimal contribution of $\text{β}$ (ranging from 0 to 1, step 0.01) is obtained by minimizing $L\left(\mathrm{\beta }\right)$. ${AF}_{Dorsal}$ is represented by ${AF}_{TBR1-observed}$ given the known dorsal origin and ${AF}_{Ventral}$ is represented by ${AF}_{COUPTF2-observed}$ given the high potential of ventral origin demonstrated in Fig. 3, respectively.

Permutation test for the significance of the snMPAS result.

Regarding MVs presented more than 2 nuclei and a maximum of 14 nuclei from Fig. 4b, detection events of each MV were randomly re-assigned within a total of 118 nuclei (16 and 17 InNs in F and T respectively; 40 and 45 ExNs in F and T, respectively), maintaining the original detection frequency. The number of inhibitory neurons sharing MVs exclusively in one lobe and shared with at least two other local cells, including one excitatory neuron, were used as the outcome and a null distribution was generated from the 10,000 permutations. The probability of having more than or equal to 15 inhibitory neurons sharing MVs exclusively in one lobe and shared with at least two other local cells, including one excitatory neuron was calculated and used as the one-tailed permutation p-value.

Single-nucleus RNA sequencing with Chromium platform.

Nuclei that underwent MFNS (16000 nuclei) were resuspended in a sorting buffer to make the desired concentration (800~1000 nuclei/ul) targeting 10000 nuclei per reaction. Gel beads emulsion (GEM) generation, cDNA, and sequencing library constructions were performed in accordance with instructions in the Chromium Single Cell 3’ Reagent Kits User Guide (v3.1). Each library pool was sequenced with 200 million read pairs using NovaSeq 6000.

Single-nucleus RNA-seq bioinformatics pipeline.

For the snRNA-seq data made with Chromium platform, fastq files from single-nucleus libraries were processed through Cell Ranger (v6.0.2) analysis pipeline with –include-introns option and hg19 reference genome. Seurat (v4.0.5) package was used to handle single nuclei data objects. Nuclei passed a control filter (nCount > 400, nFeature_RNA < 2000, percentage of mitochondrial gene < 10%) was used for downstream analysis. A total of 15,896 protein-coding genes were used for further downstream analysis. To balance the nucleus number for each group, a total of 500 random nuclei for each DAPI⁺, DLX1⁺, TBR1⁺ and COUPTFII⁺ group were selected. Data were normalized and scaled with the most variable 1000 features using the ‘ScaleData’ functions. Dimensionality reduction by PCA and UMAP embedding was performed using runPCA and runUMAP functions. Clustering was performed by FindNeighbors and FindClusters functions. For the full-length transcript snRNA-seq through ResolveOME platform, quality control were carried out for raw FASTQ from single nucleus RNA sequencing from ResolveOME files using fastqc (v0.11.8). Preprocessing was carried out with cutadapt (v1.16). Cleaned FASTQ files were aligned to GRCh38 human genome and gencode (v27) gtf annotation using STAR (v2.6.0c). Aligned BAM files were indexed with SAMtools (v1.7). PCR duplicates were marked with Picard (v2.20.7) MarkDuplicates. Post alignment quality control was carried out with Picard (v2.20.7) CollectRnaSeqMetrics, CollectInsertSizeMetrics, CollectGcBiasMetrics as well as qualimap (v2.2.2-dev). Raw read counts were collected with featureCounts (v2.0.0). Transcripts were collected with rsem (v1.3.1) with seed 12345 using the same gtf file. We also employed trimmed-means data from Human Multiple Cortical Areas SMART-seq dataset (https://portal.brain-map.org/atlases-and-data/rnaseq/human-multiple-cortical-areas-smart-seq) for reference group. Seurat (v4.0.5) package was used for further analysis. Nuclei passed a control filter (nCount > 1000, nFeature_RNA >500, percentage of mitochondrial gene < 30%) was used for downstream analysis (Supplementary Data 10). SCTransform function was used for normalizing and scaling data. For both snRNA-seq data made with Chromium platform and ResolveOME, cell type identification was performed using known cell type markers expressed in the brain including excitatory (RORB, CUX2, SATB2), inhibitory neuron (GAD1, GAD2), astrocyte (SLC1A2, SLC1A3), oligodendrocyte (MOBP, PLP1), oligodendrocyte precursor cell (PDGFRA), microglia (PTPRC), and endothelial cell markers (CLDN5, ID1) as well as using positive markers found by FindAllMarkers function with 1000 most variable features in scaled data. The final visualization of various snRNA-seq data was performed by r-ggplot2 (v3.3.5).

Phylogenetic tree analysis.

From the 68 MVs in 118 cell-type-resolved single nuclei in Fig. 4b, the alleles at each genomic position were combined into one ‘pseudo-sequence’ for each sample, and a sequenced-based phylogenic tree was reconstructed to deconvolve the clonal relationship between single cells. Multiple sequence alignments were carried out with MUSCLE ⁶⁵, and maximum likelihood phylogenetic tree was constructed with Molecular Evolutionary Genetics Analysis (MEGA, v11.0.13⁶⁶. Maximum likelihood fit-based model selection was carried out for 24 different nucleotide substitution models, and Kimura 2-parameter model was elected as the best-performing substitution model. After 1000 bootstrap replications, and Nearest-neighbor-interchange heuristic method, the bootstrap consensus phylogenetic tree was generated, as shown in Extended Data Fig. 9. Cell types were labeled based on transcriptomic information from the same cell.

Statistical tests and packages for customized plots.

Hierarchical clustering with p-value via bootstrap resampling was performed using r-pvclust (v2.2.0) package. Pearson’s product-moment correlation is calculated using cor.test() in R. One-way ANOVA with Tukey multiple comparisons of means was performed with aov() and TukeyHSD() function. Various heatmaps with dendrograms and sidebars were generated by ComplexHeatmap (v2.16.0) package. Various plots including violine plots, scatter plots, contour plots, bar plots, upset plots lolliplots were generated using r-ggplot2 (v3.4.3). The oncoplot was generated using maftools (v2.16.0) in R. UMAP analysis was performed with r-umap (v0.2.10.0) package. Single-nucleus RNA-seq data was analyzed and plotted using Seurat4 (4.2.0) package in R.

Extended Data

Extended Data Fig. 1. Tissues collected from ID01 and ID05.

Red dots indicate approximate sites of punch biopsies. Abbreviations: PF, prefrontal cortex; F, frontal cortex; P, parietal cortex; O, occipital cortex; T, temporal cortex; I, insular cortex; Cb, Cerebellum; CC, cingulate cortex; mOC, medial occipital cortex; Cau, Caudate; Put, Putamen; Thal, Thalamus; GP, globus pallidus; EC, entorhinal cortex; HIP, hippocampus; AMG, amygdala; POA, preoptic area; Cl, Claustrum.

Extended Data Fig. 2. Bisulfite sequencing in sorted TBR1⁺ and DLX1⁺ nuclear pools correlate with excitatory and inhibitory neuron methylome signatures.

(a-c) MFNS gating strategy on 30,000 single brain nuclei using DLX1, TBR1, and COUPTFII antibodies. X-axis: 488 channel intensity for monitoring autofluorescence signals. Y axis: Fluorescence intensity from antigen-bound antibodies. (d) snRNA-seq of post-MFNS nuclei confirming enrichment of targeted nuclear types. (e) Marker expression in assigned nuclear types correlating with targeted nuclear types. (f) UMAP plot NR2F2 expression pattern (encoding COUPTFII) highlighting a subpopulation of inhibitory neurons (compare with Fig. 1c). (g-i) Reference excitatory and inhibitory neuronal methylome signatures (aggregated from an available public single-nuclei methylome dataset) compared to methylome signatures of sorted nuclei and a bulk heart sample. Normalized relative methylation levels (y-axes) and genomic positions (x-axes) of genes listed at top. (g) Methylome signature of SLC17A7 encoding VGLUT1, an excitatory neuronal marker in the brain, showing reduced methylation (i.e. representing activation) across the gene body and especially near the transcription start site (TSS, red box) in TBR1⁺ excitatory neuron samples. (h) Methylome signature of SLC6A1 encoding VGAT1, an inhibitory neuronal marker in the brain, showing reduced methylation across the gene body and especially near the TSS (red box). (i) Methylome signature of RBFOX3 encoding NEUN, a mature neuronal marker in the brain, showing reduced methylation at the TSS in neurons compared with bulk heart. Ref ExN, reference excitatory neurons; Ref InN, reference inhibitory neurons. (j) Heatmap and dendrograms based on cosine similarities of global methylation patterns between groups. Two different TBR1⁺ or DLX1⁺ samples were aggregated. The TBR1⁺ nuclear pool was clustered with Ref ExN while the DLX1⁺ clustered near the pool with Ref InN. The control heart bulk sample was distant from either group.

Extended Data Fig. 3. Quality controls of MPAS results.

(a-b) Violin plot distribution of log-transformed total read depths (y-axes) of individual variant positions in 321 or 147 samples from ID01 or ID05 (x-axes), respectively. (c-d) Correlation between sqrt-t AF of individual variants from WGS and MPAS. Blue horizontal dashed lines: Lower bound for binomial distribution detection threshold. r and p-values (two-tailed) from Pearson’s Product-Moment correlation. Identity lines (red).

Extended Data Fig. 4. Basic characteristics of positively validated MVs from the cMVBA pipeline.

(a) ID01. (b) ID05. CTX, cortex; BG, Basal ganglia; THAL, thalamus; HIP, hippocampus; AMG, amygdala; CB, cerebellum; SUB, subiculum; CLA, Claustrum; POA, preoptic area; OLF, olfactory bulb. (c) Mutational signature analysis using 368 brain-specific sSNVs from ID01 and ID05 using Mutalisk. Clonal sSNVs show clock-like signatures such as SBS 1 and 5, reflecting embryonic developmental origins. (d-e) AF distributions of organ-shared early embryonic MVs in ID01(d) and ID05 (e) reflect the asymmetric clonal division in early human embryos. Vertical dashed lines (red): expected peaks (AF=25%) from the first symmetric cell division, absent in observed distribution, suggesting asymmetric divisions.

Extended Data Fig. 5. UMAP relationships between samples from the brain based on AFs of validated MVs.

Clustering by the same hemisphere validates lateralization of brain-derived cell clones except for the independent origin of microglia (marked by PU.1, arrow). (a-c) UMAP clustering in ID01 samples labeled by (a) cell type, (b) gross region, or (c) subregion, respectively. Clustered samples tend to show similar AF patterns. (d-f) UMAP clustering in ID05 samples labeled by (d) cell type, (e) region, or (f) subregion, respectively. Although PU.1 cells were not sorted in ID05, other findings are similar between ID01 and ID05.

Extended Data Fig. 6. Evidence for HIP lineage restriction occurring prior to CTX or BG in ID01 sorted nuclear pools.

(a) Heatmap with 17 sorted nuclear samples based on sqrt-t AFs of 121 informative MVs from ID01, similar to Fig. 2c, showing greater HIP lineage separation compared with CTX or BG (purple compared with green or yellow). (b) Contour plot (at center) with 121 informative MVs derived from (a) and two kernel density estimation plots (at periphery). Axes show the absolute normalized difference value for each MV between the average AF of CTX and BG (CTX-BG) or CTX and HIP regions (CTX-HIP). Solid line: identity. Red dot: averaged x and y values of individual data points. sqrt-t AF, square-root transformed allele fraction; CTX, cortex; BG, basal ganglia; HIP, hippocampus; Cau, caudate; DG, dentate gyrus; HIP and Hip, hippocampal tissue; I, insular cortex; O, occipital cortex; P, parietal cortex; PF, prefrontal cortex; Put, putamen; T, temporal cortex; GP, globus pallidus.

Extended Data Fig. 7. Confidence for dendrograms with sorted nuclei from cortical areas.

(a) Bootstrapping results of ID01. (b) Bootstrapping results of ID05. The percentage of 10,000 replicates showing relationships between sqrt-t AFs for TBR1⁺ and DLX1⁺ nuclei in the same geographic region were more similar than TBR1⁺ nuclei from two different geographic regions (arrow for example). COUPTFII⁺ nuclei clustered among themselves, outside of the DLX1 and TBR1 clusters. Approximated unbiased p-value > 95% (red): the hypothesis “the cluster does not exist” rejected with a significance level (< 5%). (c-d) Heatmaps and hierarchical clustering results after computational deconvolution of DLX1⁺ nuclei (grey) from Fig. 3b (c) and Fig. 3c (d). (e-f) Heatmaps and hierarchical clustering results after the simulated TBR1⁺ nuclei contamination for COUPTFII⁺ nuclear pools (black) from Fig. 3b (e) and Fig. 3c (f). (g) The estimated proportion of dorsally derived cortical inhibitory neurons within deconvolved DLX1⁺ nuclei of each lobe. The least square method is used (Methods). 11, 13 cortical lobes for ID01 and ID05, respectively. Median, thick horizontal line at the center; 95% confidence intervals, the notch of the box plot; 75 and 25% quantiles of data, upper and lower bounds of the box; Whiskers, maxima and minima excluding outliers.

Extended Data Fig. 8. Quality controls of the ResolveOME dataset in ID05.

(a) A UMAP plot of snRNA-seq using 225 NEUN ⁺ nuclei and 121 aggregated reference cell types. F, frontal; T, temporal; HIP, hippocampus; REF, reference dataset. (b) UMAP labeled by cell types. Note that UMAP clusters separate by cell type (ExN, InN or Other) more than by location. (c) Relative expression of cell type markers within clusters, confirming cell identity. (d) Hierarchical clustering based on sqrt-t AFs of 34 informative MVs shared in 5 to 29 cells in single-nuclear data. F- NEUN, sorted frontal NEUN⁺ nuclei pool; F-sc, pseudo-bulk snMPAS data from a frontal lobe punch; T-sc, snMPAS data from a frontal (F) lobe punch. (e) Correlation between sqrt-t AFs of MVs between F- NEUN and F-sc. (f) Correlation between sqrt-t AFs of MVs between F- NEUN and T-sc. In e and f, linear regression with upper and lower 95% prediction intervals displayed by blue solid lines and gray surrounding area; sqrt-t (AF), sqrt-t AF. Pearson’s Product-Moment correlation with r and p-values (two-tailed) in e and f. (g) Null distribution of the frequency of the number of inhibitory neurons carrying MVs exclusively detected in one lobe and shared with at least two other local cells, including one excitatory neuron within the same lobe. 10,000 permutations. The portion to the right of the red dashed line, compared to the entire distribution, represents the probability (p < 0.0001, one-tailed permutation test) of having 15 or more InNs. (h-m) RNA expression levels of informative genes between InN1 (n = 17) and InN2 (n = 16) (Fig 4b) in snRNA-seq. (h) Comparable expression levels of inhibitory neuronal markers between both groups. (i) Decreased tendency for the expression of CGE-derived cell markers in InN2 compared to InN1, implying COUPTFII+ inhibitory neurons are unlikely InN1, consistent with previous observations in sorted nuclear populations. (j) RELN⁺ inhibitory neuronal marker showed decreased expression tendency in InN2 compared to InN1. (k) Increased expression tendency for parvalbumin-positive (PV⁺) inhibitory neuronal marker in InN2 compared to InN1, implying dorsally derived inhibitory neurons include PV⁺ neurons. (l, m) top 3 genes increased (l) or decreased (m) in InN2 compared to InN1 among the most variable 3000 protein-coding genes.

Extended Data Fig. 9. Phylogenic tree analysis.

(a) Phylogenic tree generated after 1000 bootstrap replications based on the 68 MVs in 118 single nuclei in Fig. 4b. Bootstrap values supporting each edge are labeled beside branches of the tree. (b) The number of pairs diverging from the latest branch that has the local highest-confident edge is shown based on the lobe and cell type. For example, the number of excitatory-excitatory neuron pairs within the same lobe clustered with the local highest-confident edge was 20.

Extended Data Fig. 10. UMAP plots with sorted nuclear pools based on sqrt-t AFs of 186 informative MVs from Fig. 5.

Colors of data points correspond to the spatial information in the grey box.

Data availability statement

Raw whole genome sequencing and massive parallel amplicon sequencing data (MPAS) and single nucleus MPAS (snMPAS) are available through SRA (accession number: PRJNA799597) and NDA (accession number: study 919) for ID01 and ID05. The 300× WGS panel of normal is available on SRA (accession number: PRJNA660493).

human_g1k_v37 reference: http://ftp.1000genomes.ebi.ac.uk/vol1/ftp/technical/reference/

gnomAD: https://gnomad.broadinstitute.org/

Human Multiple Cortical Areas SMART-seq dataset: https://portal.brain-map.org/atlases-and-data/rnaseq/human-multiple-cortical-areas-smart-seq