henhaobio
Sign in

bioRxivNeurosciencePreprint

A consensus atlas of human brain development defines cell type-specific maturation trajectories across the lifespan

Sridevi VenkatesaniD, Patricia R. NanoiD, Jonathan M. WerneriD, Sonia MalaiyaiD, Brian R. HerbiD, Yuan Gao, L. WangiD, Aparna BhaduriiD, Tomasz Jan NowakowskiiD, Carlo ColantuoniiD, Seth A. AmentiD, Jesse A GillisiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Human neurodevelopment is a continuous process that begins prenatally and extends into postnatal life. Current transcriptomic datasets are limited, fragmented across analytical frameworks, precluding comprehensive reconstruction of cellular trajectories linking developmental states to mature cell types. Here we present a consolidated cellular-resolution transcriptomic atlas of human brain development from the onset of neurogenesis to adulthood, covering [~]2.2 million cells from 156 donors across nine studies. All data were reprocessed from raw sequencing reads and annotated within a unified cell-type taxonomy, enabling reliable mapping across the lifespan. Cell types were highly replicable across heterogeneous datasets, enabling us to chart their maturation and cortical layer localization over time. We identify dynamic gene programs predictive of cell-type maturation, validate gene modules tracking known fate transitions, and leverage our atlas scale to characterize rare populations, including microglia. This resource establishes a standardized reference of human brain development and maturation gene modules for future comparisons across model systems, species, and disease states.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (17 Jul 2026) · CC BY 4.0

Reprocessing nine human brain datasets (2,157,582 cells, 156 donors, GW6 to adulthood) under one marker-based taxonomy yields 35 replicable cell types with cell type-specific timing of adult-like maturation.

Published single-cell atlases of human brain development sample different ages, mix cell and nucleus profiling and use separate pipelines, so they disagree on when cell types reach adult transcriptional states. Here every dataset is reprocessed from raw reads and labelled per cell with meta-analytic marker sets instead of integration and clustering, placing all cell types on one lifespan time axis.

Key results

  1. ShownPer-cell marker-enrichment annotation, without clustering or cross-dataset integration, yields 35 cell types (progenitors, newborn and mature neuronal subclasses, glia, vascular/immune) that replicate across studies, ages and cell vs nucleus technologies.

    Marker-set expression specificity: average AUROC 0.96 across 28 cell types (newborn types excluded). MetaNeighbor mean AUROC 0.78 overall; vascular/immune 0.96, glia 0.88, progenitors 0.8, GABAergic 0.8, glutamatergic 0.74, newborn glutamatergic/GABAergic 0.71/0.78, unassigned 0.53; stricter one-vs-next-best mean 0.61. · Uniform processing and annotation strategy to…; Replicable cell types show variable rates of…; Cell type annotation; MetaNeighbor

  2. ShownHuman types mature on distinct schedules: vascular/immune cells, L6 CT, L6b, L5 ET and Pax6/Vip/Sst/Pvalb interneurons prenatally; IT neurons except L6 IT, Lamp5, Chandelier cells and oligodendrocytes postnatally. Mouse types mature faster, most aligned to adult human states within 10 days after birth.

    Human segmented-fit inflections: L6 CT by GW15; L6b, L5 ET GW20-25; L6 IT near birth; other IT 3-8 months postnatal; Pax6 21 GW, Vip 28 GW, Sst/Pvalb GW 29-35; Chandelier, Lamp5, Lamp5 Lhx6 1-2 months postnatal; oligodendrocytes 3-4 months postnatal. Human-mouse (Allen developing visual cortex) mean AUROC 0.87 glutamatergic, 0.94 GABAergic; most mouse types aligned within 10 days of birth, L2/3 IT by P17. · Replicable cell types show variable rates of…; Discussion; MetaNeighbor

  3. ShownProjecting atlas labels onto age-matched developing-cortex MERFISH sections with Tangram reproduces inside-out lamination, even though prenatal subclass labels come from adult-derived markers, independently supporting the annotation.

    202,000 (GW14-16), 326,000 (GW20) and 63,000 (GW30-36) atlas cells projected to GW15, GW20-22 and GW34 sections. Hypergeometric enrichment in author-defined zones (Fisher-combined P): neural progenitors in vz and L6 CT/L6b in subplate at GW15; newborn glutamatergic in iz and L4 IT in l4 at GW20; L2/3 IT in l2 and Sncg/Pax6 in marginal zones at GW34. · Spatial transcriptomic validation reveals…; Spatial Projection; Discussion

  4. ShownPer-cell-type elastic net clocks on prenatal pseudobulks predict age; their positive-weight genes are cell type-specific and richer in autism risk genes, whereas negative-weight genes are shared proliferation and differentiation programs with stronger evolutionary constraint.

    27 prenatal models: median predicted-vs-actual correlation 0.78, median error 3 weeks; 13 (Lamp5) to 270 (L5 IT) genes per model. Positive genes enriched for own markers in 17/24 cell types; negative genes have zero marker overlap. Top 25% by importance: LOEUF mean 0.66 (negative) vs 0.79 (positive); SFARI overlap 24% (positive) vs 8% (negative). · Cell type-specific maturation genes reveal…; Genes important for cell type maturation are…; Cell type-specific maturation models

  5. ShownSubclustering 35,111 immune cells resolves nascent, intermediate and mature microglial states along a pseudotime trajectory; nascent microglia resemble yolk sac myeloid progenitors, and mature states are reached only in mid-childhood.

    MetaNeighbor AUROC of Mic-nascent vs a GW4-8 human yolk sac atlas: 0.97 with common myeloid progenitors, 0.92 with HSPCs. Generalized additive model of state proportions vs donor age: nascent decline, mature rise; about 10% of adult-donor microglia stay intermediate. Mic-intermediate splits into 15 subgroups and Mic-mature into 12, including a synapse-gene population peaking in early childhood. · Evaluation of immune cell trajectories in the…; Microglia trajectory analysis

Methods, data and reproducibility

Raw reads from nine single-cell/nucleus studies were reprocessed with Optimus and each cell labelled by its most enriched meta-analytic marker set (MetaMarkers). MetaNeighbor scored replicability and similarity to adult cells; Tangram projection onto MERFISH, per-cell-type elastic net age clocks, Monocle3 microglia pseudotime and EGAD-scored co-expression meta-modules followed.

  • Nine published single-cell and single-nucleus RNA-seq studies of human brain (cortical regions, forebrain, ganglionic eminences), first trimester to adulthood, reprocessed from FASTQ · human · n = 2,157,582 cells; 156 donors; 9 studies
  • MERFISH spatial transcriptomics of developing human cortex with author-defined cortical zones (GW15, GW20-22, GW34 samples) · human
  • Developing mouse visual cortex single-cell dataset (Allen Institute), combined with the atlas on 15638 one-to-one orthologs · mouse
  • Human yolk sac single-cell RNA-seq atlas (GW4-8) used to compare nascent microglia with hematopoietic progenitors · human
  • Adult cross-region subclass differential expression statistics (Jorstad et al., Supplementary Table 4) used to build adult marker sets
  • NMF pattern loadings for cortical development and adult cortical layers (Sonthalia et al. 2026)
  • Neurodevelopmental risk gene sets: SFARI genes (including score 1), ASD and NDD genes from Fu et al. 2022; LOEUF constraint scores · n = 185 ASD genes; 664 NDD genes
Methods & tools
Optimus (WARP), scDblFinder, MetaMarkers, CONCORD, MetaNeighbor (pyMN), Tangram, glmnet (caret), Seurat, Monocle3, MrTree, projectR, EGAD
Reproducibility gaps
CellxGene release is described as future ('will be available'); the brain-development browser has no URL in the text; some FASTQs were obtained directly from authors rather than an archive; no software versions are given.
Limitations (4)
  • Inter-regional differences were not analysed in detail because cross-study comparisons were underpowered, with each region sampled in at most two studies. · stated by the authors
  • Higher-resolution clusters and subtypes were not resolved; the analysis stays at predefined subclasses chosen for replicability across datasets and conservation with mouse. · stated by the authors
  • Adult subclass markers were refined using DE consistency in three of the atlas's own postnatal datasets (Herring, Zhu, Velmeshev), so the marker-specificity AUROC of 0.96 is not fully independent of the annotation inputs. · noted in this summary
  • Age clocks are trained on prenatal pseudobulks only, and the text does not say whether the reported accuracy (median correlation 0.78) is measured on held-out donors or studies. · noted in this summary

Read this if you need a uniformly processed human brain development reference (GW6 to adulthood) for timing cell type maturation, benchmarking organoid or mouse data, or mining age-predictive and co-expression gene modules.

Citations
1

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISZonca '26Venkatesan '25Polioudakis '18Braun '23Jayakumar '24Eze '21
74% semantic matchNeuroscience
Asia Zonca, Erik Bot, José Dávila-Velderrain·Cell Reports·May 28, 2026