By Simone Webb, Antony Rose, Chuan Xu, Lloyd Steele, Masami A Kuri, Emily Stephenson, Kemal Inecik, Daniyal Jafree, April R Foster, Daniela Basurto-Lozada, Nana-Jane Chipampe, Anna V Pournara, Marc-Antoine Jacques, Ken To, Chloe Admane, Efpraxia Kritikaki, Marta A Chroscik, David Horsfall, Julia Foreman, Koen Rademaker, Julia Karjalainen, Anna Laddach, Shaista Madad, John E G Lawrence, Vitalii Kleshchevnikov, Steven Lisgo, Jimmy Tsz Hang Lee, Joséphine Blevinal, Ahlam Alqahtani, Stanislaw Makarchuk, Jake Jackson, Ekin Ucuncu, Teresa P Silva, Valentina Lorenzi, Fereshteh Torabi, Rachel A Botting, Kenny Roberts, Bayanne Olabi, Keerthi P Chakala, Leander Dony, Gabriele Dall’Aglio, Ana-Maria Cujba, Holly J Whitfield, Christine Hale, Justin Engelbert, Tong Li, Elena Prigmore, Nusayhah Hudaa Gopee, Michael W Mather, Ben Talks, Lijiang Fei, Diana Pereira, Corina Moldovan, Rasa Elmentaite, Krzysztof Polanski, Alexander V Predeus, Pasha Mazin, Vicky Rowe, Semih Bayraktar, Vincent R Knight-Schrijver, Menna Clatworthy, Kazumasa Kanemaru, James Cranley, Andrew Guo, Sanjay Sinha, David McDonald, Andrew Filby, Janet Kerwin, Peng He, Inês Sequeira, Roser Vento-Tormo, Malte D. Luecken, Alain Chédotal, Deanne Taylor, Silvia Martin-Almedina, Pia Ostergaard, Sam Behjati, Elizabeth Robertson, Nicola K Wilson, John Marioni, Helen V Firth, Caroline F Wright, Olivier Pourquie, Laure Gambardella, Paula Alexandre, Berthold Gottgens, Kerstin B Meyer, MS Vijayabaskar, Mo Lotfollahi, Sarah A Teichmann, Fabian J Theis, Fränze Progatzky, Omer Ali Bayraktar, Muzlifah HaniffaSourcebioRxiv
Fork
Clone
git clone https://openversions.com/git/index/biorxiv-10-64898-2026-03-30-714220v2.git
Cite
Version link
https://openversions.com/index/biorxiv-10-64898-2026-03-30-714220v2/commit/2108d842b489986c445643bc296bc61020ed0d19
Version identifier (commit hash)
2108d842b489986c445643bc296bc61020ed0d19

Fork to create your own copy with credit to the source. Clone to get a Git copy on your computer.

Latest version

Checks complete

Archive from bioRxiv Source: https://www.biorxiv.org/content/10.64898/2026.03.30.714220v2 License: https://creativecommons.org/licenses/by/4.0/ Manuscript unchanged.

Committed

Claims

Findings

Headline finding

The Human Developmental Cell Atlas (HDCA) integrates published and unpublished sc/snRNA-seq datasets across 21 prenatal organs with newly generated single-cell, multiome, and spatial transcriptomics data of 6PCW intact human embryos, spanning 4-22 post-conceptional weeks (~4.6 million cells/nuclei, 454 refined cell types) and charting 114 tissue niches.

Benchmarking atlas integration with scIB and scTRAM demonstrates that scVI achieves the highest combined score (0.494) compared to scPoli (0.437), TarDis (0.421), PCA, Harmony, and Scanorama, providing the strongest balance between trajectory fidelity and batch integration.

Unsupervised NicheCompass analysis of 8 human embryo Xenium sections (3.1 million cells) identifies 114 tissue niches—including 37 focal anatomical structures (5 novel niches not distinguishable by H&E), 30 CNS niches, and 47 distributed niches—and reference-query mapping of 3 additional embryo sections identifies 90 niches without requiring expert histological annotations.

Joint integration of ~3.6 million predominantly mesenchymal cells from scRNA-seq and Xenium reveals 10 previously undefined early fibroblast progenitor populations broadly expressing FRZBFRZB and/or HIC1HIC1 with distinct regional patterning, alongside 3 lateral plate mesoderm (LPM) subtypes, 4 umbilical populations, and 2 extraembryonic mesenchymal populations.

Mapping published human iPS-derived skin organoid scRNA-seq data to the HDCA recapitulates the newly defined ZIChiZIC^{hi} and RUNX2+RUNX2^+ mesenchymal progenitor populations and resolves fine cell subtypes within the off-target cartilage population.

Analysis of 94,681 endothelial cells (ECs) from the HDCA and 163,973 ECs from 6PCW Xenium data resolves 37 EC types and demonstrates that organ-specific capillary transcriptional programs in the brain, heart, and liver are already established as early as 6PCW.

Subclustering of 3,932 lymphatic endothelial cells (LECs) reveals 7 LEC subsets, including early LECs (FOXO1+CALCRL+FOXO1^+ CALCRL^+) enriched at 6-10PCW and detectable in the human developing cranium as early as 6PCW, and shows that congenital/early-onset primary lymphatic anomaly (PLA) genes are enriched at embryonic stages whereas late-onset PLA genes are enriched at fetal stages.

RNAscope in 5-6PCW human embryos validates PIEZO1PIEZO1 expression in cells co-expressing the pan-LEC markers PROX1PROX1 and FLT4/VEGFR3FLT4/VEGFR3 within and immediately adjacent to the cardinal vein.

Temporal and spatial reconstruction of 159,763 peripheral nervous system (PNS) cells shows that sensory and autonomic NCC/SCP populations co-exist at 6PCW and universally co-express SCP markers (MPZMPZ, PLP1PLP1), that the human trigeminal ganglion consists purely of NCC-derived sensory neurons, and that mesenchymal and vascular niches form an early scaffold guiding axonal/SCP migration and neuronal maturation via spatially organised ligand-receptor signalling.

The bespoke Cherita web portal integrates the HDCA with the Developmental Disorders Gene2Phenotype (DDG2P) dataset, enabling gene-cell and gene-niche queries across 21 organs and whole-embryo spatial data.

Methods

Community sc/snRNA-seq datasets were compiled from 21 datasets, realigned to GRCh38 where necessary, and harmonised to HGNC gene symbols (Ensembl release 107) using the IDTrack algorithm.

Source excerpts

“To resolve assay- and release-specific disparities in gene nomenclature for external datasets, we employed the pre-release IDTrack algorithm181 (see https://github.com/theislab/idtrack), a graph-based identifier harmonization framework designed to reconcile discrepancies in gene nomenclature across large-scale transcriptomic datasets.” · Show in paper

Batch-aware highly variable gene (HVG) selection was performed using scanpy.pp.highly_variable_genes across 192 batches defined by donor identity, biological unit, sample status, and sequencing platform, retaining the top 12,288 HVGs for scVI integration.

Source excerpts

“The batch_key was defined by the Cartesian product of 4 covariates —donor identity, biological unit, sample status, and sequencing platform, resulting in 192 batches.” · Show in paper

“To integrate the entire public and embryo data after per-system annotation harmonisation as above, we used the above-mentioned batch_key to identify 12,288 HVGs. Highly variable genes were selected in a batch-aware manner using scanpy.pp.highly_variable_genes, with dispersion normalized within each batch and aggregated globally to mitigate batch-driven effects.” · Show in paper

Atlas integration methods (scVI, scPoli, TarDis, Harmony, Scanorama, PCA) were benchmarked using a dual framework combining scIB metrics and scTRAM (single cell trajectory representation metrics) across three curated ground-truth lineage graphs.

Source excerpts

“To systematically evaluate how effectively different integration methods and model configurations preserve both batch correction and developmental biology in the multi-organ human developmental cell atlas, we implemented a dual-framework benchmarking strategy combining the established scIB44 protocol with scTRAM45, a trajectory-fidelity metric suite specifically designed to assess preservation of developmental progressions in integrated representations.” · Show in paper

Whole-embryo scRNA-seq (6PCW, n=3) was performed by dissecting each embryo into 12 anatomical regions, digesting with type IV collagenase and trypsin, FACS-sorting live DAPI-negative cells, and sequencing 10x Genomics 5' v1/v2 libraries on Illumina NovaSeq S4.

Source excerpts

“Each of the three whole human embryo samples for scRNAseq/multiome were dissected into twelve anatomical regions using a scalpel (Supplementary Table 2; Fig. S1a-e).” · Show in paper

Whole-embryo 10x Genomics Multiome (snRNA-seq + snATAC-seq) was performed on frozen cells from F137, F147, and F158, aligned with cellranger-arc v2.0.2, and analysed using Scanpy, SnapATAC2, MACS3, and cell2fate.

Source excerpts

“Multiome data (sn-RNA-seq and sn-ATAC-seq) were aligned to GRCh38-2020-A reference using cellranger-arc v2.0.2 pipeline with the default parameters from 10X CellRanger suite.” · Show in paper

Spatial transcriptomics was performed on 6PCW FFPE embryo sections using 10x Genomics Visium CytAssist (n=2 sections) and Xenium Prime 5000-plex (n=11 sections across SOB25 and SOB26), with cell label transfer performed using scANVI.

Source excerpts

“These slides were then taken forward for Visium analysis by destaining and decrosslinking the stained sections according to the standard 10X CytAssist workflow (CG000520) then proceeding immediately to perform the Visium CytAssist Spatial Gene Expression according to the standard 10X CytAssist workflow (CG000495).” · Show in paper

“We then used SCANVI (number of layers: 1; number of latent dimensions: 10) using labelled scRNA-seq and unlabelled Xenium data, restricted to ~5000 genes found in both datasets.” · Show in paper

Unsupervised niche identification was performed on 8 reference Xenium sections using NicheCompass (v0.3.0) with Squidpy 8-nearest-neighbour spatial graphs, followed by fine-tuning on 3 query sections.

Source excerpts

“We ran NicheCompass51 (v0.3.0), a graph neural network-based approach, to obtain niche embeddings. We split whole embryo sections into 8 reference and 3 query sections (Supplementary Table 5), to illustrate how our reference can be used to obtain niche embeddings for future datasets.” · Show in paper

Cell-cell communication in the DRG was inferred on Xenium data using LIANA+ (v1.5.0) rank_aggregate, and gene set enrichment scores were computed using decoupleR (v2.1.1) Univariate Linear Model (ULM).

Source excerpts

“For the cross-system analysis within the Dorsal Root Ganglion (DRG), cell–cell communication analysis was performed using LIANA+187,188 v1.5.0 on the spatially resolved Xenium dataset restricted to DRG cells found in Fig. S24d.” · Show in paper

“Gene set enrichment scores for the PNS system were computed using the Univariate Linear Model (ULM) method implemented in the decoupleR189 Python package (v2.1.1) in Fig. 5m, and Fig. S23f, S23j, S24e.” · Show in paper