Human BioMolecular Atlas Program (HuBMAP): 3D Human Reference Atlas construction and usage

Abstract

The Human BioMolecular Atlas Program (HuBMAP) aims to construct a 3D Human Reference Atlas (HRA) of the healthy adult body. Experts from 20+ consortia collaborate to develop a Common Coordinate Framework (CCF), knowledge graphs and tools that describe the multiscale structure of the human body (from organs and tissues down to cells, genes and biomarkers) and to use the HRA to characterize changes that occur with aging, disease and other perturbations. HRA v.2.0 covers 4,499 unique anatomical structures, 1,195 cell types and 2,089 biomarkers (such as genes, proteins and lipids) from 33 ASCT+B tables and 65 3D Reference Objects linked to ontologies. New experimental data can be mapped into the HRA using (1) cell type annotation tools (for example, Azimuth), (2) validated antibody panels or (3) by registering tissue data spatially. This paper describes HRA user stories, terminology, data formats, ontology validation, unified analysis workflows, user interfaces, instructional materials, application programming interfaces, flexible hybrid cloud infrastructure and previews atlas usage applications.

Fig. 1. Human Reference Atlas components and linkages.

a, The ASCT+B tables document the nested ‘part_of’ structure of organs (for example, cells that make up FTUs, successively larger anatomical structures, an entire organ such as the kidney, which is part_of the body). The cells that make up (are ‘located_in’) each of the anatomical structures are organized in a multilevel cell-type typology with ‘cell’ at the root and more and more specialized child nodes and ‘is_a’ relationships between nodes. The biomarkers used to characterize cell types might have one of five types: genes, proteins, metabolites, proteoforms and lipids organized in a biomarker typology. Gray arrows indicate crosswalks that connect other HRA DOs to ASCT+B tables. b, HRA 3D reference objects represent the shape, size, location and rotation of 1,192 3D anatomical structures with 516 unique Uberon IDs for 65 organs with crosswalks to ASCT+B tables. Shown are ‘renal papilla’ and ‘renal pyramid’ in the kidney. c, 2D reference illustrations document the shape, size and spatial layout of 3,742 2D cells of 116 types for 22 FTUs in ten organs with crosswalks to ASCT+B tables. Shown is the kidney nephron. d, Labeled training data exist for FTUs in five organs with crosswalks (gray arrows) to anatomical structures and cell types in the ASCT+B tables. e, 13 OMAPs are linked to 197 AVRs and there exist crosswalks to cell types and biomarkers in ASCT+B tables. f, Ten Azimuth references for healthy adult organs plus crosswalks to cell types and biomarkers in ASCT+B tables. g, HRApop reports cell type populations for anatomical structures compiled from experimental data. Exemplarily shown is the left atrium (blue) and the interventricular septum (orange) of the female heart plus a bar graph with the cell types that have the highest percentage across these two anatomical structures (annotated with Azimuth). Note that some cell types appear only in one anatomical structure. h, The HRAlit database links HRA DOs to existing ontologies (for example, Uberon and CL), expert ORCID, publication evidence, funding and experimental data used for HRApop computation.

Fig. 2. Mapping experimental data to the HRA.

a, A tissue block is 3D spatially registered and semantically annotated using the RUI or millitome (i). A smaller part of the tissue block might be used for sc/snRNA-seq analysis (not shown) or cut into tissue sections (ii). Tissue sections are analyzed using one or more assay types (iii). Shown are single-cell transcriptomics (for example, sc/snRNA-seq), OMAP-aligned spatial proteomics (for example, CODEX and Cell DIVE) and high-resolution hematoxylin and eosin (H&E)-stained histology images. Spatial alignment of different assay types for the very same or different tissue sections is non-trivial (iv). H&E data are used to segment FTUs using trained machine-learning models (v). A 3D reconstruction of tissue volumes is accomplished by aligning data from multiple serial tissue sections computationally (vi) followed by 3D segmentation and annotation (vii). The 2D or 3D data are analyzed to identify the distance of different cell types to the vasculature (VCCF visualizations) as a multiscale CCF from which no other cell is very distant (viii). b, Single-cell/nucleus data (sc/snRNA-seq) is stored as a cell-by-gene matrix; cell types are annotated using Azimuth or other cell type annotation tools; results are aggregated to cell-type-by-gene biomarker expression value matrices that are aligned with the ASCT+B tables; and are used in diverse HRA user interfaces (for example, EUI and FTU Explorer). c, OMAP-aligned spatial data generated using validated antibody panels linked to AVRs are analyzed to compute cell type by protein biomarker expression value matrices that are aligned with the ASCT+B tables using semi-automated workflows. d, The EUI provides full provenance for donors (sex, age and body mass index), data providers (upload date, contact name and affiliation), tissue blocks and sections (size, number, date and contact info for RUI registration) and datasets (assay type) with links to raw data in the HuBMAP Data Portal, other data portals or publications. e, CWL workflows detail which tools (yellow) are run on which input/output data (blue). Shown is the Azimuth cell type annotation workflow.

Fig. 3. Human Reference Atlas usage.

a, HRA can be used to compare the distribution of parenchymal cells including endothelial, epithelial and muscle that compose the blood vessels, airways and gas exchanging functional lung structures and resident immune cells including macrophages, to local vasculature (VCCF visualization) in healthy (top) and diseased (bottom) lung using multiplexed immunofluorescence microscopy images with bronchiole (br) and an accompanying small pulmonary artery (pa). Scale bars, white 5 mm; red 200 µm; and yellow 100 µm. The graphs on the right show distance distributions for cell types present in the healthy lung (top) and diseased BPD lung (bottom); the violin plot (middle) shows a comparison between distance distributions for cell types common in both datasets. Datasets are on GitHub. b, Multilevel cell neighborhoods can be computed to analyze and communicate the structure and function of FTUs; tissue image with cell type annotations and zoom into H&E with FTU segmentations (red outlines) and zoom into the multiplexed image (CODEX) is shown in left, neighborhoods are given in the middle; hierarchy of FTUs, neighborhoods, communities and cell types are shown on the right. Datasets are on GitHub. ICC, Interstitial cells of Cajal; TA, transit-amplifying; NK, natural killer; DC, dendritic cell; IEL, intraepithelial lymphocytes.