HuBMAP: Human BioMolecular Atlas Program

• HuBMAP is a multi-institutional NIH initiative creating a detailed 3D atlas of healthy human tissues at single-cell resolution.
• It integrates spatial transcriptomics, proteomics, metabolomics, and imaging data to systematically catalog cellular and molecular phenotypes.
• Standardized protocols, advanced computational pipelines, and open access ensure reproducibility and interoperability with related global atlasing projects.

The Human BioMolecular Atlas Program (HuBMAP) is a multi-institutional, NIH Common Fund initiative launched with the aim of constructing a comprehensive, three-dimensional, multi-modal reference atlas of the healthy human body at single-cell resolution. HuBMAP integrates spatially resolved transcriptomic, proteomic, metabolomic, imaging, and other high-content molecular datasets from diverse adult donors to systematically catalog cellular and molecular phenotypes in their precise anatomical context. The program emphasizes openness, reproducibility, and data harmonization, and is tightly connected to related consortium efforts, including the Human Cell Atlas, Human Protein Atlas, and organ-specific mapping projects (Snyder et al., 2019, Hemberg et al., 13 Aug 2024).

1. Objectives, Vision, and Consortium Structure

HuBMAP's mission is the production of foundational 3D tissue maps covering both sexes, multiple ancestries, and a broad adult age range, capturing both cellular identities and the molecular states—genomic, epigenomic, transcriptomic, proteomic, metabolomic—of cells in situ. The goal extends to delineating spatially explicit architectures of cell types, subtypes, and states; capturing their interactions and microenvironments; and providing baselines for comparative analyses in disease, development, and regenerative medicine (Snyder et al., 2019, Weber et al., 2019, Hemberg et al., 13 Aug 2024).

The project is organized into distributed Tissue Mapping Centers (TMCs) responsible for standardized tissue procurement and assay; Transformative Technology Development (TTD) and Rapid Technology Implementation (RTI) groups for methodological innovations; and a Data Coordination Center (DCC, also called HIVE) for data harmonization, integration, visualization, and public dissemination (Turner et al., 7 Nov 2025, Hemberg et al., 13 Aug 2024).

HuBMAP is explicitly collaborative, working to align protocols, ontologies, and reference frameworks with parallel efforts and the biomedical research community, with policies for continuous, FAIR-compliant data releases and programmatic data access (Snyder et al., 2019).

2. Foundational Technologies and Methodological Integration

HuBMAP supports and scales a suite of orthogonal measurement modalities:

• Single-cell 'omics: Massively parallel dissociation-based assays (scRNA-seq, ATAC-seq, SNARE-seq2) with per-cell mRNA quantification sensitivity down to ~1 transcript/cell for highly expressed genes, sequencing depths >50,000 reads/cell, and typical detection of >2,000 genes/cell, spanning thousands of individuals (Snyder et al., 2019, Turner et al., 7 Nov 2025).
• Multiplexed spatial transcriptomics/proteomics: High-plex platforms (seqFISH, MERFISH for RNA; CODEX, Immuno-SABER for proteins) enabling direct detection of up to 1,000 analytes per section at sub-micron (200–500 nm) resolution. Lumiphore-coupled fluorescence extends multiplexing via lanthanide emission differentiation. Calibration is performed with reference beads and dilution series, and quantitative signals are mapped across serial tissue sections (Snyder et al., 2019).
• Imaging mass spectrometry/cytometry: MALDI and nano-DESI IMS (spatial metabolomics, proteomics; 5–10 µm resolution) and imaging mass cytometry (IMC, ~30 protein markers, ~1 µm resolution) allow multi-modal mapping of cellular environments, with protocols for quantitative normalization and false discovery control (Snyder et al., 2019).
• Image analysis and deep learning: Automated segmentation of functional tissue units (FTUs) and microvasculature is performed by state-of-the-art architectures (U-Net, U-Net++, Detectron2, DeepLabV2, FPN), with domain adaptation and semi-supervised learning explicitly mitigating cross-protocol domain gaps (Sydorskyi et al., 2023, Sultan et al., 2023, Keller et al., 22 Oct 2025).

Each technology cycle generates quantitative QA/QC metrics (e.g., Dice coefficient, IoU for segmentation; spatial resolution Δx; molecular sensitivity) and is fully detailed in standardized, publicly accessible protocols.

3. Spatial Mapping and the Common Coordinate Framework (CCF)

HuBMAP mandates the use of a Common Coordinate Framework—a multi-layer reference system providing semantic ("what"), spatial ("where"), and clinical ("who") annotation for each tissue sample and corresponding data product (Börner et al., 2020, Weber et al., 2019).

• CCF Clinical Ontology encodes specimen provenance (e.g., donor age, sex, BMI), acquisition metadata, and experimental history with OWL/JSON-LD expressivity and W3C PROV interoperability.
• CCF Semantic Ontology provides a partonomy over anatomical structures (AS), associated cell types (CT), and biomarker/feature mappings (B), enforcing tree-like "part_of" hierarchies and explicit assignment of ASCT+B triplets (Börner et al., 2020).
• CCF Spatial Ontology defines 3D Cartesian reference spaces (origin, axes, scaling, Euler-angle-based rotations, translations), linking specimen-local coordinates to organ- and body-level reference frames. Rigid and non-rigid registration algorithms map experimental images and spatial molecular signals into CCF-aligned volumes [

$\mathbf{x}' = \mathbf{R} \mathbf{x} + \mathbf{t}$

where $\mathbf{R}$ is the $3 \times 3$ rotation matrix and $\mathbf{t}$ is a translation vector], enabling direct spatial querying and inter-donor cross-comparison.

A key innovation is the proposal and partial implementation of a vasculature-based CCF, where vascular architectures are modeled as rooted trees (G = (V, E, w)) with hierarchical (loop, branching, arc length, circumferential) coordinates that facilitate multi-scale (organ, FTU, cell) registration and pan-and-zoom-style exploration (Weber et al., 2019).

4. Data Acquisition, Curation, and Portal Infrastructure

As of October 2025, the HuBMAP Data Portal (https://portal.hubmapconsortium.org/) supports access to 5,032 datasets (primary raw and processed), covering 27 organ classes and 310 adult donors, with a mean of ~8.2 samples per donor organ. Modalities include single-cell and nucleus RNA-seq (N=1,469), ATAC-seq (N=1,023), SNARE-seq2, CyTOF, histological WSIs, spatial transcriptomics and proteomics (CODEX, MIBI, Visium, 2D/3D IMS, light-sheet) (Turner et al., 7 Nov 2025).

Data curation and processing are performed via standardized, reproducible pipelines (CWL, Docker, Airflow, AnnData/OME-TIFF output), with provenance fully tracked (Donor→Sample→Assay→Dataset→Processing). Quality control is enforced through both automated schema validation (CEDAR) and manual review of key metrics, ensuring biological differences are not confounded by pipeline inconsistencies (Turner et al., 7 Nov 2025, Hemberg et al., 13 Aug 2024).

Data are indexed (Elasticsearch, UBKG), accessible via open APIs, searchable via both faceted metadata and data-driven queries (e.g., by gene, protein, or cell type), and linked to interactive visualization engines (Vitessce, JupyterLab Workspaces, EUI) supporting direct in-browser analysis of >1,500 datasets (Turner et al., 7 Nov 2025, Keller et al., 22 Oct 2025).

Community contributions are enabled via EPICs (Externally Processed Integrated Collections), which allow external labs to contribute results (cell-type annotations, segmentations, analyses) alongside official pipeline outputs.

5. Computational and Analytical Innovations

HuBMAP drives development and integration of advanced computational methods to support high-throughput, multi-modal analysis and interpretation:

• Segmentation pipelines: Domain-adapted, semi-supervised segmentation architectures combining CNNs (EfficientNet, ResNet), vision transformers (MiT-B3/B5), and ensemble strategies have achieved near state-of-the-art Dice coefficients for FTU segmentation (e.g., kidney Dice up to 0.96, lung up to 0.49) across HPA and HuBMAP datasets (Sydorskyi et al., 2023). Focal loss and feature pyramid networks provide improvements in class imbalance and multi-scale feature extraction (Sultan et al., 2023).
• Probabilistic atlases: Bayesian frameworks for probabilistic cell identification, supporting joint modeling of cell positions, morphology, and molecular features, offer scalable assignment mechanisms for integrating new data and managing uncertainty in cell annotation (Bubnis et al., 2019).
• Generative models: Conditional GANs with U-Net generators, guided by SSIM-based channel clustering, have demonstrated synthesis of missing multiplexed proteomics channels at >2.5× the scale of prior work (29–100 channels), with image-level SSIM up to 0.97 for DAPI and 0.85–0.92 for major surface markers (Saurav et al., 2022).
• High-content multiplex imaging: The open IBEX protocol supports >65-plex iterative imaging using commodity hardware, chemical bleaching (LiBH₄), and automated affine registration, with SNR >10–100:1 and spatial resolution down to 0.16 µm XY, providing the basis for reproducible, spatially resolved phenotyping across multiple tissue types (Radtke et al., 2021).

These computational tools are released under open-source licenses, and their outputs are natively integrated into HuBMAP's polymorphic data standards (e.g., OME-TIFF, AnnData, MuData, SpatialData) for interoperability.

6. Applications, Benchmarking, and Data Exploration

Interactive exploration and analysis are central requirements. HuBMAP's web-based tools (Vitessce, EUI) deliver facility for viewing, querying, and annotating large-scale spatial data—including per-channel rendering of FTU segmentations, multi-modality overlays, and direct access to quantitated pathomics features (morphology, texture, color, distance transforms; e.g., area, circularity, aspect ratio, Haralick statistics) (Keller et al., 22 Oct 2025). Standardized file conventions explicitly link image, segmentation mask, and feature tables, facilitating both research and clinical applications.

HuBMAP's platforms meet or exceed all 29 criteria in biomedical FAIRness assessments. The data portal supports rapid collaborative discovery, bulk data transfer (Globus endpoints), live Jupyter workspaces, and explicit API-driven integration with other major cell atlas consortia (Turner et al., 7 Nov 2025).

Quantitatively, the segmentation and machine-learning pipelines achieve expert pathologist-level accuracy on specific FTU types (Dice >0.8 for glomeruli, tubules), and the generative imaging models approach photo-realistic image synthesis benchmarks (SSIM ≥0.97 for critical markers) (Sydorskyi et al., 2023, Saurav et al., 2022, Keller et al., 22 Oct 2025).

7. Challenges, Limitations, and Future Directions

Major anticipated challenges include tissue heterogeneity, assay sensitivity for low-abundance analytes, and the management of petabyte-scale data. Ongoing work includes iterative SOP refinements, adaptive power analyses (for rare cell state sampling), and scalable cloud storage/computation. The spatial resolution of CCF reference organs is currently constrained by baseline imaging (∼1 mm, Visible Human), with efforts underway to enhance this via micro-CT, MRI consensus organs, and further automation (Snyder et al., 2019, Börner et al., 2020).

Batch correction, scaling of integration algorithms, and metadata completeness remain active areas, with ongoing development of streaming, out-of-core analytics, and dynamic, user-driven harmonization frameworks for massive single-cell and spatial-omics datasets (Hemberg et al., 13 Aug 2024).

Future phases of HuBMAP will expand into pediatric and disease-context atlases, consider translational and AI-driven applications, and further formalize governance for clinical and ethical issues arising from synthetic imaging and atlas-driven decision support (Snyder et al., 2019, Saurav et al., 2022).

Overall, HuBMAP stands as a foundational resource, defining both infrastructure and standards for spatially explicit, harmonized, multi-modal human tissue atlasing, with direct implications for tissue engineering, precision therapeutics, and computational biology at scale.