Every diagnosis starts with the same question: what does normal look like? For most of medical history, doctors answered from experience and textbooks. On November 2, researchers gave them something stronger: a free, open HuBMAP Data Portal holding more than 9,000 datasets from healthy human tissue, built so scientists and AI systems can measure disease against a shared reference. The work was described in Nature Communications by Turner and colleagues.
This human tissue atlas changes three things at once. For a future patient, it means the tests behind a diagnosis rest on better baselines. For a researcher, it means ready-made data instead of building a collection from scratch. For science itself, it means one common baseline instead of scattered lab collections. No single lab could have built anything at this scale. That is the story here: scientific data made widely available.
This atlas does not treat or cure any disease. It is infrastructure, not a medicine: think of it as a survey map of the healthy body. Its value is what others can build on it.
Why your future diagnosis starts with healthy tissue
To spot what is wrong, you need a clear picture of what is right. A pathologist reading a biopsy, a researcher hunting early cancer signals, and an AI model flagging abnormal cells all need the same thing: a reliable definition of healthy.
That definition has been patchy. Tissue samples lived in separate labs, processed in different ways, and hard to compare. The new portal tackles that problem directly. It gathers two kinds of data from healthy donors across dozens of organs: spatial data, which shows which molecules sit where, and single-cell data, which goes down to the level of individual cells. Researchers can search by organ, gene, protein, or cell type, and view results in a web browser.
This continues a project Intelligent Living has followed since 2020, when we covered the Human Cell Atlas effort to map every cell in the human body. Think of that earlier vision as the ambition. This new release is a working reference library from a different consortium, with a live portal researchers can open now.
What the atlas actually contains
The program is called HuBMAP, short for the Human BioMolecular Atlas Program. It is funded through the US National Institutes of Health Common Fund and started in 2018. More than 600 members at more than 60 institutions contributed over eight years.
As reported in the paper, the portal held 9,316 public datasets as of August 2026, spanning 26 data types and 29 organ classes from 501 donors. The live counter on the site has since moved higher, listing more than 9,700 datasets from more than 510 donors. Use the paper figures as the peer-reviewed baseline and the portal counter as the current total.
- Donors: 284 female, 209 male, and 8 of unknown sex, ages 0 to 90. Race data skew White, with Black or African American donors the next largest group, plus Asian donors and a large unknown category.
- Organs: placenta, lung, kidney, uterus, small and large intestine lead by dataset count. Uterus came from 87 donors, the most sampled organ by donor count.
- Data types: single-cell and single-nucleus RNA sequencing (reading which genes are active), chromatin accessibility (ATACseq, which parts of DNA are open for reading), protein imaging (MIBI, CODEX, Cell DIVE), spatial transcriptomics (Visium, Xenium, GeoMx, seqFISH, Slide-seq), mass spectrometry imaging (MALDI, DESI), light sheet microscopy, and liquid chromatography mass spectrometry, among others.
- Access: free under a Creative Commons Attribution 4.0 license. Raw identifiable sequencing data sits behind a controlled approval process.
Three things anyone can do in the portal today
The portal is built for use, not just storage. A Pittsburgh Supercomputing Center announcement describes it as a data repository and 3D map for studying the body from organs down to cells. A Filter and Browse interface is the simplest front door. Here are three practical entry points.
1. Look up a gene or cell type across healthy organs. The Biomarker and Cell Type Search indexes gene expression, chromatin accessibility, and cell type prevalence. A translational scientist, one who moves lab findings toward the clinic, can check what normal looks like before studying a tumor. A molecular biologist can see which cells in which organs express a gene of interest.
2. Explore tissue visually without downloading anything. Interactive Vitessce visualizations cover more than half of all datasets across 19 data types, including 2D and 3D spatial views. Organ pages add cell population plots for heart, kidney, and lung. As with rare disease work that depends on good reference data, the ability to inspect before downloading saves months.
3. Analyze data next to the data. Workspaces provide browser-based Jupyter notebooks in Python or R with free computing from PSC, plus guided templates. Researchers can also query through an API, pull bulk files through Globus, or request controlled sequencing data through dbGaP. A natural-language helper called Say and See lets you ask questions in plain English instead of building filters one by one.
What AI-ready actually means
Press coverage calls this atlas AI-ready. That phrase deserves a plain translation.
AI-ready here means standardized, versioned, and findable. Each data type runs through its own uniform processing pipeline, and each processed dataset lists the pipeline version so results stay comparable over time. Metadata follows shared formats. Every dataset traces back to its sample and donor through provenance records, a documented chain from sample to result. A cited 2024 review found HuBMAP the only repository of 28 surveyed to satisfy all 29 FAIR criteria. FAIR is shorthand for findable, accessible, interoperable, and reusable.
That standardization is what lets AI researchers train and benchmark single-cell and spatial foundation models, meaning large general models adapted to many tasks, and lets others define normal for anomaly detection. It does not mean the AI answers are automatically fair or correct. Models learn the gaps in their training data. Which brings us to the honest part.
What is missing: who is underrepresented and why one organ dominates
The paper warns plainly that its donor pool does not fully reflect human diversity. Sampling reflects which tissues were available and which NIH-funded projects prioritized them, not a population sampling plan. The authors state that findings may not generalize across demographic groups or tissue contexts. Independent reporting on the release, including R and D World analysis of the limitations, stresses the same point for AI training sets.
The clearest skew is organ coverage. About 34 percent of datasets, 3,188 of 9,316, came from just 29 placentas, driven by two high-throughput assays, methods that process many samples at once. Reproductive and pregnancy-related tissues account for nearly half the atlas. By contrast, 37 hearts yielded 191 datasets, while skin yielded 96 datasets from 774 samples. The paper calls spatial sampling uneven and biased by research questions.
None of this makes the resource useless. It makes the use instructions matter: treat it as the best open healthy reference available, check who and what is thinly sampled before generalizing, and expect the collection to keep growing as the consortium publishes more data. What would strengthen future releases:
- Broader donor recruitment across ages, ancestries, and geographies.
- Balanced organ coverage beyond placenta and heavily studied tissues.
- Continued pipeline versioning so older and newer datasets stay comparable.
A practical note: the portal currently displays a banner stating the repository is under review for potential modification in compliance with administration directives. Availability was confirmed at the time of writing.
How HuBMAP differs from the other atlases
Names overlap and confuse search results. This table separates four resources readers often mix up.
| Resource | What it is | How it differs |
|---|---|---|
| HuBMAP Data Portal | NIH program portal for healthy human tissue, spatial plus single-cell, with browser tools | The subject of this article: open reference with standardized pipelines and analysis workspaces |
| Human Cell Atlas | International effort to map every human cell type | Broader global ambition, covered on this site in 2020; HuBMAP complements it with deep spatial data and a live portal |
| Human Protein Atlas | Protein-focused maps, widely used for protein location | Protein-centered, not the multimodal healthy-tissue reference described here |
| GTEx (Genotype-Tissue Expression) | Bulk-tissue gene expression resource | Bulk averages without the spatial and single-cell resolution HuBMAP adds |
Frequently Asked Questions
Is this human tissue atlas free to use?
Yes. Published datasets carry a Creative Commons Attribution 4.0 license. Tools for search, visualization, and analysis run in the browser, with bulk download through Globus. Only raw sequencing data that could identify donors requires an approval process.
How many datasets and donors does it cover?
The peer-reviewed paper reports 9,316 public datasets from 501 donors as of August 2026. The live portal counter now shows higher totals as data releases continue. Always cite which date a number belongs to.
Does the atlas cure disease?
No. It is a reference for healthy tissue. Researchers and AI developers use it to compare diseased samples against normal, to train models, and to run cross-organ analyses. Any medical benefit comes later, through studies and tools built on top of it.
What are its biggest limits?
Demographic representation and organ balance. Underrepresented groups are thinly sampled, and placenta and a few well-studied organs dominate dataset counts. The authors advise caution when generalizing.
A shared baseline for the next decade of medicine
Most medical progress looks like a new drug or device. This looks like plumbing: shared pipes for data, standard formats, and tools anyone can open. Its strength is openness: thousands of standardized healthy samples available to any researcher with a browser.
The measure of success will not be headlines about the atlas itself. It will be quieter: a study that found its control group in hours instead of years, a model trained on comparable data instead of a patchwork, a diagnosis that arrived sooner because normal was finally well defined.
