Harvard-Led HuBMAP Atlas Is AI-Ready, but 501 Donors Fall Short

The NIH-funded Human BioMolecular Atlas Program (HuBMAP) Data Portal contains 9,316 datasets from 501 donors as of August, and the Harvard Medical School-led team that built it declared the resource "AI-ready" in a Nov. 2 paper in Nature Communications. The same paper warns that the donors do not fully reflect human demographic diversity — a caveat that matters directly for any lab planning to use HuBMAP as a training set or healthy reference.

Researchers at Harvard Medical School and five partner institutions — the Pittsburgh Supercomputing Center, the University of Pittsburgh, Carnegie Mellon University, the University of Pennsylvania and Stanford University — built the portal over eight years with more than 600 consortium members at more than 60 institutions. Harvard Medical School's Nov. 2 release calls it "one of the most comprehensive" atlases of healthy human tissue. Access is free and open, though sequencing data that could identify donors requires an approval process.

The portal's own data overview, last updated Oct. 29, already lists 9,727 datasets from 512 donors, and the paper says the consortium expects the volume to keep growing after publication.

What does 'AI-ready' mean here?

Morgan Turner, the paper's first author and director of scientific operations in the HIDIVE Lab at Harvard Medical School, said the consortium's data standards "make the data ready for AI." The paper states that computational biologists can use the portal to find "large uniformly processed datasets across multiple organs to train AI models or benchmark computational tools."

That readiness rests on engineering choices R&D managers can verify: each data type runs through its own standardized processing pipeline, each processed dataset lists the pipeline version that produced it, and every dataset traces back to its sample and donor. A 2024 review cited in the paper found HuBMAP was the only one of 28 biomedical databases to document everything on a 29-question FAIR checklist covering whether data is findable, accessible, interoperable and reusable.

How uneven is organ coverage?

Very. The measured distribution as of August:

• 3,188 of 9,316 datasets (34%) come from 29 donor placentas; GeoMx spatial profiling and probe-based RNA sequencing each produced 1,362 placenta datasets.
• Reproductive and pregnancy-related tissues — placenta, uterus, ovaries and fallopian tubes — account for about 4,480 datasets, nearly half the atlas.
• Uterus samples came from 87 donors, more than any other organ class; left and right kidneys count separately at 64 and 53 donors.
• Lungs are dense: 17 donor right lungs produced 800 datasets and six donor left lungs produced 641 — about 107 datasets per left lung.
• Hearts are thin: 37 donor hearts produced 191 datasets, about five per heart. Kidneys sit in between at 1,321 datasets from 117 donors, roughly 11 per kidney.
• Skin is thinnest per sample: 96 datasets from 774 samples.

The paper states plainly that organ-level spatial sampling is "uneven and biased by research questions," and that donor tissue availability and the priorities of NIH-funded projects limit the atlas's scope. Users treating HuBMAP as a healthy reference should expect findings may not generalize across demographic groups or tissue contexts, the authors caution.

Who is in the data?

Of the 501 donors counted in August, 284 were female, 209 male and eight of unknown sex, spanning ages 0 to over 90, with more than a third in their 20s and 30s. As of Nov. 4, the portal listed 284 of its 512 donors as White, 108 as Black or African American, 86 as unknown and 26 as Asian. The authors note "relatively low sampling from underrepresented groups."

The portal's own donor diversity page frames representation as central: to serve "as a reference atlas for future studies in human disease," HuBMAP needs donors spanning age, sex, race and ethnicity. On Nov. 4, that page — like the rest of the portal — carried a banner reading "This Repository Is Under Review For Potential Modification In Compliance With Administration Directives."

Why the gaps matter for AI models

Researchers use atlases like HuBMAP to train and benchmark single-cell and spatial foundation models and to define what counts as "normal" for anomaly and disease detection. Kuan-lin Huang, an associate professor at the Icahn School of Medicine at Mount Sinai, led a July audit in Cell Genomics of more than 13,500 samples across three other major single-cell atlases — the Human Cell Atlas, the Human Tumor Atlas Network and the PsychAD brain dataset. The audit found those atlases overrepresented people of European ancestry, and nearly 70% of Human Cell Atlas samples lacked any ancestry data.

"These gaps can be passed into AI models trained on the datasets," Huang said, adding that this often happens without users realizing it.

The consortium continues to publish data, so the composition of the atlas — and the bias any model inherits from it — will shift with each release.