Demographic Biases Threaten the Fairness of Global Single-Cell Human Atlases

The single-cell "atlases" that are increasingly used to map the human body and as training data for artificial intelligence models, may not represent the world's populations fairly, according to a study led by researchers at the Icahn School of Medicine at Mount Sinai. The analysis found that people of European ancestry were consistently overrepresented, while Asian and Latino individuals were underrepresented, and that a large share of samples lacked any record of ancestry at all. The findings were published July 20 in Cell Genomics.

Single-cell technologies allow scientists to profile biology one cell at a time, revealing rare cell types and disease-related changes that older, bulk methods miss. Large international efforts have used these tools to build reference maps, expansive datasets meant to serve as universal references for research and medicine. But the researchers say the demographic makeup of these resources had not been systematically examined, raising the question of whether the maps reflect all of humanity or only part of it.

The team reviewed more than 13,500 samples from three major single-cell resources: the Human Cell Atlas, the Human Tumor Atlas Network, and the PsychAD Consortium. They curated the reported ancestry, race and ethnicity, and sex of the sampled individuals, then compared each dataset against global population data, U.S. cancer incidence data, and disease-specific reference data, examining patterns across tissues, cancer types, and brain disease categories.

"These atlases are becoming the reference maps for biology and medicine, and they are increasingly used to train the AI models that will shape future research and care," says senior corresponding author Kuan-lin Huang, PhD, Associate Professor of Genetics and Genomic Sciences, and Artificial Intelligence and Human Health, at the Icahn School of Medicine.

We wanted to ask a simple but important question: do these maps represent people from different populations fairly, or are some groups being left out? What we found is that some of the most important datasets are not as representative as they should be."

Kuan-lin Huang, Associate Professor of Genetics and Genomic Sciences, and Artificial Intelligence and Human Health, The Icahn School of Medicine at Mount Sinai

The gaps were consistent across all three resources. In the Human Cell Atlas, nearly 70 percent of samples had no ancestry information recorded at all. Among the samples that did, European ancestry individuals were overrepresented roughly sixfold compared with global expectations, while Asian, African, and Latino individuals were underrepresented. In the study, "Latino" refers to ethnicity reported in the available data, rather than a single genetic ancestry, and its underrepresentation highlights the need for single-cell datasets to better reflect the communities they aim to serve.

Tumor samples in the Human Tumor Atlas Network were about 69 percent European, and the PsychAD brain dataset was nearly two-thirds European. Several cancer types also showed differences by sex that went beyond what disease incidence would predict. 

The researchers note that the European overrepresentation in the Human Cell Atlas held up even under the most conservative assumptions about the missing data, meaning it cannot be explained away by incomplete records alone.

"The most surprising finding was how much ancestry information was simply missing from some of these costly studies" says Dr. Huang. "These gaps can be passed into AI models trained on the datasets, often without users of the AI models realizing it, which is why building diversity and complete demographic information into single-cell studies from the start matters so much."

The concern, the authors say, is one of fairness and reliability. If future research, AI tools, biomarkers, or treatments are built on datasets that do not represent everyone, the benefits of those advances may not reach everyone equally. The problem echoes a long-standing issue in human genomics, where studies conducted mostly in people of European ancestry have produced risk scores that transfer poorly to other populations.

The study is among the first systematic checks of who is represented in major single-cell atlases. Rather than only asking how large or detailed the datasets are, the researchers asked whether they are fair and generalizable across populations. The team also provides a practical, field-specific checklist that research groups can use when designing future studies. The checklist covers how to plan recruitment, record demographic information, balance samples, and report whether AI models perform equally well across ancestry and sex groups.

"A key takeaway is that representation matters. If these are going to be reference maps of human biology, they should reflect the diversity of humanity," says Dr. Huang. "These resources are already very valuable, and we wanted to help make sure the field can improve them so they serve more people fairly, which ultimately means more reliable science and more equitable precision medicine."

The analysis was carried out by a team of student researchers, co-first authors Catrina Yang of the University of Oxford, Kavitharini Saravanan of the University of North Carolina at Charlotte, and Aryan Saharan of Saint Louis University, working with Dr. Huang at the Icahn School of Medicine.

The authors acknowledge that the study relied on the demographic information reported in public datasets rather than direct measurement of genetic ancestry, and that broad ancestry categories cannot capture the full complexity of identity. Because the work focused on three major public consortia, the findings may not reflect every single-cell dataset in the field. As a next step, the team plans to expand this type of audit to additional datasets, track whether representation improves over time, and evaluate whether AI models trained on single-cell data perform differently across populations.