Single-cell atlases used to map the human body and train artificial intelligence models may not represent global populations fairly, according to a study led by researchers at the Icahn School of Medicine at Mount Sinai on July 20. The analysis found that individuals of European ancestry were consistently overrepresented, while Asian and Latino individuals were underrepresented, and a significant portion of samples lacked any record of ancestry.
The findings, published in Cell Genomics, are based on a review of more than 13,500 samples from three major resources: the Human Cell Atlas, the Human Tumor Atlas Network, and the PsychAD Consortium. Researchers curated reported ancestry, race and ethnicity, and sex for each sample before comparing these datasets against global population statistics as well as disease-specific reference data.
The study found consistent gaps across all three resources. In the Human Cell Atlas specifically, nearly 70 percent of samples had no recorded ancestry information. Among those with available data, European ancestry was overrepresented by about sixfold compared with global expectations; Asian, African, and Latino groups were underrepresented. Tumor samples in the Human Tumor Atlas Network were approximately 69 percent European; brain dataset samples from PsychAD were nearly two-thirds European. Some cancer types also showed sex differences beyond what would be predicted by disease incidence.
Researchers said that even when accounting for missing data using conservative assumptions, overrepresentation of Europeans remained clear. Dr. Huang said, "The most surprising finding was how much ancestry information was simply missing from some of these costly studies." Huang added, "These gaps can be passed into AI models trained on the datasets... which is why building diversity and complete demographic information into single-cell studies from the start matters so much." The authors warned that if research or AI tools are built on unrepresentative datasets, "the benefits of those advances may not reach everyone equally," echoing concerns seen previously in human genomics research.
This study is among the first systematic assessments examining representation within major single-cell atlases rather than just their size or detail level. The team provided a practical checklist for future studies covering recruitment planning, demographic recording practices, sample balancing strategies, and reporting standards for AI model performance across groups.
Dr. Huang concluded, "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... we wanted to help make sure the field can improve them so they serve more people fairly." The authors noted limitations including reliance on publicly reported demographic categories rather than direct genetic measurement; they plan further audits to track improvements over time.
