IBM Research Exposes How Biased DNA Data Corrupts Biomedical AI Models
A study of 4,514 omics papers finds under 3% report ancestry, threatening biomedical foundation models with cascading population bias.
Takeaways
- Audit of 4,514 omics papers shows only 2.7% report ancestry, 3.1% report sex or gender.
- European-ancestry samples make up 78% of GWAS participants but 16% of the global population.
- CellxGene and GEO, key training sources for single-cell models, are heavily skewed demographically.
- Warfarin dosing algorithms trained on European data underperform for patients of African descent.
- Automated pipeline validated with 97.8% specificity but only 12.5% recall, so numbers are conservative.
- Authors propose a Provenance, Openness, and Evaluation Transparency framework modeled on CONSORT trial reporting.
Biomedical foundation models are inheriting a bias problem that starts long before a single weight is trained. A new perspective paper from researchers at IBM Research, Hebrew University, Tel Aviv University, and the Barcelona Supercomputing Center audits the demographic transparency of the omics literature and finds it essentially nonexistent, with the datasets feeding today's largest biological models skewing heavily toward European-ancestry samples.
The team scanned 4,514 PubMed-indexed omics publications from 2015 to 2024, examining reporting across multiple demographic dimensions. The results are stark, and they matter because the same corpora underlie models like scGPT, scBERT, DNABERT, ESM-3, and AlphaFold 3.
The transparency gap, by the numbers
Across genomics, transcriptomics, proteomics, epigenomics, and microbiome research, demographic metadata almost never surfaces in abstracts. Only 2.7% of publications report ancestry or ethnicity information, 4.5% report geographic origin, and 3.1% report sex or gender composition. Rates vary considerably by field: genomics leads on ancestry reporting at 6.0%, while transcriptomics (0.6%) and proteomics (1.0%) rarely mention it.
Even when metadata exists, the underlying populations are lopsided. Individuals of European ancestry account for over 78% of participants in genome-wide association studies, despite representing only 16% of the global population. The most widely used single-cell repository has the same problem: CellxGene, the largest known source of human single-cell RNA-seq data to date, remains skewed toward samples from adults of European or unknown ethnicity. A parallel audit of the Gene Expression Omnibus found significant bias in ethnicity representation, with a predominance of White and African American participants.