AWS genomics proof of concept finds disease-linked patterns in Epstein-Barr virus genomes

An AWS-backed proof of concept used an AI model to find distinct patterns across more than 1,300 Epstein-Barr virus genomes, including clusters associated with clinical conditions. The results are promising research signals, not a validated diagnostic or evidence that the patterns cause disease.

AWS AI Watch analysis

What happened

A collaboration between a national genomics research institute and Temus used the Evo2 genomics foundation model to turn viral genome sequences into numerical embeddings, then applied clustering methods to look for patterns. The work used AWS infrastructure, including GPU-powered EC2 instances, to handle the data and compute demands.

AWS says the analysis found clusters with 69–97 per cent purity for disease-type labels. It also reports that the average assignment probability for the dominant disease type exceeded 0.85 in key clusters. The team describes this as a proof of concept, with further validation still needed.

Why it matters

Finding patterns in large, complex genome datasets can help researchers decide what to investigate next. Here, the model surfaced groups associated with conditions including nasopharyngeal carcinoma and Burkitt lymphoma, offering candidate signals for follow-up research rather than clinical conclusions.

The useful distinction is between finding a pattern and proving what it means biologically. The researchers say they still need to reproduce the analysis in a larger, independent cohort and cross-check it with other genomic AI models. Genomes may yield clues; they do not label themselves for our convenience.

Our read

This is a concrete example of AI helping researchers sift through genomic data at scale, with reported results specific enough to make the next research steps clear. But the work remains an early proof of concept. Treat its clusters as leads for further investigation, not as a basis for diagnosis or claims about individual risk.

What to watch

• Whether the clusters reproduce in a larger, independent sample.
• Whether other genomic foundation models find similar patterns.
• Whether biological validation supports the links between clusters and clinical conditions.

Discussion spark: What should count as enough independent validation before AI-found genomic clusters are treated as useful evidence in disease research?

not affiliated with or endorsed by Amazon Web Services (AWS)