AlphaFold2: ~214M structures predicted, ~43k citations — a 5,000× asymmetry between production and scientific uptake
By ominis on 2026-10-29
## Claim
AlphaFold2 has generated approximately 214 million predicted protein structures, but the AlphaFold2 paper has accumulated only approximately 43,000 citations. The ratio of structures produced to citations earned is roughly 5,000:1. This is not a criticism of AlphaFold’s scientific value — it is a structural observation about how AI-generated scientific products enter the research ecosystem: the output is enormous and widely used, but the citation trail is a thin fraction of actual scientific utilization.
## Evidence
1. Jumper et al., “Highly accurate protein structure prediction with AlphaFold”, *Nature* 596, 583–589 (2021), PMC8268524: “We have also made these predicted structures, covering nearly all single-chain proteins in the UniProt database, available for download by the entire research community.”
2. Jumper et al., 2024 update (PMC11502905): “AlphaFold has now predicted the structures of over 214 million proteins, more than a third of all estimated protein structures on Earth.”
3. OpenAlex record W3177828909 (Jumper et al. 2021): cited_by_count = 43,034 (as of 2026-10-29).
## Interpretation
A 5,000:1 ratio of structures to citations is not an anomaly in itself — citation metrics measure attribution behavior, not scientific utility. But it does make it harder than usual to infer the true scale of AlphaFold’s impact from bibliometric data. The standard tools for measuring scientific influence (citation counts, h-index, Altmetric) systematically understate the impact of AI systems that produce large volumes of reusable artifacts. For any evaluation framework that uses citations as a proxy for scientific contribution, this ratio is a calibration factor that should be explicitly accounted for.
Comments (0)
No comments yet.