Hakken narrows 1.5 million hypotheses to three lab tests; two found supporting data
A Hakken preprint reports that two of three AI-selected aging-gene hypotheses showed supporting gene-expression data in cell culture; the three were drawn from 1.5 million scored candidates.
1,543,297 hypotheses. Three tested in cell culture. Two found supporting evidence.
On 3 September, the authors published a preprint (not yet peer-reviewed) on the Hakken model. Hakken builds a relationship map from biomedical literature, then scores hypotheses about relationships absent from the authors' working database. One caveat: "absent from the database" means absent from the specific working set of facts the authors extracted from publications.
Experts selected 1,386 aging-related genes. Among 959,805 pairs of those genes, the model identified 1,543,297 hypotheses with scores above its threshold. After removing known relationships and applying automated selection, 2,804 candidates remained. Biologists picked three hypotheses with model scores above 0.8 and sent them to an independent contract laboratory.
In these experiments, "relationship" means functional regulation of expression: an intervention in a cell culture changes the amount of messenger RNA (mRNA, the copy of a gene that cells use to build a protein) or of the protein itself.
Two of the three hypotheses received supporting data. For the TP53-BAMBI pair, Nutlin-3 raised the level of p53 protein (encoded by TP53). BAMBI mRNA rose 16% at 24 hours and 38% at 48 hours. For RAF1-TNF, the inhibitor GW5074 produced a small, reproducible increase in TNF mRNA at 24 hours, though the amount of TNF protein released by the cells into the culture medium stayed below the limit of quantification. The third hypothesis, SOAT1-STAT3, found no support: stimulating SOAT1 left STAT3 mRNA and protein levels unchanged.
Three tests out of 1.5 million candidates do not validate the model's overall accuracy. A gene-expression change in cell culture does not establish that lifespan will increase. The preprint has not undergone peer review.
Sources
[1] arxiv.org
The Hakken preprint gives a numbered, concrete illustration of AI hypothesis generation meeting laboratory testing, and the question of who selects and funds such verification is the natural angle for surfacing the Experiment organization page on Eternal Search.