Hakken preprint reports laboratory tests of three hypotheses selected from 1,5 million biomedical hypotheses; two yielded supporting evidence

Hakken preprint reports laboratory tests of three hypotheses selected from 1,5 million biomedical hypotheses; two yielded supporting evidence

On 3 September, the authors published a preprint on the Hakken model, a manuscript that has yet to undergo peer review. The model looks for relationships between biomedical entities that are absent from the authors’ working database of facts extracted from publications. Biologists selected three of these hypotheses for laboratory testing.

Hakken builds a map of relationships from publications: each record connects two biomedical entities, specifies their relationship, and records when that relationship appeared in the literature. The model compares this structure with article passages that mention the entities separately and scores hypotheses absent from the authors’ working database. For each candidate, Hakken provides a short chain of established facts to explain its high model score.

For an evaluation focused on aging, experts selected 1 386 aging-related genes. Across 959 805 pairs, Hakken identified 1 543 297 hypotheses with scores above a specified threshold. Relationships already recorded in the database were removed, and an automated selection of the three highest-scoring hypotheses for each entity left 2 804 candidates. Biologists selected three hypotheses with model scores above 0,8 and submitted them to an independent contract laboratory for testing.

In these experiments, a “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 the cell uses to make a protein, or the amount of the protein itself. For the TP53–BAMBI gene pair, Nutlin-3 increased the level of p53 protein, which is encoded by TP53. BAMBI mRNA increased by 16% at 24 hours and by 38% at 48 hours. For the RAF1–TNF pair, GW5074, a compound that inhibits RAF1 activity, produced a small, reproducible increase in TNF mRNA at 24 hours. The amount of TNF released by the cells into the culture medium remained below the limit of quantification.

The third hypothesis, SOAT1–STAT3, received no support under the experimental conditions used: stimulation of SOAT1 left STAT3 mRNA and protein levels unchanged.