A 13-author research team has introduced HypoEvolve, a multi-agent system that treats scientific hypothesis formation as a population search. In a drug-repurposing test spanning 34 cancer types, it produced hypotheses with the highest mean support on two external biological databases among six baselines. On the panels shared by every method, its DepMap score was 0.171 versus 0.115 for Tree of Thoughts, while its Open Targets score was 0.426 versus 0.329. The result tests whether team structure can improve what identical language models propose.
Drug repurposing asks whether an existing medicine could work in a different disease. A fluent rationale is insufficient evidence, so the researchers converted each proposed drug into its annotated target genes and scored those targets after final selection. DepMap measures whether the genes are unusually important to survival in cell lines from the named cancer relative to other cancers. Open Targets aggregates genetic, clinical, literature and other evidence connecting a target to a disease. Neither score shows that the drug will treat patients.
For each cancer, every method used gpt-5.4-mini, the same 61-drug vocabulary and the same literature-retrieval setup. HypoEvolve began with six hypotheses. Agents compared every pair for cancer specificity, supporting evidence and testability; a tournament then chose parents. Agents combined or revised rationales to produce six offspring. Parents and offspring were scored together, the top six remained, and the cycle ran for three generations. DepMap and Open Targets were withheld from this search and used only to score the final choice.
The study ran HypoEvolve two or three times per cancer, yielding 94 runs, and compared it with six baselines including single-pass generation, debate and Tree of Thoughts. Across 26 cancers with complete DepMap coverage, HypoEvolve led single-pass generation in 19 and trailed in seven. In 27 held-out cancer types, it also exceeded single-pass generation by 0.111 on DepMap for the 24 covered types and 0.280 on Open Targets. The configuration was fixed before held-out evaluation.
| Method | DepMap (26 cancers) | Open Targets (29 cancers) |
|---|---|---|
| HypoEvolve | 0.171 | 0.426 |
| Tree of Thoughts | 0.115 | 0.329 |
| Single pass | 0.039 | 0.163 |
An ablation separates the contribution of search policy from agent capability. When the researchers replaced fitness-guided parent selection with random selection while keeping prompts, agents, retrieval and hypothesis count fixed, guided selection raised the average population score by 0.075 on DepMap and 0.128 on Open Targets. The best score in each population did not change significantly, and the DepMap margin for the final selected hypothesis was not statistically conclusive. The system also used 206 model calls per run, 167 for pairwise comparisons.
These findings support a narrower claim than autonomous drug discovery: explicit rules for selecting, varying and retaining ideas improved database support in this benchmark. The paper is an unreviewed preprint, the hypotheses had no wet-lab or patient testing, and models may have encountered some drug-disease associations during training. The Open Targets advantage over Tree of Thoughts was not statistically resolved, and a fixed-drug control outscored HypoEvolve on that database's mean. Independent, blinded laboratory tests must now show whether the better-ranked hypotheses produce more useful experiments.
Sources
- Jieyuan Liu and colleagues, “HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses,” arXiv, September 14, 2026. Abstract, authorship, submission date and headline result.
- Full paper. Search procedure, comparison conditions, external metrics, ablations, statistical tests, costs and limitations. The comparison table was created by Global Tech News from Figure 3 values. Lead photograph: Dr. Jason Kang / National Cancer Institute, public domain; the image is illustrative and does not show the reported study.
- Cancer Dependency Map. Official portal for cancer dependency data and CRISPR screening resources used by the study.
- Open Targets Platform. Official description of the target-disease association resource used for the second external measure.