HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses
Jieyuan Liu, Mengzhou Hu, Jefferson Chen, JungHo Kong, Pratibha Jagannatha, Yiming Gao, Dexter Pratt, Hsin-Yuan Lee, et al.
cs.CL · cs.CE · cs.MA · cs.NE
Quick verdict
A generational genetic algorithm coordinating LLM agents raises drug-repurposing hypothesis scores over six baselines, but the strongest-baseline gap is unresolved on one metric and a constant-drug control wins held-out Open Targets.
Summary
HypoEvolve casts scientific hypothesis discovery as a finite population search. A generation agent initializes a population of mu = 6 literature-grounded hypotheses; each of G = 3 generations samples parents by size-2 fitness tournaments, produces lambda = 6 offspring through LLM crossover (probability 0.6, combination or inspiration variants) and LLM mutation (probability 0.15, drug substitution or out-of-box), scores parents and offspring jointly with pairwise LLM comparisons, fits a Bradley-Terry model to obtain fitness, and truncates back to the top mu. Agents are gpt-5.4-mini with Tavily retrieval. Evaluation is drug repurposing across 34 cancer types against two held-out external measures: DepMap CRISPR selectivity and Open Targets association, both applied only after the search and never used as fitness. HypoEvolve reaches 0.171 DepMap selectivity and 0.426 Open Targets association, versus 0.115 and 0.329 for Tree of Thoughts, the strongest of six baselines, and 0.039 and 0.163 for single-pass generation. Gains hold on 27 held-out cancer types not used in protocol development. Fitness-guided parent selection improves population mean and minimum external scores with all scientific operators fixed. The paper's honest negatives matter: the Open Targets margin over Tree of Thoughts is not statistically resolved (p = 0.126), and a constant-drug control selected on development types scores 0.4827 on held-out Open Targets against HypoEvolve's 0.4290.
Review complete
Assessment at a glance
4
risk flags identified
● high 1
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Formulating hypothesis development as a generational genetic population search makes agent collaboration an explicit, testable design variable separate from the agents' scientific roles.
- HypoEvolve achieves the highest mean DepMap selectivity (0.171) and Open Targets association (0.426) among seven methods across 34 cancer types, versus 0.115 and 0.329 for the strongest baseline, Tree of Thoughts.
- Gains over single-pass generation generalize to 27 held-out cancer types excluded from protocol development and interim inspection (+0.2798 Open Targets, p = 8.2e-7; +0.1114 DepMap, p = 2.8e-4).
- With scientific operators and hypothesis count held fixed, fitness-guided parent selection raises population mean and minimum external scores, establishing a contribution of search design itself.
- Crossover and mutation produce 87 of the 94 final hypotheses, and drug-cancer matching residuals improve from generation 0 to generation 3 in 21 of 31 and 25 of 34 cancer types.
Strengths
Collaboration is isolated as an experimental variable
Section 4.4 replaces fitness-guided tournament selection with uniform random selection while holding generation, crossover, mutation, retrieval, evaluation and hypothesis count fixed. That is the cleanest test in the paper: population minimum rises by +0.0879 on DepMap (p = 6.6e-4) and +0.2181 on Open Targets (p = 1.2e-6), and the mean by +0.0753 and +0.1283, attributing the effect to search design rather than to agent capability.
External evidence is genuinely held out
DepMap selectivity and Open Targets association are computed only after each method has committed to one hypothesis, never enter search fitness, and each method contributes one externally unselected answer. The paper also refuses to take an externally chosen maximum over its own candidate pool, which blocks the common selection-on-the-test-set leak.
Raw dependency scores were diagnosed and corrected
Lineage and provenance are recorded for every offspring
Negative results are reported rather than buried
Weaknesses
A constant-drug control beats HypoEvolve on held-out Open Targets
The headline comparison to the strongest baseline is unresolved on one of two metrics
Fitness is model self-assessment with no validation against external evidence
Ablations are underpowered and openly non-significant
Cost and budget comparability are limited
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks a question that agent-for-science work usually skips: given fixed LLM agents with fixed scientific roles, does the *collaboration rule* - how proposals are combined, revised and retained - change hypothesis quality? HypoEvolve answers by making collaboration a generational genetic algorithm over a population of mu = 6 hypotheses for G = 3 generations, with lambda = 6 offspring per generation, and by holding prompts, roles, retrieval and evaluation fixed while varying search decisions. The instantiation is drug repurposing across 34 cancer types, and the assessment is deliberately external: DepMap CRISPR selectivity and Open Targets association, computed only after each method commits to a single answer and never entering search fitness. That design is the paper's best asset.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is HypoEvolve and how does it differ from other LLM scientist agents?
HypoEvolve is a framework that coordinates specialized LLM agents through a generational genetic algorithm over a fixed-size population of hypotheses. Agents handle generation, pairwise scientific comparison and semantic variation (crossover and mutation), while the genetic algorithm decides which hypotheses reproduce and which survive. The point is that the collaboration rules are explicit, so they can be changed while the agents' scientific roles, prompts and criteria stay fixed - making collaboration itself an experimental variable rather than an implementation detail.
How were the generated hypotheses evaluated?
Through drug repurposing across 34 cancer types. Each hypothesis proposes one of 61 approved drugs and a mechanism, which implies that the drug's targets matter for that cancer. That implication is tested against two external sources applied only after the search finishes: DepMap CRISPR knockout screens (measured as selectivity relative to each target's pan-cancer median dependency) and Open Targets drug-target-disease association scores. HypoEvolve reaches 0.171 DepMap selectivity and 0.426 Open Targets association versus 0.115 and 0.329 for the strongest baseline.
Are the hypotheses experimentally validated?
No. The evaluation measures whether the proposed drug's annotated targets have external genetic, literature or clinical support for the named cancer. The authors state explicitly that prospective experiments are needed to establish the full mechanism and therapeutic benefit, and that expert review remains necessary before any proposal informs therapeutic decisions.
What are the main limitations of the study?
Computational cost: pairwise comparison accounts for 167 of 206 model calls per run and grows quadratically with pool size. Fitness depends on LLM judgment rather than expert or experimental evidence. Most ablations cover only eight cancer types with one seed each and none survives multiple-testing correction. And the margin over the strongest baseline, Tree of Thoughts, is not statistically resolved on Open Targets association (p = 0.126).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-8 could not be inspected; all numbers were read from the body text, table captions and appendix tables, and Figure 3's per-baseline values were reconstructed from the axis labels in the extracted text.
- The 61-drug vocabulary is truncated to 25 entries in Appendix D.4, so the exact task constraint is not fully specified.
- No code artifact is released and the backbone model is a commercial API with live web retrieval, so results were not re-run.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.