Source-linked AI summary
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, Zhiting Hu, Trey Ideker, Wei Wang, Eric P. Xing, Zhen Wang
TL;DR
HypoEvolve addresses the open question of how agent collaboration affects scientific hypothesis quality by separating scientific roles from the rules that update a hypothesis population. It uses a generational genetic algorithm with specialized agents and evaluates drug-repurposing explanations against external biological evidence. Across cancer types, it achieves the highest scores among six baselines on both DepMap selectivity and Open Targets association, with gains over single-pass generation generalizing to held-out types.
Problem
It remains unclear how agent collaboration affects hypothesis quality independently of agents’ scientific capabilities, motivating explicit and controlled search rules.
Method
HypoEvolve coordinates specialized LLM agents with a generational genetic algorithm that selects, varies, scores, and replaces structured scientific hypotheses.
Results
Across cancer types, HypoEvolve achieves the highest mean scores among six baselines on both DepMap selectivity and Open Targets association.
Takeaways & Limitations
The framework makes collaboration a testable search-design variable and supports study of scientific teams beyond individual models.
Takeaways & Limitations
Pairwise comparison accounts for 167 of 206 model calls and grows quadratically with the evaluated pool, while token budgets are not matched across methods.
Abstract
from arXiv · showhide
Scientific agents contribute to hypothesis discovery by synthesizing evidence, assessing proposals, and developing new explanations. Recent systems combine scientific agents with evolutionary search through critique, comparison, and revision. However, how different forms of agent collaboration affect hypothesis quality remains an open question. Answering this question requires separating the effects of agents' scientific capabilities from those of their collaboration. A framework must therefore preserve agents' scientific roles and support rules for combining, revising, and retaining hypotheses. Building on this view, we introduce HypoEvolve, which makes collaboration explicit through successive updates to a hypothesis population. Specifically, we propose a generational genetic algorithm to coordinate specialized large language model (LLM) agents that integrate mechanistic arguments, reconsider assumptions, and assess evidence and testability. Each generation specifies how scientific judgments and new proposals reshape the population, making collaboration effects on hypothesis quality directly testable. Moreover, we design our evaluation around scientifically meaningful hypotheses that explain how a proposed intervention could work. Drug repurposing links these explanations to target-level biological claims assessed against external evidence. Specifically, we adapt DepMap and Open Targets into complementary external measures grounded in experimental, genetic, and clinical evidence. Across 34 cancer types, HypoEvolve achieves the highest scores against six baselines on both measures. DepMap selectivity reaches 0.171, versus 0.115 for the strongest baseline. Gains over single-pass generation also generalize to held-out cancer types. HypoEvolve advances a vision of autonomous science in which AI research teams achieve a capacity for discovery beyond that of individual models.
1 Introduction
HypoEvolve makes agent collaboration an explicit, controllable part of evolutionary hypothesis search while preserving specialized scientific roles. It evaluates this framework through drug-repurposing hypotheses and external biological evidence across cancer types.
- 1 Introduction: HypoEvolve separates scientific roles from search rules, allowing collaboration to vary while prompts, criteria, and agent capabilities remain fixed.Each population update determines which proposals agents receive and which outputs enter the next round.
- 1 Introduction: A generational genetic algorithm coordinates specialized agents that evaluate, vary, and replace hypotheses across successive population updates.Crossover combines mechanistic arguments, mutation reconsiders assumptions, and parent-offspring competition retains promising proposals.
- 1 Introduction: The evaluation links drug-repurposing explanations to target-level claims assessed with DepMap selectivity and Open Targets association across 34 cancer types.Both external measures are reserved for assessment after the search under a shared task and retrieval protocol.
- 1 Introduction: HypoEvolve achieves the highest mean scores among six baselines on both external biological measures.DepMap selectivity reaches 0.171, while the supplied introduction passage reports the strongest baseline at 0.115.
2 Related Work
Prior work combines scientific-agent workflows with evolutionary search, but HypoEvolve distinguishes itself by making hypothesis collaboration an explicit, fixed-size generational process.
- 2 Related Work: Scientific hypothesis systems use specialized agents for generation, criticism, synthesis, reviewing, prioritization, and experimental feedback.Examples include literature-grounded proposal refinement, chemistry and materials systems, drug-combination hypotheses, and experiment-linked discovery.
- 2 Related Work: HypoEvolve differs from nearby scientific systems by using fixed-size generational search with explicit parent selection, semantic variation, joint replacement, and recorded lineages.This design makes collaboration rules directly inspectable and controllable.
- 2 Related Work: Evolutionary search has been applied to prompts, programs, equations, molecules, and diverse text artifacts through evaluation, selection, crossover, and mutation.These approaches treat language-model outputs as population members that can be iteratively improved or diversified.
3 Method
HypoEvolve formulates scientific hypothesis discovery as a finite population search in which specialized agents generate, compare, and evolve structured explanations. A genetic algorithm updates the population by selecting parents, producing offspring, scoring the combined pool, and retaining the highest-fitness candidates.
- Problem Formulation: Each candidate is a structured document containing a title, summary, hypothesis statement, and supporting rationale.The search targets scientifically grounded explanations addressing a natural-language research goal.
- Problem Formulation: The search maintains µ candidates, produces λ offspring per generation, runs for G generations, and returns the highest-fitness final hypothesis.Fitness is inferred from task-specific comparisons within each generation.
- Generational Search with Comparative Fitness: The update Qt = Pt−1 ⊎Ot, ft = Score(Qt), Pt = Topµ(Qt; ft) pools parents and offspring, scores them, and retains the top µ candidates.Distinct candidate identities are preserved, and lineage records trace parentage and operator choices.
- Algorithm Overview: A generation agent initializes literature-grounded hypotheses, while comparison and evolution agents operate on their claims and rationales.Three specialized agents implement generation, comparative judgment, and semantic variation.
- Comparative Scientific Judgment: Pairwise comparison evaluates specificity, target evidence, and falsifiable predictions, with Bradley-Terry strengths converted into search fitness.Fitness directs parent selection and population replacement, while external biological evidence is applied only after search.
- Semantic Variation: Crossover combines or redirects mechanisms from two parents, whereas mutation revises an intervention or reconsiders an explanation’s assumptions.The supervisor repeats these operations each generation and records the resulting parentage and operator provenance.
4 Experiments
HypoEvolve is evaluated on drug-repurposing hypotheses across cancer types using external biological evidence, held-out testing, evolutionary trajectories, and ablations. It achieves the strongest overall benchmark performance, while evolution improves fitness and drug–cancer matching, with mixed sensitivity to search settings and limited gains from additional calls or scaffold feedback.
- 4.2 Main Results: 0.171 DepMap selectivity and 0.426 Open Targets association are HypoEvolve’s highest means, exceeding Tree of Thoughts at 0.115 and 0.329.The comparison covers 26 shared cancer types for DepMap and 29 for Open Targets.
- 4.2 Main Results: +0.280 on Open Targets and +0.111 on DepMap over single-pass generation show gains on 27 held-out cancer types.The held-out analysis excludes seven cancer types used for protocol development or interim inspection and applies the frozen configuration.
- 4.2 Main Results: +0.334 on genetic association is HypoEvolve’s largest Open Targets margin over single-pass generation, exceeding margins of +0.247 for known drugs and clinical trials and +0.313 for literature.The genetic-evidence comparison records 25 wins and 4 losses.
- 4.3 Hypothesis Evolution: 87 of 94 final hypotheses arise through crossover or mutation, while crossover produces 60 and mutation 27.Seven final outputs are unchanged parent copies; across all offspring records, inspiration, out-of-box mutation, combination, and drug substitution account for 27.6%, 24.8%, 22.9%, and 17.1%.
- 4.4 Ablations and Computational Cost: Search-setting changes and operator removal produce mixed metric shifts without comparisons surviving multiple-testing correction, while scaffold feedback changes scores by −0.017 and −0.010.Repeating the withheld-feedback condition changes scores by +0.012 and +0.028; HypoEvolve uses 206 model calls per run, including 167 pairwise comparisons.
- 4.4 Ablations and Computational Cost: Fitness-guided selection raises weakest-member scores by 0.088 and 0.218 and population means by 0.075 and 0.128 on DepMap and Open Targets, respectively.Maximum external scores show no statistically detectable change.
5 Conclusion
HypoEvolve treats scientific hypothesis development as explicit population search, separating agent reasoning from generational coordination. Its drug-repurposing study demonstrates a reusable architecture while motivating broader evaluation across tasks and models.
- HypoEvolve couples reasoning over scientific claims and rationales with genetic selection, variation, and replacement across hypothesis populations.The framework assigns agent contributions defined roles and makes collaboration rules available for controlled study.
- Fitness-guided parent selection improves mean and minimum scores on both external measures while scientific operations and hypothesis count remain fixed.This isolates a contribution of search design to the biological support of generated hypotheses.
- The drug-repurposing implementation uses domain-specific retrieval, fitness criteria, semantic operators, and external evidence within the reusable population-search architecture.The same genetic control flow is proposed for domains such as single-cell data, cell-type annotation, and materials simulation.
A.2 Limitations and Future Directions
The paper identifies computational cost, limited operator-study coverage, and risks from unsupported assumptions as important boundaries. It proposes more efficient comparisons, broader replication, adaptive search, and expert or experimental validation.
- Pairwise comparison accounts for 167 of 206 model calls and grows quadratically with the evaluated pool, making computational cost the main practical constraint.Sparse comparison schedules could redirect budget toward larger populations or longer search horizons.
- The operator study covers eight cancer types, so broader replicated analyses are needed to clarify the separate contributions of crossover and mutation.Adaptive operator rates and diversity-aware replacement are suggested as ways to allocate search effort.
- LLM-based selection may propagate unsupported assumptions across generations, directing attention toward persuasive hypotheses with weak evidence.Expert review of supporting literature and independent experiments remain necessary before therapeutic use.
- HypoEvolve’s generational procedure uses tournament selection, crossover with pc = 0.6, mutation with pm = 0.15, and joint parent-offspring replacement.The algorithm records offspring identity, parents, and operators during population updates.
C.1 Statistical Details for Main Results
Statistical analyses compare HypoEvolve with baselines, assess evidence channels and configurations, and test scaffold feedback. Results support gains over single-pass generation, while configuration and feedback effects are mixed or unresolved.
- Comparisons with single-pass generation are significant across known drugs and clinical trials, literature, and genetic association evidence channels.The reported paired p values are 1.4 × 10^-5, 1.3 × 10^-8, and 9.8 × 10^-6, respectively.
- Table 3 reports mean-score margins, cancer-type wins and losses excluding ties, and paired Wilcoxon p values for main and held-out comparisons.Its caption states that DepMap comparisons favor HypoEvolve, while the Open Targets difference from Tree of Thoughts remains statistically unresolved.
- No search configuration differs significantly from the default on either metric in the eight-cancer-type configuration study.Increasing population size yields four wins and no losses on Open Targets, while four or five types retain the same drug across settings.
- Scaffold feedback yields 3 wins and 10 losses on DepMap and 6 wins and 8 losses on Open Targets versus withholding feedback, with neither comparison significant.Approximately 60% of cancer types receive the same drug under both conditions.
C.2 Fitness Improves Across All Cancer Types
Fitness improves across all 34 evaluated cancer types under the generational search. Across 94 runs, both mean and best fitness increase consistently from the common generation-0 baseline.
- 124.2% mean fitness improvement occurs in 94 of 94 runs, while best fitness improves 54.9% in 94 of 94 runs.The mean and best-fitness changes are reported with paired t-test p values of 1.7 × 10^-49 and 1.3 × 10^-32.
- Mean fitness increases in all 34 cancer types from generation 0 to generation 3.Figure 6 measures percentage changes relative to the common initial fitness scale.
C.3 DepMap Selectivity Improves in Most Cancer Types
HypoEvolve outperforms single-pass generation on DepMap selectivity across most evaluated cancer types, extending its aggregate advantage across cancer contexts.
- 19 of 26 cancer types favor HypoEvolve over single-pass generation on DepMap selectivity.Single-pass generation leads in the remaining 7 cancer types.
- Figure 7 compares per-cancer selectivity scores for HypoEvolve and single-pass generation.Positive selectivity indicates greater target dependency in the matched cancer than in the pan-cancer reference.
C.4 Evolution Improves Drug-Cancer Matching
The cancer-specificity analysis measures drug–cancer matching after removing each drug’s average performance across other cancers. HypoEvolve improves within-generation matching and generally compares favorably with multi-candidate baselines and controls.
- The residual removes a drug’s average advantage across other cancer contexts, isolating cancer-specific matching.The analysis covers 31 DepMap types and 34 Open Targets types; the main comparison uses the smaller common set across methods.
- +0.0612 DepMap selectivity and +0.0658 Open Targets association are the generation-3 residual means.Residuals increase from generation 0 to generation 3 in 21 of 31 DepMap cancer types and 25 of 34 Open Targets cancer types.
- HypoEvolve directionally exceeds four multi-candidate baselines on both residuals, but none survives multiple-testing correction.
- The constant-drug control gives HypoEvolve a +0.0917 DepMap advantage, while its Open Targets mean is 0.4290 versus 0.4827 for the control.HypoEvolve records 14 wins and 8 losses on DepMap, but 8 wins and 17 losses on Open Targets.
C.5 Comparison and Scoring Details
The evaluation standardizes methods, scoring, and biological mappings while comparing retrieval-sensitive Tree of Thoughts and externally selected single-pass candidates with HypoEvolve.
- Each method commits to one hypothesis before external scoring, and cancer-type scores are averaged within type for paired comparisons.Single-pass scores average six independent draws within each cancer type.
- A retrieval-free Tree of Thoughts configuration scores 0.1183 on DepMap selectivity and 0.3556 on Open Targets association.HypoEvolve’s margins over this variant are +0.0577 on DepMap and +0.0651 on Open Targets; the latter is not statistically resolved.
- +0.032 DepMap and +0.120 Open Targets are the mean margins favoring HypoEvolve over the externally best of six single-pass draws.The DepMap difference remains statistically unresolved, while the Open Targets margin is p = 0.019.
- DepMap maps drugs to curated target genes and cancers to cell lines, then subtracts each target’s pan-cancer median before taking the maximum target score.These mappings define the biological coverage and target-level interpretation of the selectivity comparison.
- HypoEvolve develops hypotheses through pairwise fitness comparison, crossover that combines parent mechanisms, and mutation that generates distinct explanations.Crossover is instructed to preserve strong mechanisms while addressing weaknesses; mutation uses analogy rather than replication.
- Hypotheses use one of 61 FDA-approved drugs with known targets covered by both Open Targets and DepMap CRISPR data.