Research map / Large language models / Agents, code and math
MOOSE-Chem: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MOOSE-Chem, an agentic LLM framework that decomposes chemistry hypothesis generation into inspiration retrieval, evolutionary hypothesis composition, and ranking, and evaluates it on a 51-paper 2024 benchmark (TOMATO-Chem).
Problem
Whether LLMs can autonomously discover novel and valid chemistry research hypotheses given only a research background (question and/or survey), without data contamination.
Approach
Decomposes P(hypothesis | background) into alternating inspiration retrieval and hypothesis composition steps (an MDP view), implemented with LLM screening of a literature corpus, an evolutionary mutation/refinement/recombination unit, and LLM-based multi-aspect scoring with beam search.
Key results
- MOOSE-Chem reaches Top MS 4.020 (GPT-4o judge) vs 2.882 for MOOSE and 2.549 for SciMON.
- Inspiration retrieval hit ratio reaches over 75% of ground-truth inspirations when screening only the top 4% of a 300-paper corpus with GPT-4o (83.7%).
- Expert evaluation of top hypotheses: 21 of 51 benchmark items rated 3 or higher on the Matched Score scale (19 rated 3, 2 rated 4, 0 rated 5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MOOSE-Chem instead of MOOSE | MOOSE-Chem · TOMATO-Chem | Top MS: 2.882 -> 4.020 | Helped |
| Use MOOSE-Chem instead of MOOSE | MOOSE-Chem · TOMATO-Chem | Average MS: 2.464 -> 2.564 | Helped |
| Add multi-step inspiration retrieval and composition (ablation) | MOOSE-Chem · TOMATO-Chem | Top MS: 3.765 -> 4.020 | Helped |
| Add multi-step inspiration retrieval and composition (ablation) | MOOSE-Chem · TOMATO-Chem | Average MS: 2.730 -> 2.564 | Hurt |
| Add the evolutionary unit (mutation and recombination) (ablation) | MOOSE-Chem w/o multi-step · TOMATO-Chem | Top MS: 2.863 -> 3.765 | Helped |
| Add the evolutionary unit (mutation and recombination) (ablation) | MOOSE-Chem w/o multi-step · TOMATO-Chem | Average MS: 2.578 -> 2.730 | Helped |
| Use a screening window size of 15 rather than 60 (ablation) | GPT-4o inspiration retrieval · Corpus size 300, top 4% after screening | Hit Ratio: 53.9% -> 83.7% | Helped |
Surprising findings
- Mutation and recombination lowers average Matched Score versus the MOOSE baseline, though it raises top Matched Score.
- Llama-3.1 scaling gives an emergent-then-plateau pattern for inspiration retrieval; 70B and 405B do not clearly beat 70B on the top-4% metric.
Limitations
- Evaluation relies largely on GPT-4o-based Matched Score, with expert evaluation limited to top hypotheses per background.
- Benchmark is small (51 papers) and mostly polymer and organic chemistry.
- LLM ranking ability is only weakly established; ranking signal is confounded by whether ground-truth inspirations were used.
Experimental setup
Datasets: TOMATO-Chem 3000 most-cited Nature chemistry papers (inspiration corpus)
Benchmarks: TOMATO-Chem
Metrics: Matched Score Top MS Average MS Hit Ratio average rank ratio
Baselines: SciMON MOOSE biomedical hypothesis discovery framework
Builds on: MOOSE SciMON self-refine Swanson linking
Closest papers by meaning
- MOOSE-Chem2: Exploring LLM Limits in Fine-Grained Scientific Hypothesis Discovery via Hierarchical Search (NeurIPS 2025 · similarity 0.89)
- Can LLMs Solve Molecule Puzzles? A Multimodal Benchmark for Molecular Structure Elucidation (NeurIPS 2024 · similarity 0.89)
- Beyond Chemical QA: Evaluating LLM's Chemical Reasoning with Modular Chemical Operations (NeurIPS 2025 · similarity 0.87)
- Structured Chemistry Reasoning with Large Language Models (ICML 2024 · similarity 0.87)
- Can Agents Price a Reaction? Evaluating LLMs on Chemical Cost Reasoning (NeurIPS 2026 · similarity 0.86)
- A Sober Look at LLMs for Material Discovery: Are They Actually Good for Bayesian Optimization Over Molecules? (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.