Reproduce “LabAgent: Customize Any Research Hubs for Scientific Discoveries Using AI Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lei Liu, Yikun Zhang, Jialin Chen, Wanjia Zhao, Rex Ying, Wengong Jin, Hua Xu, James Zou and 2 more
LabAgent builds a laboratory-specific, executable knowledge base from papers, repositories, documentation, and prior troubleshooting records to support reproducible scientific analysis. Across biomedical benchmarks, it reproduced 32 of 65 ADMET results within published variation, scored 74.4 on BiomniBench-DA, achieved the strongest mean performance on GeneBench-Pro among compared agents, and selected protein variant-effect predictors reaching a mean Spearman correlation of 0.443 across 13 assays.
Key insights
- Executable laboratory knowledge is more useful than a static collection of papers or code: validated skills encode data preparation, environment requirements, quantitative checks, and failure modes in a reusable format.
- Persistent troubleshooting memory improved the agent’s ability to recover from failures; 116 of 145 failing-turn transitions were followed by a successful subsequent command in the ADMET experiments.
- Scientific performance depends on decisions beyond code execution, including predictor selection and identification of technical variables. In a cis-eQTL example, accounting for donor-level ambient RNA contamination moved the CXCL10 estimate into the target range.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.