MR-MAS: An End-to-End Multi-Agent System for Autonomous Mendelian Randomization Analysis

Understanding disease-related causal mechanisms is essential for biomedical research, yet the practical use of Mendelian Randomization (MR) remains constrained by fragmented analytical workflows and a strong dependence on specialized expertise. These challenges limit the efficient, standardized, and reproducible execution of end-to-end MR analyses.

We propose MR-MAS, an end-to-end automated MR analysis system built on a multi-agent architecture. MR-MAS leverages the reasoning and coordination capabilities of large language models (LLMs) and integrates external tools to unify data acquisition, causal effect estimation, results interpretation, and manuscript preparation into a single workflow, thereby reducing manual effort. MR-MAS further incorporates a structured Mendelian Randomization Knowledge Base (MRKB) to support MR method implementation and targeted repair of MR analysis code. Across diverse test scenarios, MR-MAS demonstrated stable workflow execution, reproduced several literature-supported causal signals, and maintained consistent causal estimates across repeated runs and different model backends. Additional evaluations showed robust trait retrieval and improved code-repair performance with MRKB guidance. These results suggest that MR-MAS may provide practical support for scalable, standardized, and reproducible MR analyses in biomedical research.

The source code, MRKB, and benchmark data are freely available at https://github.com/Zaiwen/MR-MAS, with the version used in this study archived at https://doi.org/10.5281/zenodo.20342636.

Supplementary data are available at Bioinformatics online.