MR-MAS: an end-to-end multi-agent system for autonomous Mendelian randomization analysis

“As summarized in Table S1, it covers conventional and robust MR methods such as IVW regression (Smith and Ebrahim 2003), MR-Egger regression (Bowden et al. 2015), weighted median (Bowden et al. 2016), weighted mode (Hartwig et al. 2017), MR-PRESSO (Verbanck et al. 2018), MRMix (Qi and Chatterjee…”

“…conventional and robust MR methods such as IVW regression (Smith and Ebrahim 2003), MR-Egger regression (Bowden et al. 2015), weighted median (Bowden et al. 2016), weighted mode (Hartwig et al. 2017), MR-PRESSO (Verbanck et al. 2018), MRMix (Qi and Chatterjee 2019), GSMR (Zhu et al. 2018),…”

“With the availability of large-scale genomic resources such as the UK Biobank (Bycroft et al. 2018), OpenGWAS (Elsworth et al. 2020), and the GWAS Catalog (Buniello et al. 2019), MR has been widely adopted to elucidate causal relationships, identify actionable risk factors and prioritize therapeutic targets.”

“For each pair, we applied five core MR methods: IVW, MR-Egger regression, weighted median, weighted mode, and MR-PRESSO, and recorded the corresponding P -values and effect directions.”

“A full analysis typically involves multiple steps, including GWAS data retrieval, instrument selection, causal effect estimation, sensitivity analyses, and result interpretation ( Lawlor et al. 2008, Hemani et al. 2018), often requiring frequent switches between different tools and platforms.”

“Mendelian Randomization (MR) offers a powerful alternative by leveraging the random allocation of genetic variants and using them as instrumental variables for causal inference (Smith and Ebrahim 2003).”

“A full analysis typically involves multiple steps, including GWAS data retrieval, instrument selection, causal effect estimation, sensitivity analyses, and result interpretation ( Lawlor et al. 2008, Hemani et al. 2018), often requiring frequent switches between different tools and platforms.”

“A growing body of work has demonstrated that LLM-driven agent systems (Wei et al. 2022, Yao et al. 2023) can perform complex scientific tasks via role specialization and collaborative reasoning.”

“Specifically, it provides programmatic access to the GWAS Catalog for GWAS study retrieval and summary statistics download, and supports PubMed (Canese and Weis 2013) and PubMed Central retrieval for literature evidence collection.”

“However, traditional observational studies are often affected by unmeasured confounding and reverse causation (Greenland et al. 1999, Lawlor et al. 2008, Toonsi et al. 2024).”

“In recent years, large language models (LLMs) have shown substantial potential for building ‘AI scientists’ through their capabilities in natural language processing, reasoning, code generation and task planning (Achiam et al. 2023, Chang et al. 2024).”

“Mendelian Randomization (MR) offers a powerful alternative by leveraging the random allocation of genetic variants and using them as instrumental variables for causal inference (Smith and Ebrahim 2003).”

“It then combines keyword-based recall with semantic ranking based on TF-IDF similarity (Aizawa 2003) to identify the most relevant trait candidates.”

“As summarized in Table S1, it covers conventional and robust MR methods such as IVW regression (Smith and Ebrahim 2003), MR-Egger regression (Bowden et al. 2015), weighted median (Bowden et al. 2016), weighted mode (Hartwig et al. 2017), MR-PRESSO (Verbanck et al. 2018), MRMix (Qi and Chatterjee 2019), GSMR (Zhu et al…”

“With the availability of large-scale genomic resources such as the UK Biobank (Bycroft et al. 2018), OpenGWAS (Elsworth et al. 2020), and the GWAS Catalog (Buniello et al. 2019), MR has been widely adopted to elucidate causal relationships, identify actionable risk factors and prioritize therapeutic targets.”

“ChemCrow (Bran et al. 2024) tightly integrates LLMs with a chemistry toolset to achieve automated reaction planning and experimental task execution.”

“…on real GWAS research questions, we assessed its performance using exposure–outcome pairs stratified into five prior-evidence categories (Morrison et al. 2020): considered causal, supported by literature, unknown or conflicting evidence, implausible or unsupported, and considered…”

“In biomedical research, accurately identifying causal relationships between traits, such as risk factors, lifestyle behaviours, or biomarkers, and diseases is fundamental for understanding disease mechanisms and designing effective interventions (Russo and Williamson 2007).”

“For analytical consistency, MR-MAS was run with gpt-4o (Hurst et al. 2024) on exposure–outcome pairs spanning five prior-evidence categories.”

“Virtual Lab (Swanson et al. 2025) further demonstrates the innovative capacity of agent systems in real biomedical scenarios, where newly designed SARS-CoV-2 nanobodies exhibit strong performance in experimental validation.”

“For instance, CRISPR-GPT (Qu et al. 2025) can autonomously plan and analyse gene-editing experiments, SciAgents (Ghafarollahi and Buehler 2025) combines multiple agents with knowledge-graph reasoning to generate hypotheses and predict material properties, and AutoBA (Zhou et al. 2024) leverages…”

“TraitMatcher is evaluated using Top-k hit rates under noisy query settings, with the reported results averaged across all four base model backends: o1 (Jaech et al. 2024), gpt-4o, o1-mini (Jaech et al. 2024), and DeepSeek-V3.”

“AI Scientist-v2 (Yamada et al. 2025) constructs a fully automated framework that spans hypothesis generation, experiment design and execution, result analysis and manuscript writing, and further advances the ‘AI scientists’ paradigm by incorporating a multi-agent collaboration mechanism guided by tree search.”

“MRAgent reported three definitive positive causal associations, whereas MR-MAS reported none.”

“However, traditional observational studies are often affected by unmeasured confounding and reverse causation (Greenland et al. 1999, Lawlor et al. 2008, Toonsi et al. 2024).”