MARRVEL-MCP: An agentic interface for Mendelian disease discovery via tool-augmented context engineering

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence in Medicine

Background:

• Variant interpretation for rare diseases is complex, requiring expertise across multiple genomic databases.
• Existing tools like MARRVEL have usability barriers due to strict input formats and heterogeneous outputs.
• This complexity poses challenges for both non-experts and specialists, increasing cognitive load.

Purpose of the Study:

• To develop MARRVEL-MCP, a natural-language interface for large language models (LLMs) to automate end-to-end variant interpretation.
• To demonstrate the effectiveness of tool-augmented context engineering in enhancing LLM capabilities for genomics.
• To improve the accessibility and efficiency of rare disease variant analysis.

Main Methods:

• Developed MARRVEL-MCP, integrating LLMs with 44 diverse genomic tools via structured function interfaces.
• Enabled LLMs to perform named-entity recognition, identifier normalization, and multi-database synthesis from clinical queries.
• Utilized tool-augmented context engineering to guide LLM decision-making without hard-coded workflows.

Main Results:

• Lightweight LLMs (3B-20B parameters) equipped with MARRVEL-MCP matched or exceeded the performance of larger models without tool access on 100 expert-curated questions.
• A 20B-parameter model achieved a 94% success rate with MARRVEL-MCP, significantly outperforming the 41% success rate without it.
• Demonstrated that contextual guidance can compensate for limited model capacity in complex genomic tasks.

Conclusions:

• MARRVEL-MCP significantly enhances LLM performance in variant interpretation through structured tool access and context engineering.
• Context engineering is a fundamental principle for advancing biomedical AI and integrating LLMs with genomic resources.
• This approach supports scalable and more accessible rare disease diagnosis and research.