MARRVEL-MCP enables natural language variant interpretation through autonomous workflow construction

Abstract

AO_SCPLOWBSTRACTC_SCPLOWRare disease variant interpretation requires navigating multiple genomic databases with strict input formats and synthesizing heterogeneous evidence, creating barriers for non-experts and cognitive burdens even for specialists. MARRVEL exemplifies this challenge by requiring precise queries (e.g., HGVS notation) and returning complex, difficult-to-synthesize outputs. To address this input-output asymmetry, we developed MARRVEL-MCP, a natural language interface enabling large language models (LLMs) to perform end-to-end variant interpretation via structured tool access. This work demonstrates the impact of context engineering--the deliberate design of domain-aware tool environments and information scaffolding--in reshaping the role of model scale in genomics. MARRVEL-MCP equips LLMs with 39 tools spanning gene and variant utilities, pathogenicity databases, phenotype resources, expression atlases, ortholog data, and literature APIs. Without hard-coded pipelines, LLMs autonomously infer workflows, performing named-entity recognition, identifier normalization, and multi-database synthesis from narrative queries. Using 45 expert-curated questions, lightweight models (3B-20B parameters) with MARRVEL-MCP matched or outperformed larger models lacking tool access. A 20B-parameter model (gpt-oss-20b) achieved a 95% pass rate versus 33% without MARRVEL-MCP, approaching state-of-the-art proprietary performance. Although accuracy remains below the 100% required for autonomous diagnosis and token usage is considerable, these results show that well-designed context can compensate for limited model capacity. These findings establish context engineering as a core principle for biomedical AI, enabling efficient, scalable integration of LLMs with curated genomic resources. MARRVEL-MCP is available at https://github.com/hyunhwan-bcm/MARRVEL_MCP/.