ToxPipeMCP: An AI Access Solution for Toxicology Databases Based on the MCP Protocol

The ToxPipeMCP project, open-sourced by the U.S. National Institute of Environmental Health Sciences (NIEHS), builds a standardized interface layer based on the Model Context Protocol (MCP). It provides large language models (LLMs) with query capabilities for the ChemBioTox toxicology database and literature retrieval through the FastMCP framework, enabling seamless integration between AI and professional scientific databases. This addresses the pain points of low efficiency in traditional toxicology data retrieval and LLMs' difficulty accessing professional databases, providing a powerful AI-assisted tool for environmental health science research.

In environmental health science and toxicology research, traditional retrieval requires switching between multiple databases and manually integrating information, which is inefficient and prone to data omission. With the development of LLM technology, researchers hope AI can directly access professional databases. However, general-purpose LLMs lack the ability to access professional databases, and simple interfaces are difficult to seamlessly integrate with AI's dialogue capabilities. The ToxPipeMCP project emerged as a solution, building a standardized interface layer based on the MCP protocol proposed by Anthropic, allowing LLMs to query the ChemBioTox database and retrieve literature as if calling functions.

Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an open standard launched by Anthropic, providing AI with a standardized way to connect to external data sources. Its core components include MCP Host (AI runtime environment), Client (communication component), Server (service side), etc. The workflow is: AI identifies needs → Host sends requests via Client → Server executes queries → returns results → AI generates answers.

Technical Implementation

• FastMCP Framework: A Python library that simplifies MCP Server development via decorators, automatically handling underlying details like communication and serialization.
• ChemBioTox Integration: Provides chemical substance retrieval (name/CAS number), toxicology data query (acute/chronic toxicity, etc.), safety limit query, and links to related literature.
• Literature Retrieval: Supports searching academic databases like PubMed to obtain abstracts and filter relevant studies.

1. Rapid Toxicology Information Query: Ask AI for toxicology data of a chemical (e.g., acute inhalation toxicity of benzene), and AI will obtain and present accurate data via ToxPipeMCP.
2. Literature Review Assistance: Query the research status of a chemical (e.g., studies on the health effects of microplastics in the past five years), and AI will retrieve literature to provide an overview, key findings, and a list of relevant studies.
3. Safety Assessment Support: Ask about workplace exposure limits and occupational risks of chemicals, and AI will integrate database information to provide support.

Technical Advantages

• Standardized Interface: Can integrate with any AI application that supports MCP via the MCP protocol, reducing costs.
• Safe and Controllable: MCP Server runs locally, allowing control over data access scope and permissions.
• Scalable: Easy to add new features based on FastMCP.
• Transparent and Auditable: Queries are explicitly traceable, ensuring results are verifiable.

Scientific Significance

• Improve Research Efficiency: Obtain information directly via natural language without needing to be familiar with the syntax of multiple databases.
• Reduce Information Omission: AI queries multiple data sources simultaneously for cross-validation.
• Promote Interdisciplinary Collaboration: Non-toxicology professionals can also access professional information.
• Support Decision-Making: Quickly obtain data to support chemical management and policy formulation.

Current Limitations

• Data Coverage: Mainly integrates ChemBioTox; other toxicology databases need to be expanded.
• Query Complexity: Complex joint query and analysis functions need improvement.
• Result Interpretation: AI's interpretation of professional data requires expert verification.

Future Directions

• Multi-Database Integration: Integrate databases from institutions like EPA and ECHA.
• Advanced Analysis: Support toxicology prediction models and dose-response analysis.
• Visualization: Generate data charts for presentation.
• Collaboration Features: Support team sharing of queries and results.