ToxPipe
Parker K. Combs1,4, Trey O. Saddler1,3, Amlan Talukder2,5, Jonathon F. Fleming1,6, Jeremy N. Erickson1, David M. Reif1, Charles P. Schmitt2, Scott S. Auerbach1
Affiliations
1. Predictive Toxicology Branch (PTB), NIEHS, NIH, Durham, NC, USA
2. Office of Data Science (ODS), NIEHS, NIH, Durham, NC, USA
3. Dynanet Corporation, Elkridge, MD, USA
4. Axle Informatics, North Bethesda, MD, USA
5. 22nd Century Technologies, McLean, VA, USA
6. GAP Solutions, Herndon, VA, USA
DOI: https://doi.org/10.22427/DATA-500-110-003-000-0
Publication
Abstract
ToxPipe is a modular, fully open-source software stack designed to accelerate toxicological data analysis and decision-making by integrating large language model (LLM) capabilities with workflow, monitoring, and user-interface services. ToxPipe leverages existing open-source technologies like LibreChat, LiteLLM, and Langflow to provide centralized and streamlined access to different LLMs for end users and developers and logging and auditing of LLM use for organizations. ToxPipe also allows for open-source/local model support. While the ToxPipe ecosystem also supports advanced AI functions such as MCP server integration, RAG and literature search, and agent creation, we did not observe clear benefits using these configurations relative to the GPT-5.5 baseline assertion coverage scores. ToxPipe is available as a series of Docker containers for rapid and reproducible deployment.
Data Sets
Main Code Repository
https://github.com/NIEHS/ToxPipe
Evaluations Repository
https://github.com/NIEHS/ToxPipe-Model-Comparison
RAG Embeddings Database Repository
https://github.com/NIEHS/ToxPipeRAG