Retrieval-Augmented Language Models Enable Scalable Chemical Source Classification in Metabolomics Workflows

There is a growing need for scalable chemical classification to support the interpretation of exposomics and metabolomics data. While structural categorization has been largely automated, functional and exposure-based labeling of chemicals remains a manual and time-consuming process. Here, we present chemsource, a flexible framework that integrates large language models (LLMs) with retrieval-augmented generation (RAG) to automate chemical classification. chemsource retrieves descriptive text from Wikipedia or PubMed abstracts based on chemical names and prompts LLMs to assign user-defined categories based on the retrieved content. We demonstrate classification into five exposure categories: endogenous metabolites, food molecules, drugs, personal care products, industrial chemicals, and combinations of these possibilities. Benchmarking against manually curated labels for 4,953 compounds showed 75% overall agreement, with category-level recall exceeding 75% across all classes. Expert review indicated that most discrepancies could be attributed to prompt ambiguity and incomplete manual labels rather than model failure. To demonstrate the utility of chemsource in metabolomics workflow, we applied it to eight public untargeted metabolomics data sets, revealing distinct exposure patterns across human biospecimens, mouse tissues, environmental dust, and consumer product extracts. chemsource is customizable via prompt editing, enabling diverse classification tasks without requiring coding expertise. The tool is freely available as a Python package (https://pypi.org/project/chemsource/). Text retrieval is free; classification requires user-supplied LLM API access.