Asking the Data: Conversational Interrogation of LC-MS/MS Proteomics Data Sets with an LLM Agent

A conversational LLM agent for LC-MS/MS proteomics data set interrogation exposed multiple analysis failures across four datasets, including 99.2% matching peptide assignments and r = 0.93 fold-change correlation on reanalysis.

• Why it matters: Proteomics pipelines answer predefined questions, but emerging analytical problems are hard to test without extra bioinformatics, risking misinterpretation.
• What they did: The framework translates researcher questions into executable tests with explicit controls and dated, traceable code, applied retrospectively to four LC-MS/MS/MS datasets and reanalyzed a public TMT dataset from raw files in 24 min.
• The result: It changed interpretations of three in-house studies (e.g., missing symbiont in the search database, swapped TMT channels, inflated statistics, and false deamidation from 13C monoisotopic selection) and found failed/partial cysteine alkylation in all three.