ScProteoAgent enables natural-language-driven single-cell proteomics analysis and interpretation

Abstract

Single-cell proteomics requires computational choices to remain aligned with experimental design as research questions evolve. Here we present ScProteoAgent, which translates natural-language requests into domain-specific calculations and preserves research intent, analysis design, matrix-processing history and statistical outputs as reusable analytical state. Follow-up requests reuse applicable inputs and initiate new calculations when contrasts or experimental units change. We assembled a benchmark comprising 55 tasks and 33 reference conclusions from 11 published studies, combining progressive questions within studies with complementary biological and analytical settings. In a retrospective comparison of 88 archived outputs from eight systems, ScProteoAgent achieved the highest mean score of 93.67 under a common six-dimensional rule-based evaluation. A HeLa migration case connected phenotype composition with proteomic clusters, candidate differences and original-observation support. In liver zonation, a researcher-guided follow-up reused a saved matrix to test mouse-level spatial contrasts and compare full-matrix and observed-only results. Applications in brain development, sample preservation and hematopoiesis illustrated how interpretation depends on processing choices and experimental units. ScProteoAgent connects research questions with statistical results and their analytical context, enabling existing analyses to support more specific biological investigations.