EV-PROTEOMICS: A Database-Driven Computational Framework for Refining Extracellular Vesicle-Associated Protein Profiles in Genotoxicity and Carcinogenicity Research

Abstract Carcinogenicity assessment using rodent bioassays is constrained by interspecies differences, cost, duration, and ethical concerns, highlighting the need for human-relevant New Approach Methodologies (NAMs). Extracellular vesicles (EVs) are potential sources of molecular biomarkers, but heterogeneous EV-associated proteomic data can complicate the identification of mechanistically informative patterns. We present EV-PROTEOMICS, a database-driven computational framework for refining curated EV-associated protein profiles across predefined toxicological reference categories. Protein profiles were represented as binary data and processed using a subtractive refinement strategy. Twelve reference substances representing genotoxic and non-genotoxic categories were evaluated. Refinement reduced the genotoxicity-associated protein set from 245 to 119 proteins. Jaccard dissimilarity and Principal Coordinates Analysis showed that the proportion of variation represented by the first two axes increased from 31.2% in the initial dataset to 66.3% after refinement. Further mechanistic exploration prioritized candidates, including NDRG2 and NFE2L2, in profiles associated with aromatic amine/nitrosamine reference substances. These findings provide an in silico proof of concept for refining EV-associated protein datasets and generating candidate signatures for mechanistic toxicology research. Independent datasets and experimental studies are required to assess reproducibility and determine discriminatory or predictive performance.