Exosomes are extracellular vesicles that propagate in the body as a form a cell-to-cell communication, implicated in many diseases such as cancer and neurodegeneration. To understand the impacts of exosomal messages, it is important to determine the message source: the organ system that initially secreted them. To do so, we develop a new technique based on protein language models (PLMs); PLMs with Transformer neural architecture now learn powerful protein representations in a self-supervised manner. Learned protein representations can be used to estimate the source organs of a protein. Using a pre-trained Transformer-based PLM as a feature extractor and fine-tuning a prediction model over the extracted features to predict source organs, yields reasonable predictive accuracy. We apply this new analysis tool to bulk exosomal proteomics data to understand differences between healthy aging and neurodegenerative disease.