Machine Learning for Individual EV Classification Based on Highly Sensitive Multiplexed Mass Spectrometry Measurements.

Here, we report on combining Random Forest (RF) classification with nanoprojectile secondary ion mass spectrometry (NP-SIMS) to analyze single extracellular vesicles (EVs) isolated from human liver cancer (HEPG2) and Normal liver cell lines. EVs were tagged with antibody-lanthanide (Ln) tags specific to marker proteins and dispersed on a surface, enabling NP-SIMS to produce millions of individual EV mass spectra. Previously, EVs were manually classified based on Ln-tag signals, and as a result, only 5% could be confidently classified as either from cancer or normal cells. Using a random forest model, optimizing data preprocessing, and expanding the spectral features resulted in a 60-fold increase in classification efficiency over manual analysis. We also performed untargeted RF classification, where both supervised and untargeted RF analyses resulted in consistent outcomes, showing the compatibility of RF with the NP-SIMS data for individual EV classification. The results from the untargeted analysis suggest that NP-SIMS with RF could aid in marker discovery in systems where limited sample quantities are available. Overall, using RF and NP-SIMS enables single-EV classification and provides a promising pathway for EV-based disease diagnostics.