Automated Cryo-EM and Supervised Machine Learning Enable Reproducible Characterization of Extracellular Vesicles and Co-Isolating Particles

Abstract

Cryo-electron microscopy (cryo-EM) is the only technique capable of visualizing the lipid bilayer of extracellular vesicles (EVs), enabling their distinction from non-EV particles. However, the application of cryo-EM for EV sample characterization has been limited by a combination of low imaging throughput and complex image analysis of the structurally diverse EVs. To address these challenges, we developed a workflow combining automated cryo-EM image acquisition with supervised machine learning (sML)-assisted particle detection and classification. Automated image acquisition facilitates the routine acquisition of thousands of cryo-EM images with consistent quality, enabling the imaging of hundreds of EVs. sML-assisted particle detection enabled efficient and reproducible identification, size measurement, and structural classification of EVs. Furthermore, using sML we are able to differentiate EVs from non-EV particles, such as lipoproteins and protein aggregates, which might co-isolate due to overlapping physical properties or by physical association with EVs. In mixed EV-lipoprotein samples, we demonstrate that our pipeline can distinguish EVs and differentiate between high-density (HDL), low-density (LDL), and very low-density (VLDL) lipoproteins. Our automated cryo-EM and sML workflow overcomes key limitations of EV characterization using cryo-EM by increasing imaging throughput and enabling reproducible EV characterization. Furthermore, this method provides a tool for analysing EV heterogeneity, sample purity, and co-isolated contaminants, advancing the field of EV research.