Machine Learning-enabled Classification of Extracellular Vesicles Using Nanoplasmonic Microfluidics - PROJECT SUMMARY/ABSTRACT Ovarian cancer is a diverse group of malignancies that can vary greatly in molecular biology, etiology, and presentation of symptoms. While it accounts for 2.5% of all cancers among women, it results in roughly 5% of cancer-related deaths due to its high fatality rate. This is because 75% of patients are diagnosed with advanced disease, largely attributed to its relatively late presentation of symptoms and a lack of reliable detection and monitoring strategies. Extracellular vesicles (EVs) are released by all cells, including ovarian cancer, and their cargo reflects their cells of origin. They have shown immense potential as stable biomarkers, however their low abundance compared to EVs from healthy cells and a lack of sufficiently sensitive characterization tools has limited their clinical translation. Surface-enhanced Raman spectroscopy (SERS) is sensitive enough to biochemically fingerprint even single EVs, and the information-rich spectra produced in EV SERS can be fed into machine learning (ML) algorithms to classify them based on their latent spectral features. Despite early progress in EV SERS, the highly heterogeneous nature of EVs indicates that their separation into distinct subpopulations prior to SERS analysis may help improve disease diagnosis, classification, and monitoring. Microfluidic devices are uniquely capable of separating EV s into subpopulations of interest while simultaneously enabling SERS spectral acquisition in a single device. In the proposed research, during the mentored phase, ML-enabled inverse design will be combined with high resolution nanofabrication techniques to improve the signal strength and spectral quality attainable from EV SERS. Preliminary data indicates that the SERS enhancement is highly dependent on the substrate’s nanoscale geometry, which is particularly important for EVs compared to conventional chemical analysis. Towards the end of the mentored phase, once the improved substrates have been thoroughly tested using bioreactor-produced EVs, they will be incorporated into two distinct microfluidic devices and tested throughout the independent phase using both cell culture and patient EV samples. One device will capture different subpopulations of EVs directly onto the microfluidic SERS substrates based on specific surface- antigens for multiplexed characterization, while the other will separate EVs precisely by size and flow them over the microfluidic SERS substrates to produce EV SERS barcodes. In parallel, ML algorithms tailored specifically to these platforms will also be developed to process and classify the acquired spectra. This proposal is multidisciplinary, utilizing advanced ML and micro-nanofabrication techniques as well as EV production, isolation, and characterization strategies. These experiments are innovative and significant because they will develop ML inverse design architectures specifically for EVs as well as microfluidic EV SERS for characterizing and classifying ovarian cancer.
