Diagnosing Disease Using High Throughput Parallelized Nanofluidics And Machine Learning
I have developed microchip-based diagnostic platforms to enrich for specific rare blood based cells and extracellular vesicles and extract useful clinical diagnoses and prognoses from the multi-dimensional molecular cargo that these biomarkers contain. In my thesis, I have demonstrated the power of this approach by focusing on two applications, badly in need of improved diagnostics and prognostics, the early diagnosis of pancreatic cancer and the prognosis of traumatic brain injury. This work presents miniaturized, low cost, and clinical deployable technologies that are designed to personalize medicine and greatly improve treatment efficacy and patient outcomes. Blood-based biomarkers contain rich molecular information with high diagnostic and prognostic values but they are challenging to work with due to the rarity, heterogeneity, small sizes, and the presence in the high background. Microfluidics has been extremely successful sorting cells by matching the scale of the device to the scale of the cells, which allows precise sorting, control, and measurement of the cells. However, there have been several key challenges to translate microfluidics to clinical applications, including susceptibility to clogging, low throughput, and poor scalability to the nanoscale. To solve these challenges and achieve the goal of building a robust tool that can be translated to clinical settings, I developed a strategy to incorporate millions of nano/microfluidic devices onto microchip platforms that operate in parallel, increasing throughput by a million-fold and eliminating susceptibility to clogging while preserving precise sorting. We achieved high specificity by coating millions of nano/microfluidic devices with a magnetic material, NiFe, to specifically enrich for the materials of interest using immunomagnetic capture. To resolve the biological heterogeneity inherent to individuals and their diseases, we measured and analyzed multiple nucleic cargos of the blood-based biomarkers using machine learning algorithms to identify signatures that persist across the subject-to-subject variability, endemic to all diseases. Using this approach, we were able to isolate specific subsets of rare blood-based biomarkers (circulating tumor cells, exosomes), profile their nucleic acid cargo, and detect within this data multidimensional molecular signatures for specific disease states. The identified molecular signatures can classify disease states that are currently difficult to predict for pancreatic cancer and traumatic brain injury in pre-clinical models (e.g. pre-cancerous lesions versus pancreatic cancer, characterization of severity, history, and time since brain injury). To test the translatability of the developed diagnostic platforms, we applied the platforms to clinical samples and accurately classified pancreatic cancer patients from healthy controls and traumatic brain injury patients with different injury levels and elapsed time since injury from healthy controls. By combining microfluidics and machine learning, we can discover new biomarkers for diseases that are encoded in the complex molecular data packaged inside of nanoscale vesicles, which circulate amongst the vast background of information present in patient blood. This work has broad potential to be further developed for use on a wide range of cancers, brain related diseases, and beyond. The ability to measure and identify specific molecular signatures by resolving the inherent biological variability provides an important addition to the development of personalized medicine that can guide patients to the right treatment and improve therapeutic efficacy of novel and emerging treatments.