Pushing the Boundaries of Biomolecule Characterization through Deep Learning

Abstract: The importance of studying biological molecules in living organisms can hardly be overstated as they regulate crucial processes in living matter of all kinds. Their ubiquitous nature makes them relevant for disease diagnosis, drug development, and for our fundamental understanding of the complex systems of biology. However, due to their small size, they scatter too little light on their own to be directly visible and available for study. Thus, it is necessary to develop characterization methods which enable their elucidation even in the regime of very faint signals. Optical systems, utilizing the relatively low intrusiveness of visible light, constitute one such approach of characterization. However, the optical systems currently capable of analyzing single molecules in the nano-sized regime today either require the species of interest to be tagged with visible labels like fluorescence or chemically restrained on a surface to be analyzed. Ergo, there exist effectively no methods of characterizing very small biomolecules under naturally relevant conditions through unobtrusive probing. Nanofluidic Scattering Microscopy is a method introduced in this thesis which bridges this gap by enabling the real-time label-free size-and-weight determination of freely diffusing molecules directly in small nano-sized channels. However, the molecule signals are so faint, and the background noise so complex with high spatial and temporal variation, that standard methods of data analysis are incapable of elucidating the molecules' properties of relevance in any but the least challenging conditions. To remedy the weak signal, and realize the method's full potential, this thesis' focus is the development of a versatile deep-learning based computer-vision platform to overcome the bottleneck of data analysis. We find that said platform has considerably increased speed, accuracy, precision and limit of detection compared to standard methods, constituting even a lower detection limit than any other method of label-free optical characterization currently available. In this regime, hitherto elusive species of biomolecules become accessible for study, potentially opening up entirely new avenues of biological research. These results, along with many others in the context of deep learning for optical microscopy in biological applications, suggest that deep learning is likely to be pivotal in solving the complex image analysis problems of the present and enabling new regimes of study within microscopy-based research in the near future.