| Summary: | The advent of data science in many facets of science is a clear testament to its ability to revolutionize modern scientific discoveries. This thesis explores its application in surface-enhanced Raman scattering (SERS), a powerful spectroscopic technique that offers molecule-specific readout with high sensitivity. Despite having immense potential, practical SERS sensing applications remains hindered by poor surface affinities of the target analytes and complexity of the media they are present within. From a unique data science perspective, we design a strategy which leverages multiple molecular receptors that aim to induce receptor-analyte chemical interactions with different facets of the analyte at the plasmonic surface. The collective spectral output forms a holistic SERS ‘super-profile’ which accumulates all subtle variances embedded within and bolsters machine learning (ML) predictive models. Crucially, we demonstrate improved analyte specificities in detecting flavor compounds at the laboratory scale and breath volatile organic compounds in an actual clinical trial even in the presence of matrix interferences. To facilitate smart receptor selection, we introduce a ML-driven recommender system that maximizes SERS variance within the super-profile by selectively excluding excess uninformative features. Finally, we explore data augmentation techniques in overcoming class imbalance issues and construct robust predictive models that can be swiftly deployed for mass screenings during infectious disease outbreaks. Overall, these findings highlight the synergistic relationship between SERS and data science and are key in accelerating the practical translation of SERS sensors for diverse applications.
|
|---|