BioE PhD Thesis Presentation - Rajas Poorna

Advisors:

Saad Bhamla, Ph.D. (School of Chemical and Biomolecular Engineering, Georgia Institute of Technology)

Marcus Cicerone, Ph.D. (School of Chemistry and Biochemistry, Georgia Institute of Technology)

Committee Members:

Nicholas Hud, Ph.D. (School of Chemistry and Biochemistry, Georgia Institute of Technology)

Mark Prausnitz, Ph.D. (School of Chemical and Biomolecular Engineering, Georgia Institute of Technology)

Francisco E Robles, Ph.D. (Wallace H. Coulter Department of Biomedical Engineering, Georgia Institute of Technology and Emory University)

Bringing Universal Diagnostics to the Point-of-Care: Raman Spectroscopy and Machine Learning

Raman spectroscopy of biofluids, cells, and tissues can diagnose a wide range of diseases – including diabetes, malaria, tuberculosis, celiac disease, cancer, and Alzheimer's – and quantify analytes such as glucose, albumin, and urea. This versatility has led to its description as a potential “universal” medical diagnostic. This thesis advances this vision along three fronts.

First, it introduces a method to improve the biological state resolution inferable by Raman imaging. While existing approaches commonly use spatial averaging to obtain a single spectrum, this work uses nonlinear dimensionality reduction to incorporate spatial information into the inferred cell state. Applied to imaging of egg maturation of a live C. elegans with BCARS – a highly sensitive and reproducible Raman imaging technique – this approach enables state resolution comparable to transcriptomics. This establishes the possibility that BCARS imaging combined with machine learning could serve as an omics surrogate in live cells.

Second, this machine learning approach is generalized into ScaleMAP, a dimensionality reduction technique that preserves local scale information typically erased by state-of-the-art methods such as UMAP, while retaining their neighborhood quality. Across diverse datasets, ScaleMAP recovers important structures that these techniques miss, including rare spectral “spikes” in Raman imaging, transitional "bridge" cells connecting distinct cell types in transcriptomics, and density structure spanning 17 orders of magnitude in flow cytometry.

Third, this thesis confronts the central barrier to real-world Raman diagnostics: the absence of an instrument that is simultaneously sensitive, reproducible, and affordable. The Raman diagnostics literature commonly uses instruments such as the Renishaw inVia Qontor, which can cost upwards of $200,000, yet exhibit limited reproducibility. This work demonstrates a proof-of-concept low-cost biofluid Raman spectrometer that surpasses the Qontor in both sensitivity and repeatability, at a current bill-of-materials of $1,300, with a roadmap to $150. It then leverages this repeatability to develop a proof-of-concept low-cost field calibration method for reproducibility across point-of-care centers.

Together, these contributions advance Raman diagnostics in two respects: the diagnostic information extractable from each measurement, and the translational path to obtaining measurements reliably in the real world.