(101g) AI-Assisted Raman Spectroscopy for Cell Analysis

Cellular senescence, cancers, and a wide range of other developmental and pathological processes often result in alterations in the chemical composition or metabolism of cells. The study of these changes continues to provide novel biomarkers and fundamental insights into these phenomena. The study of the dynamics and distributions of varying chemical features within cells and tissues has emerged as a powerful predictor of disease development and clinical outcomes. Despite its potential, the deployment of this analysis approach has been slow due to limits surrounding the number of metabolites and chemical features traditional methods are capable of assessing. The speed and cost of traditional methods is also a major concern. Techniques like mass spectrometry or chromatography destroy cell samples in the process of analysis, preventing further in vivo assessments which may be useful to researchers. Probe based methods are live-cell compatible but are inherently reagent dependent and are limited in terms of target specificity and multiplexing. To accelerate the pace of metabolic diagnostics we have developed a platform for the chemical analysis of single cells and bulk tissues using artificial intelligence (AI) and Raman spectroscopy. We capture intensity and wavenumber data which relay the relative concentration and chemical features found within cells that are capable of undergoing changes in polarizability (Raman active). Our platform consists of a preprocessing pipeline that removes artifacts, baseline corrects, and normalizes Raman spectra. We then apply linear (PCA) and nonlinear (UMAP) dimensionality reduction methods to sample spectra. The transformed datasets capture the variance of spectral features within individual samples while separately preserving local and global structures that exist across larger sample groups. These results are helpful in identifying unique chemical compositions within cells and may be used to discover metabolic subpopulations. Post preprocessing, we assess supervised (3 total) and unsupervised (5 total) machine learning model-candidates tasked with identifying cell types. Where applicable, we perform a grid search to find the optimal hyperparameters for each model. Platform users are then provided several performance metrics by which they can determine the optimal model for their specific study. Our best candidate model achieved 92.7% accuracy, 94% precision, and 93% recall, when discriminating between NIH3T3 and HEK293 cells. Our non-destructive, reagent-free method marks a significant step forward in metabolic diagnostics, offering detailed, scalable analysis for single cells and bulk tissues.