Extracting Single-Cell Raman Signature in Scaffold Environments through Bayesian-Optimized Local Background Correction

Despite the fact that a hydrogel provides a scaffold on which cells experience elevation of stemness marker expression in cell culture, technologies for real-time, noninvasive monitoring of biochemical states in these 3-dimensional (3D) systems remain limited. Recent advances in high-speed Raman imaging enable label-free biomolecular mapping with high spatial resolution and chemical specificity; however, spectral interference from scaffold polymers often obscures cellular features and complicates data interpretation. Here, we introduce a Bayesian-optimized background subtraction method that employs spatially adaptive Gaussian kernels to minimize scaffold contamination that may have spatial inhomogeneities. Applied to Raman images of murine C2C12 cells infused with the synovial sarcoma fusion gene SS18-SSX1 cultured on scaffold hydrogel and quartz substrates, our method improved the overlap of cellular spectral signatures across substrates and reduced unwanted variation within substrates, as confirmed both in the principal component space and in the wavenumber space. Compared to conventional preprocessing techniques that consider only the spectral dimension in background correction, our method is shown to be more effective in suppressing substrate features in cell spectra. As a result, the random forest classification of cells between the types of scaffold was primarily driven by cellular information, whereas conventional methods led to decisions dominated by substrate information. This approach is expected to advance Raman imaging for hydrogel-based cell culture analysis, enabling more accurate metabolic profiling of single cells and may, after further validation, be extended to more complex biological systems such as organoids in complex in situ-like environments.