CASCADE: continuous wavelet transform-Adam spectral curve-fitting and decomposition engine for bio-Raman peak fitting

Omics technologies have transformed biology but remain destructive, static, and incompatible with longitudinal measurements in living systems. Raman fingerprint spectroscopy offers a compelling noninvasive omics surrogate. Broadband coherent anti-Stokes Raman scattering (BCARS) reveals coordinated biochemical changes in proteins, lipids, and metabolites at information densities comparable to transcriptomics, with subcellular spatial resolution and millisecond acquisition times. Despite strong, widely demonstrated correlations between Raman spectral features and gene program activity, generalizable interpretation models have remained elusive. A central barrier to the development of such models is that raw spectral intensities are a poor substrate for transferable machine learning; baseline artifacts and instrument-dependent offsets can dominate biologically meaningful variation. Generalizable inference requires operation in the space of physically meaningful peak parameters which are stable, chemically interpretable, and instrument-independent. With Raman spectroscopy, we can define a single peak by its amplitude, position, and width parameters. Here we introduce CASCADE (CWT-Adam Spectral Curve-fitting and Decomposition Engine), an automated pipeline that transforms Raman fingerprint hyperspectral images into quantitative peak-parameter representations at every pixel, combining continuous wavelet transform initialization with physics-constrained Adam gradient descent. Validation on synthetic datasets with known ground truth demonstrates F₁ scores exceeding 80%, position accuracy better than 0.5 cm^-1, and amplitude errors below 10% across a wide range of noise levels and peak separations. Application to BCARS images of differentiating neuroblastoma cells reveals spatially resolved metabolic signatures invisible in averaged spectra. CASCADE establishes a fast, reliable translation from raw fingerprint spectra to a peak parameter space, providing an objective common language to realize generalizable Raman-based biological inference for spectral interpretation by researchers and machine learning alike.