Deep Learning-Decoded Raman Spectroscopy for Hour-Scale iPSC Pluripotency Assessment via Lipid–Protein Biomarkers

Rapid and label-free evaluation of induced pluripotent stem cell (iPSC) pluripotency is critical for advancing regenerative medicine and clinical applications. Although traditional genomics- and proteomics-based pluripotency assessment methods are reliable, their invasive nature, reliance on labeling, and time-intensive workflows limit their suitability for dynamic monitoring. Here, we present a method combining deep learning with Raman spectroscopy for an hour-scale label-free pluripotency assessment in iPSCs. By inducing pluripotency modulation through a culture medium alteration and simultaneously acquiring correlated Raman spectra, we established spectral data sets of iPSCs at distinct pluripotent states. Using these data sets as input, we trained a one-dimensional convolutional neural network (1D-CNN) to classify pluripotent states with an average accuracy of 97.10%. Remarkably, the model achieved 98.00% accuracy in detecting pluripotency anomalies at the 1 h time point of medium perturbation─prior to observable morphological changes─establishing the fastest reported detection framework for iPSC quality control. Gradient-weighted class activation mapping (Grad-CAM) shows that lipids and proteins (with Raman peaks at 1440 and 1660 cm⁻¹, respectively) are biomarker signatures directly linked to pluripotency states. Further, the pluripotency of iPSCs was confirmed to be related to PI3K/AKT pathway activity. This integration of Raman spectroscopy and interpretable deep learning bypasses prior biomarker knowledge requirements, offering a paradigm shift toward clinical-grade noninvasive stem cell diagnostics.