Integrated UPLC-MS/MS with interpretable artificial intelligence-enhanced Raman and near-infrared spectroscopy for
Qiuya Zhou1, Meifang Jiang2, Dandan Wang2
1College of Pharmaceutical Science, Zhejiang University of Technology, Hangzhou 310014, PR China.
Abstract:
The column purification process plays a critical role in the production of ginkgo ketone ester tablets, and monitoring the concentrations of key compounds is essential for ensuring final product quality. Herein, a quality‑by‑artificial‑intelligence concept integrating real‑time spectral monitoring with predictive modeling was proposed, utilizing portable Raman and near‑infrared spectrometers (NIRS) to acquire spectral data from 217 eluate samples. UPLC‑MS/MS was employed to quantify seven key compounds, including four terpene lactones (bilobalide, ginkgolides A-C) and three flavonol glycosides (quercetin, kaempferol, isorhamnetin). Subsequently, seven quantitative calibration models were constructed using seven different artificial intelligence algorithms. Notably, the ResNet50‑based model exhibited optimal performance. For Raman spectral data analysis, the correlation coefficients (RP2) between predicted and reference values for the seven active components were 0.9944, 0.9953, 0.9945, 0.9899, 0.9961, 0.9953, and 0.9931, with corresponding root mean square error of prediction (RMSEP) values of 0.0172, 0.0199, 0.0184, 0.0265, 0.0181, 0.0176, and 0.0152, respectively. For NIRS spectral data, the R²ₚ values were determined to be 0.9712, 0.9865, 0.9826, 0.9820, 0.9851, 0.9664, and 0.9773, with RMSEP values of 0.0359, 0.0315, 0.0329, 0.0311, 0.0361, 0.0553, and 0.0300, respectively. Notably, Raman spectroscopy provided richer feature information for quantifying these compounds in column chromatography eluate, whereas NIRS spectroscopy yielded more limited discriminative bands. Furthermore, the Grad‑CAM algorithm was applied to visualize and interpret the optimal ResNet50 models, effectively identifying the most influential spectral features for each analyte. Collectively, these results demonstrate the feasibility of integrating portable Raman and NIRS spectrometers with the ResNet50 quantitative calibration model for real‑time monitoring during the production of ginkgo ketone ester tablets.
