Deep Learning Algorithm-Guided Raman Spectroscopy for Real-Time Correlation of Critical Process Parameters and Titers...

Manufacturing scalability and cost efficiency of Lentiviral vector (LV) production for gene therapy are hampered by slow offline detection of critical parameters and viral titer. To enable real-time monitoring, this study proposes a process analytical technology (PAT) based on Raman spectroscopy and an innovative deep learning algorithm to achieve real-time prediction of glucose, lactate, viable cell density, and viral titer during LV production. To address the limitations of existing Raman spectral analysis methods, this study integrated data augmentation, self-supervised learning, and time-series modeling strategies. A regression model integrating a one-dimensional Convolutional Neural Network (1D-CNN) with Squeeze-and-Excitation (SE) attention, enhanced by relative positional encoding (RPE), multi-query attention (MQA), and Top-K feature selection (termed ACNN), was constructed. This approach utilized vast amounts of unlabeled spectral data to precisely extract feature information from complex components, including viral particles. Experimental results demonstrated that the ACNN model significantly outperformed the traditional Partial Least Squares (PLS) model in predicting viral titer, cell density, glucose, and lactate concentrations (paired t-test, p < 0.05 for all parameters). Notably, titer prediction error was substantially reduced. Statistical analysis showed that the RMSE of ACNN for titer prediction was significantly lower than that of PLS (t = 3.82, p < 0.01), and the R² was significantly higher (t = 4.15, p < 0.001). This study overcomes technical bottlenecks in real-time quality monitoring and process optimization for LV production, providing a feasible transition from "black box operations" to transparent and intelligent monitoring throughout the entire process and establishing a robust foundation for large-scale intelligent production of other gene therapy vectors.