Innovative methodology for flexible monitoring of various bioprocesses by using pre-existing Raman data coupled with automated transfer learning technique.

This study introduces an innovative approach for the flexible monitoring of bioprocesses using Raman spectroscopy coupled with automated transfer learning. Traditional Raman spectroscopy requires extensive process-specific calibration, limiting its transferability across different conditions. To address this, we developed an automated method that utilizes the dynamic orthogonal projection (DOP) algorithm to use pre-existing chemometric models built from one process ("Input Process") to monitor a distinct process ("Target Process") which lacked its own Raman models. These new processes conditions varied in terms of culture mode, cell line, media, analyzer, and acquisition settings. This method used spectral data from Target Process to modify the existing spectral data from Input Process, aligning them with the new conditions. The approach was validated on Chinese Hamster Ovary cell cultures, targeting critical metabolic parameters such as glucose, lactate, glutamine, and viable cell density. The results showed that with only one batch used for the transfer, the average relative errors compared to offline values were around 10% for glucose and lactate but remained high for VCD and glutamine. After a second batch to perform the transfer on two batches, the relative errors were further reduced below 10% for most parameters. By effectively transferring models across different processes, this approach minimizes the need for extensive recalibrations, enhancing the efficiency and applicability of Raman spectroscopy in diverse bioprocess environments.