Accurate glucose monitoring has been a bottleneck in fed-batch fermentation for years, and is especially critical as biopharmaceutical manufacturers transition toward hybrid and continuous bioprocessing for advanced therapeutics.
Fusing Raman and near-infrared (NIR) spectroscopy into a multimodal soft-sensing framework appears, however, to remove that bottleneck. This new framework provides high-frequency, low-latency measurements that aren’t typically available with existing commercial, off-line sensors. Its developers say it provides “accurate, online-compatible glucose soft sensing,” and that it is also amenable for in-line applications.
More specifically, Raman spectroscopy offers crisp, highly specific chemical fingerprints, while NIR offers broad, information-laden spectral coverage. Realizing this, a team including senior authors Xiwei Tian, DEng, of East China University of Science and Technology in Shanghai and Yinlan Ruan, PhD, of Guilin University of Electronic Technology, encoded each modality separately and then merged them using dual-stream, cross-attention transformer architecture (DS-CAT).
The team tested their system using 142 preprocessed, synchronized Raman-NIR records.
Writing in Bioprocess and Biosystems Engineering, they report that their model explained 99% of the variation in true glucose concentration. Anything above 95% is considered good. Likewise, the root mean square error of 1.09 g/L was well within the level needed to ensure effective feed control. Furthermore, the mean absolute error of 0.89 g/L achieved by this model suggests it is consistently accurate. Such consistency is particularly important for real-time processing controls, where outliers could negatively affect feeding strategies. Results remained similar throughout five-fold validation, the scientists report.
Additionally, “Prediction-based perturbation…revealed strong dependence on Raman information, while Raman-block sensitivity analysis identified 1122−1335 cm-1 [which is associated with carbohydrate vibrational modes] as the interval showing the strongest prediction sensitivity,” Tien, Ruan, and colleagues point out. Those findings are comparable to other literature reports, while demonstrating the enhanced sensitivity, reduced noise, and reduced drift of this new framework.
Notably, the researchers note that the DS-CAT framework “outperform[ed] partial least squares regression, gradient boosting regression trees, convolutional neural networks, and temporal convolutional networks under the same evaluation protocol.” By performing better than these four industry-standard modeling approaches, DS-CAT has the potential to substantially improve soft-sensor capabilities for real-time glucose monitoring.
These findings suggest the DS-CAT approach to glucose sensing may enable reliable insights that lead to better control of a critical parameter in closed-loop mammalian and microbial fermentation. If so, the high-frequency and low-latency data measurements this method delivers may become a key element in a rational transition from fed-batch fermentation toward continuous manufacturing.