(T0930-11-60) Rapid Total Sialic Acid Monitoring During Cell Culture Process Using a Machine Learning Model Based Soft Sensor

Purpose: Total sialic acid content (TSA) in biotherapeutic proteins is often a critical quality attribute as it impacts the drug efficacy. Traditional wet chemical assays to quantify TSA in biotherapeutic proteins during cell culture typically takes several hours or longer due to the complexity of the assay which involves isolation of sialic acid from the protein of interest, followed by sample preparation and chromatographic based separation for analysis¹. Those assays are lengthy and do not allow for the quick parallel analysis of multiple samples. In contrast, soft sensors developed from process models can integrate and analyze process data in real time, thus providing fast turnaround TSA content measurements^2,3. Here, we developed a machine learning model-based soft sensor to rapidly predict TSA during cell culture by using typically measured process parameters.

Methods: Our study aimed at developing a soft sensor for rapid (in minutes) total sialic acid prediction in therapeutic proteins during cell culture based on readily available bioreactor process parameters such as pH, DO, nutrients and metabolites. The raw dataset was downloaded from a database which connected to a PI server for the bioreactor parameters and daily offline measurements, and then subjected to our modeling strategy. Four major steps were included: 1. Data cleaning, where we evaluated raw data for outliers and trends, and handled missing data. 2. Data pre-processing, where we transformed raw data into a format that minimizes the impact of noise and outliers while preserving the information content. 3. Model method selection, where we compared several model types. 4. Results interpretation, where we gained process understanding from model results.

Results: Multiple machine learning algorithms were assessed on our dataset, with the Random Forest model being identified as the most promising model. With random partitioning the dataset to calibration and test sets, different models were trained and tested for the TSA values on three modeled days (8, 10, and 14). The model statistics are shown in Table 1. As the observational period progressed from 8 days to 14 days, the models generally depicted an enhancement in both coherence and predictability. Feature importance analysis revealed that attributes like viable cell density (VCD), glutamate, ammonium, phosphate, and basal medium type are critical for predictions.

Conclusion: This research emphasizes the establishment of machine learning model based soft sensors, enabling a rapid and efficient tool for sialic acid prediction. Future endeavors may focus on data augmentation to further enhance model precision and exploration of process control capabilities.

References: 1. Markely, L. R. A.; Ong, B. T.; Hoi, K. M.; Teo, G.; Lu, M. Y.; Wang, D. I. C. A High-Throughput Method for Quantification of Glycoprotein Sialylation. Anal. Biochem. 2010, 407 (1), 128–133.
2. Nikita, S.; Thakur, G.; Jesubalan, N. G.; Kulkarni, A.; Yezhuvath, V. B.; Rathore, A. S. AI-ML Applications in Bioprocessing: ML as an Enabler of Real Time Quality Prediction in Continuous Manufacturing of mAbs. Comput. Chem. Eng. 2022, 164, 107896.
3. Luttmann, R.; Bracewell, D. G.; Cornelissen, G.; Gernaey, K. V.; Glassey, J.; Hass, V. C.; Kaiser, C.; Preusse, C.; Striedner, G.; Mandenius, C.-F. Soft Sensors in Bioprocessing: A Status Report and Recommendations. Biotechnol. J. 2012, 7 (8), 1040–1048.