Data-Driven Protein Purification: Machine Learning Applications in Chromatography

Sammanfattning

Protein purification is a central step in biopharmaceutical manufacturing. Modern chromatography systems generate large volumes of multivariate time-series data, yet anomaly detection remains challenging due to manual analysis and the rarity of faulty runs. This project investigates whether supervised machine learning can enable early detection of anomalies during protein purification using historical process data. A full data-processing pipeline was developed, including large-scale data extraction from Cytiva’s internal SQL databases, preprocessing of sensor signals, metadata-based labeling of alarm events, and sliding-window segmentation for model input. Using a small balanced dataset, a sanity-check experiment verified that the preprocessing and labeling pipeline performed correctly. To evaluate supervised learning under real conditions, we trained several 1D Convolutional Neural Network (1D CNN) models using two sampling strategies designed to address the severe class imbalance (0.106% positives). Although one strategy produced clear class separation during training, neither approach generalized to unseen data. Validation and test performance showed PR-AUC values close to random, and attempts to increase recall resulted in extremely high false-positive rates. These findings indicate that supervised learning based solely on metadata-derived anomalylabels is insufficient for reliable anomaly detection in this domain. Despite these limitations, the project produced a robust data-extraction framework and a functioning real-time simulation prototype, demonstrating how a future anomaly-detection model could be integrated into chromatography workflows. The overall results suggest that unsupervised or semi-supervised methods, capable of learning normal chromatographic behavior without relying on extensive labeled fault data, are likely better suited for early anomaly detection in protein purificatio