Leveraging Transfer Learning to Maximize Data Reusability in Protein A Chromatography Resin Screening

Abstract

Early identification of optimal chromatography resins is essential but challenging due to extensive experimental data requirements and process complexities. This study explores transfer learning as a method to maximize data reuse, enhancing efficiency in resin screening and design space identification. Two transfer learning scenarios were analysed: model-to-data, leveraging synthetic data from mechanistic models, and data-to-data, utilizing existing experimental data. Both methods were benchmarked against conventional data-driven approaches under conditions of varying data scarcity. Employing artificial neural networks, we demonstrate that transfer learning substantially improves prediction accuracy and reduces the experimental workload by up to 60% compared to purely data-driven methods. Our findings reveal that model-to-data transfer learning achieves superior accuracy, especially in scenarios with severely limited experimental data, whereas data-to-data transfer learning remains highly effective under moderate data constraints. Notably, model-to-data transfer learning maintained robust prediction performance even when the experimental dataset was highly restricted, highlighting its practical value under realistic process development conditions.