Harnessing Artificial Intelligence for Classifying Free-Fatty Acid Particles in Biopharmaceuticals

The degradation of polysorbates in biopharmaceutical formulations most often leads to the formation of sub-visible particles (SvPs). Upon degradation, free-fatty acids (FFAs) are released which lead to the potential loss of stabilization of the product and render the product incompliant with current regulations. Flow-imaging microscopy (FIM) is a prevalent technique for enumerating and characterizing SvPs, allowing for the acquisition of image data within the size range of two to several hundred micrometers. The substantial volume of highly useful data generated by FIM often relies on expert knowledge for correct identification, paving a clear way for the necessity of rapid and accurate automated classification of these particles.

In this talk, we introduce a customized convolutional neural network (CNN) approach for the automated classification of SvP images obtained through FIM. Specifically, we focus on categorizing FFAs, proteinaceous particles, and silicon oil droplets. Our network was trained and validated on a diverse dataset, encompassing artificially pooled test samples with varying compositions, including unlabelled data of unknown composition. Our results demonstrate the efficacy of this CNN-based approach for swift and robust classification of the most common SvPs encountered during FIM analysis. Notably, our model excels in the identification of FFA SvPs, which serve as early indicators of polysorbate degradation in biopharmaceutical formulations. This presentation highlights the potential of machine learning techniques, such as CNNs, to enhance the efficiency and accuracy of SvP classification in the context of biopharmaceutical quality control, ultimately contributing to the safety and efficacy of these critical formulations.