(672c) Machine Learning Analysis of Protein Aggregates Formed in Container-Closure Systems

Therapeutic protein formulations are exposed to numerous stresses during manufacturing, shipping, storage, and administration that promote aggregation. These different stresses result in aggregates whose morphologies are characteristic of the stress at their root cause. This morphology information can be captured by using flow imaging microscopy (FIM) to image the aggregates in a sample and applying machine learning techniques such as convolutional neural networks (ConvNets) to extract the morphological information embedded in FIM images.

In this study, we describe and demonstrate a novel algorithm that can be used to determine if two protein formulations contain similar particle morphologies and thus were exposed to the same root cause stresses. An algorithm was developed to compare collections of FIM images using ConvNets trained using techniques borrowed from facial recognition and traditional statistical techniques such as nonparametric density estimation and goodness-of-fit hypothesis testing. This algorithm provides a robust statistical method to determine if two samples exhibit significantly different FIM datasets and thus contain particles from different root cause stresses. As a demonstration, this algorithm was used to compare intravenous immunoglobulin (IVIg) formulations that were exposed to freeze-thaw and shaking stresses within several container-closure systems. The algorithm was capable of distinguishing between the populations of IVIg aggregates produced by these stresses in subtly different container-closure systems (e.g. glass vials from different manufacturers). This analysis also revealed that the particle morphologies produced a result of shaking stress were more impacted by the container-closure system than those produced by freeze-thaw cycling.