Automatic Identification of the Stress Sources of Protein Aggregates Using Flow Imaging Microscopy Images

Abstract

A novel approach to identify 5 types of simulated stresses that induce protein aggregation in prefilled syringeetype biopharmaceuticals was developed. Principal components analyses of texture metrics extracted from flow imaging microscopy images were used to define subgroups of particles. Supervised machine learning methods, including convolutional neural networks, were used to train classifiers to identify subgroup membership of constituent particles to generate distribution profiles. The applicability of the stress-specific signatures for distinguishing stress source types was verified. The high classification efficiencies (100%) precipitated the collection of data from more than 20 independent experiments to train support vector machines, k-nearest neighbors, and ensemble classifiers. The performances of the trained classifiers were validated. High classification efficiencies for friability (80%-100%) and heating at 90 C (85%-100%) are indicative of high reliability of these methods for stress-stability assays while extreme variations in freeze-thawing (2%-100%) and heating at 60 C (2.25%-98.25%) indicate the unpredictability of particle composition profiles for these forced degradation conditions. We also developed subvisible particle classifiers using convolutional neural network to automatically identify silicone oil droplets, air bubbles, and protein aggregates. The developed classifiers will contribute to mitigating aggregation in biopharmaceuticals via the identification of stress sources.

Advanced Characterization of Silicone Oil Droplets in Protein Therapeutics Using Artificial Intelligence Analysis of Imaging Flow Cytometry Data

“…CNN have previously been used in the classification of protein from silicone oil particles in protein formulations using FM data 26,27 ; however, these models have shown limited accuracy when testing samples not used in the training data sets. As an example, near-perfect classification accuracy was obtained for test images in protein-silicone mixtures when images from the same experimental sample were used for training, however inconsistency was evident for fully independent protein-silicone mixtures not used during training.…”

“…Deep learning using convolutional neural networks (CNNs), which specialize in pattern recognition using an architecture mimicking the human brain, show great promise for nonparametric classification of images across many industries, and have been applied for classification of subvisible particles in therapeutic formulations measured by FM. 26,27 These reports demonstrated accurate classification of particles as small as 1 mm when training and validation are performed on the same test samples, however classification accuracy was inconsistent for independent samples not used for training.…”

Machine learning and statistical analyses for extracting and characterizing “fingerprints” of antibody aggregation at container interfaces from flow microscopy images

“…This approach is a departure from previous techniques used to predict to which of a small set of conditions a sample was exposed (Calderon et al, 2018;Gambe-Gilbuena et al, 2020). The primary advantage of this new approach is its ability to determine, using only a small number of FIM images, if a new sample exhibits significantly different particle populations than those found under baseline conditions.…”

“…Convolutional neural networks (ConvNets) can be used to extract and analyze morphological information embedded in FIM images (Calderon, Daniels, & Randolph, 2018;Gambe-Gilbuena, Shibano, Krayukhina, Torisu, & Uchiyama, 2020). ConvNets are a family of neural networks capable of learning relevant features from a collection of images that are useful when performing tasks such as classification and dimension reduction (Calderon et al, 2018;Esteva et al, 2017;Krizhevsky, Sutskever, & Hinton, 2012;Schroff, Kalenichenko, & Philbin, 2015).…”

“…ConvNets trained on FIM data sets can accurately classify protein aggregates produced by different stresses. In Calderon et al (2018) and Gambe-Gilbuena et al (2020), a set of single, well-defined stresses (e.g., freeze-thawing and heating) was applied to protein solutions, causing aggregates to form.…”