(M1530-09-57) Combining Machine Learning and Backgrounded Membrane Imaging: A Case Study in Comparing and Classifying Different Types of Biopharmaceutically Relevant Particles

Purpose: The main purpose of this study was to develop and evaluate automated image analysis methods based on CNNs for the comparison and classification of images from different types of pharmaceutically relevant particles. Further, to understand the potential impact of BMI-specific sample preparation (vacuum application, drying on a membrane surface) on the particle morphology information and consequently on the possibility of distinguishing different types of particles. Finally, a fingerprint approach was evaluated to compare the particle classes/types but also improve studies on subvisible particle formation mechanisms in drug products and how this approach can allow for a quantitative comparison of similarities and differences between particle types.

Methods: In this study, we used BMI (Horizon), to generate images of subvisible particles from three different groups of samples containing a variety of particles including: protein-based drug products, synthetic model FA particles and ETFE particles. Particles in DPs were characterized after isolation on gold-coated membrane filters by micro-spectroscopic techniques and analyzed using SEM-EDX which allowed a semi-quantitative analysis of the elemental composition of particles. ATR-FTIR using a spectral range of 4000-6000 cm^-1 was performed on isolated particles to determine their organic nature. Some isolated particles were further characterized using LC-MS to determine their organic nature, both qualitatively and quantitatively. The classification and embedding analyses were performed on the BMI particle image sets without any detailed information on the type or composition of particles, or on applied experimental protocols. CNN model construction and architecture used for training the classification is outlined, as are the steps necessary to prepare images for pre-processing. The CNN classification model used an 80/20% train/test split and no images used for training were used in testing the network. Methods to create the probability density function, or fingerprints, and the goodness-of-fit hypothesis testing are discussed.

Results: The randomly selected particle images display both dramatic and subtle morphological differences in BMI particle images sampled. BMI images contain enough morphological information about the underlying particles to accurately distinguish between particles from the FAs, proteins and ETFE sample types. Further, the accuracy achievable by a standard CNN classifier using BMI data is encouraging and notable, given the particles are not in their native solution environment when analyzed by GMI. The fingerprinting approach was evaluated as an alternative to supervised CNN classification for the qualitative and quantitative comparison of particle types from the variety of conditions in this study. The fingerprint approach reveals differences in particle morphologies. The particle populations in the two different DPs with proteinaceous particles exhibited readily distinguishable fingerprints with well separated modes, indicating that the BMI images of protein particles generated by different mechanisms can be separated using the fingerprint approach. Finally, fingerprint analysis of BMI images from FA particles acquired over time revealed a meaningful change in particle morphology at 36-hours compared to all later time points. BMI particle image analysis by CNN classification or by the fingerprinting approach provides consistent results for comparable samples with respect to particle origin or age.

Conclusion: This work demonstrates that both traditional CNN classifiers and the fingerprinting approach using images from BMI contain representative morphological features capable of distinguishing various particle types studied. The CNN-based fingerprinting approach can be used to both qualitatively and quantitatively characterize morphological differences in particle populations. The fingerprinting approach also enables the characterization of the morphological similarity of BMI particle image distributions of different particle types. The work also demonstrates that extracted BMI particle images using proprietary software completement by the fingerprinting approach can be a valuable label-free method for high-throughput particle classification and characterization. Particle classification and characterization based on morphological features, in addition to particle count and size, can be utilized for the monitoring of particle formation over the different stages of product development. Finally, our findings illustrate that the fingerprinting approach can help to gain new insights in studies aiming to mimic and understand particle formation mechanisms related to polysorbate degradation in drug products.