Using Deep Convolutional Neural Networks to Circumvent Morphological Feature Specification when Classifying Subvisible Protein Aggregates from Micro-Flow Images

TL;DR

This paper demonstrates that deep convolutional neural networks can effectively classify subvisible protein aggregates from raw flow-imaging microscopy images, surpassing traditional morphological feature-based methods and enabling rapid, accurate process condition predictions.

Contribution

The study introduces a CNN-based supervised learning approach with an image pooling strategy for classifying protein aggregates directly from raw images, bypassing the need for predefined morphological features.

Findings

Nearly perfect prediction accuracy with as few as 20 images.

CNN approach outperforms traditional feature-based methods.

Effective in diverse protein therapeutic scenarios.

Abstract

Flow-Imaging Microscopy (FIM) is commonly used in both academia and industry to characterize subvisible particles (those $\le 25 \mu m$ in size) in protein therapeutics. Pharmaceutical companies are required to record vast volumes of FIM data on protein therapeutic products, but are only mandated under US FDA regulations (i.e., USP $\big \langle 788 \big \rangle$) to control the number of particles exceeding $10$ and $25 \mu m$ in delivered products. Hence, a vast amount of digital images are available to analyze. Current state-of-the-art methods rely on a relatively low-dimensional list of "morphological features" to characterize particles, but these methods ignore an enormous amount of information encoded in the existing large digital image repositories. Deep Convolutional Neural Networks (CNNs or "ConvNets") have demonstrated the ability to extract predictive information from raw macroscopic image data without requiring the selection or specification of "morphological features" in a variety of tasks. However, the heterogeneity, polydispersity of protein therapeutics, and optical phenomena associated with subvisible FIM particle measurements introduce new challenges regarding the application of CNNs to FIM image analysis. In this article, we demonstrate a supervised learning technique leveraging CNNs to extract information from raw images in order to predict the process conditions or stress states (freeze-thaw, mechanical shaking, etc.) that produced a variety of different protein images. We demonstrate that our new classifier (in combination with a sample "image pooling" strategy) can obtain nearly perfect predictions using as few as 20 FIM images from a given protein formulation in a variety of scenarios of relevance to protein therapeutics quality control and process monitoring.