Testing Precision Limits of Neural Network-Based Quality Control Metrics in High-Throughput Digital Microscopy

Area of Science:

• Biopharmaceutical analysis
• Microscopy and imaging
• Machine learning for quality control

Background:

• Digital microscopy images of protein aggregates contain underutilized information beyond particle counts and size.
• Convolutional neural networks (CNNs) can identify aggregation-causing stresses but require known stress types.
• A limitation exists in identifying particles formed by unknown or uncharacterized stresses.

Purpose of the Study:

• To develop an expanded CNN analysis for detecting protein aggregates induced by unknown root-causes.
• To adapt a fingerprinting algorithm for flow imaging microscopy (FIM) data.
• To assess the robustness of the algorithm against experimental variations and noise.

Main Methods:

• Utilized a fingerprinting algorithm with ethylene tetrafluoroethylene (ETFE) microparticles as surrogates for protein aggregates.
• Performed expanded CNN analysis on flow imaging microscopy (FIM) images.
• Quantified algorithm sensitivity to microscope focus and solution refractive index, and analyzed FIM sample noise effects.

Main Results:

• The algorithm reproducibly detected complex textural features in protein aggregate images.
• These features were not easily quantifiable by standard morphological measurements.
• The approach demonstrated potential for identifying particles from unknown process upsets.

Conclusions:

• The enhanced CNN analysis successfully detects subtle particle characteristics indicative of aggregation root-causes.
• This method offers a promising tool for robust quality control in biopharmaceutical manufacturing.
• The algorithm can identify shifts in protein aggregate populations stemming from uncharacterized process variations.