Application of one-class classification using deep learning technique improves the classification of subvisible...

Area of Science:

• Biopharmaceutical Analysis
• Microscopy Techniques
• Machine Learning Applications

Background:

• Flow imaging microscopy is crucial for detecting subvisible particles in biologics, which can cause immunogenicity.
• Automated methods are needed to differentiate protein aggregates from benign contaminants like silicone oil (SO).
• One-class classification offers a potential solution for identifying stable, heterogeneous distributions, but its efficacy for subvisible particles is uncertain.

Purpose of the Study:

• To investigate the effectiveness of deep learning techniques in improving one-class classification performance for subvisible particles.
• To evaluate the classification accuracy of deep learning-enhanced one-class models for silicone oil, immunoglobulin G aggregates (Agg_IgG), and albumin aggregates (Agg_Alb).

Main Methods:

• Development of datasets using silicone oil, Agg_IgG, and Agg_Alb.
• Application of deep learning techniques to enhance one-class classification models.
• Utilizing cluster analysis to assess classification effectiveness for different particle types.

Main Results:

• Deep learning significantly improved classification scores for both Agg_IgG and Agg_Alb.
• Classification performance was more satisfactory for Agg_IgG compared to Agg_Alb.
• One-class classification combined with deep learning demonstrated excellent effectiveness across most clusters for Agg_IgG identification.

Conclusions:

• Deep learning substantially enhances the one-class classification of subvisible protein aggregates (Agg_IgG and Agg_Alb).
• This combined approach shows particular promise for the accurate evaluation of immunoglobulin G aggregates.
• Deep learning-augmented one-class classification is a valuable tool for subvisible particle analysis in biopharmaceuticals.