Evaluation of a Self-Supervised Machine Learning Method for Screening of Particulate Samples: A Case Study in Liquid...

Area of Science:

• Pharmaceutical analysis
• Biotechnology
• Image analysis

Background:

• Imaging is crucial for characterizing subvisible particles in pharmaceutical formulations.
• Current methods for morphological classification often rely on predefined features or supervised learning, facing challenges with complex morphologies and extensive labeling.

Purpose of the Study:

• To evaluate self-supervised contrastive learning for analyzing particle images in therapeutic solutions.
• To assess the method's ability to extract morphological information without requiring ground truth labels.

Main Methods:

• Applied a self-supervised contrastive learning approach to particle images from therapeutic solutions.
• Learned image representations by comparing particle images and their augmentations, avoiding supervised training.

Main Results:

• The self-supervised method provides a fast and implementable tool for morphological screening.
• A small subset of data is sufficient for training a convolutional neural network encoder.
• Particle classes formed distinct clusters in the encoder's embedding space, enabling tasks like weakly-supervised classification.

Conclusions:

• Self-supervised contrastive learning is effective for analyzing particle morphology in pharmaceuticals.
• This method reduces the need for time-consuming ground truth labeling.
• The approach facilitates the identification of known and novel particle subpopulations in protein solutions.