Foundation model cascades enable zero-shot microscopy image analysis for cell therapy manufacturing.

Area of Science:

• Biotechnology
• Process Analytical Technology (PAT)
• Artificial Intelligence in Manufacturing

Background:

• Scalable cell therapy manufacturing requires advanced process analytical technologies (PAT).
• Automated analysis of microscopy images is crucial for monitoring critical quality attributes.
• Conventional machine learning models face challenges with data labeling and generalizability.

Purpose of the Study:

• To develop a robust, zero-shot method for analyzing microscopy images in cell therapy manufacturing.
• To overcome limitations of traditional machine learning models in terms of data requirements and batch effect generalizability.

Main Methods:

• A foundational model cascade integrating a large language model (LLM) and a segment anything model was employed.
• The LLM performs initial anomaly detection, triggering alerts for anomalous images.
• For non-anomalous images, instance segmentation and LLM-based classification estimate cell counts and viability.

Main Results:

• The unified, zero-shot approach achieved robust anomaly detection.
• Quantitative measures of cell count and cell health were obtained without task-specific fine-tuning.
• The method demonstrated generalizability across different batch effects.

Conclusions:

• Combining pre-trained foundation models in a cascade offers a generalizable solution for real-time process monitoring.
• This approach facilitates feedback control for scalable and automated cell therapy manufacturing.
• The developed method supports the advancement of automated biomanufacturing processes.