Regularized Single-Cell Imaging Enables Generalizable AI Models for Stain-Free Cell Viability Screening

Area of Science:

• Biomedical research
• Drug development
• Cell biology

Background:

• Cell viability assays are crucial for research and drug discovery.
• Current artificial intelligence (AI) models for predicting cell viability from images lack broad applicability.
• Generalizability issues hinder AI adoption in diverse cell-based assays.

Purpose of the Study:

• To develop a novel strategy for training generalizable AI models for cell viability assessment.
• To improve the accuracy and robustness of AI-driven cell viability prediction using microscopy.
• To enable label-free, non-destructive cell screening workflows.

Main Methods:

• Introduced "regularized imaging" by isolating single cells in nanowells for standardized image acquisition.
• Generated training data using brightfield microscopy images of live and dead cells under cytotoxic conditions.
• Trained an AI model on a limited dataset and tested its generalizability across different cell types and unseen compounds.

Main Results:

• The AI model accurately predicted cell viability for unseen compounds, matching fluorescence assay results.
• Demonstrated effective generalization across diverse cell types, including adherent and suspension cells.
• Enabled non-destructive, kinetic cell viability studies, distinguishing compound action speeds.

Conclusions:

• Regularized single-cell imaging significantly enhances AI model generalizability for cell viability prediction.
• This method provides a robust, label-free approach for cell screening.
• The developed AI models can recognize key cellular features for improved biological insights.