Information-Distilled Generative Label-Free Morphological Profiling Encodes Cellular Heterogeneity.

Area of Science:

• Biophysics
• Computational Biology
• Cell Biology

Background:

• Image-based cytometry struggles with batch variations, hindering accurate cell morphology interpretation.
• Current solutions for batch correction are limited, especially for complex cell image data.

Purpose of the Study:

• Introduce Cyto-Morphology Adversarial Distillation (CytoMAD) for self-supervised learning of cell morphology.
• Enable integrated analysis of multi-batch cell image data without extensive annotations or control samples.
• Enhance biophysical imaging cytometry for label-free analysis.

Main Methods:

• CytoMAD employs a multi-task learning strategy to distill morphological information from batch variations.
• Incorporates "morphology distillation" paired with deep-learning image-contrast translation.
• Utilizes self-supervised learning to minimize reliance on pre-existing data or manual labels.

Main Results:

• CytoMAD enables label-free classification of human lung cancer cell types.
• Accurately predicts progressive drug responses without drug concentration data.
• Reveals tumor biophysical cellular heterogeneity linked to epithelial-mesenchymal plasticity, missed by standard markers.

Conclusions:

• CytoMAD effectively integrates multi-batch cell image data for robust morphological analysis.
• Facilitates label-free classification and drug response assessment in cancer research.
• Supports wider adoption of biophysical cytometry for cost-effective diagnosis and screening.