Deep learning strategy for small dataset from atomic force microscopy mechano-imaging on macrophages phenotypes

Area of Science:

• Cellular Biophysics
• Immunology
• Computational Biology

Background:

• The cytoskeleton is crucial for immune cell function, with biomechanical properties like elasticity and adhesion serving as potential biomarkers for cell activation.
• Macrophages exhibit phenotype polarization, and their biomechanical behavior can indicate early diagnostic markers.

Purpose of the Study:

• To develop a method combining atomic force microscopy (AFM) and deep neural networks (DNNs) for classifying macrophage activation states.
• To address the challenge of small datasets in AFM experiments by enhancing data through pixel-wise analysis.

Main Methods:

• Nanomechanical maps from AFM were processed into pixelated data with localization information to create an enlarged dataset.
• A DNN was trained using multimodal fusion on this dataset, with predictions derived from voting classification.
• Permutation feature importance was used to interpret the DNN's predictions and identify key biophysical properties.

Main Results:

• The DNN algorithm achieved high performance in predicting macrophage phenotypes without relying on conventional biomarkers.
• The study successfully correlated specific biophysical properties with the local density of the cytoskeleton.
• The methodology demonstrated potential for distinguishing cell states and identifying novel diagnostic markers.

Conclusions:

• The combined AFM-DNN approach offers a sensitive and automated method for analyzing cell biomechanics and activation states.
• This technique provides insights into the relationship between cytoskeleton density and cell phenotype, paving the way for innovative diagnostic tools.
• The developed methodology is adaptable for various cell systems, highlighting its broad applicability in biomedical research and diagnostics.