Interpretable Fine-Grained Phenotypes of Subcellular Dynamics via Unsupervised Deep Learning.

Area of Science:

• Computational Biology
• Cell Biology
• Machine Learning

Background:

• Understanding cellular heterogeneity is crucial for biological processes.
• Unsupervised machine learning faces challenges in extracting interpretable and discriminative features for live cell dynamics.

Purpose of the Study:

• To develop a self-training deep learning framework for fine-grained and interpretable cell phenotyping.
• To extract features that preserve cellular heterogeneity and discriminate between biological states.

Main Methods:

• A self-training deep learning framework with an unsupervised teacher model and a student deep neural network (DNN).
• An autoencoder-based regularizer to maximize heterogeneity associated with molecular perturbations.
• Application to analyze protrusion dynamics in migrating epithelial cells.

Main Results:

• The framework acquired features with enhanced discriminatory power, preserving molecular perturbation heterogeneity.
• Successfully delineated fine-grained phenotypes in migrating epithelial cell protrusion dynamics.
• Identified specific responses to pharmacological perturbations and linked interpretable features to temporal intervals.

Conclusions:

• The developed framework provides a valuable tool for investigating cellular dynamics and heterogeneity.
• Enables the acquisition of highly interpretable features for fine-grained phenotyping.
• Facilitates understanding of cellular responses to perturbations.