Interpretable representation learning for 3D multi-piece intracellular structures using point clouds.

Area of Science:

• Cell Biology
• Computational Biology
• Machine Learning

Background:

• Understanding subcellular organization requires objective quantification of intracellular structures.
• Complex multi-piece morphologies pose challenges for traditional analysis methods.
• Existing methods lack robustness and generalizability for diverse cellular structures.

Purpose of the Study:

• To develop a morphology-appropriate representation learning framework for analyzing complex intracellular structures.
• To create orientation-independent, compact, and interpretable representations of cellular morphologies.
• To enable objective, robust, and generalizable quantification of subcellular organization.

Main Methods:

• Utilized 3D rotation-invariant autoencoders and point clouds for representation learning.
• Applied the framework to punctate (e.g., DNA replication foci) and polymorphic (e.g., nucleoli) intracellular structures.
• Systematically compared the framework against image-based autoencoders using diverse datasets, including synthetic data.

Main Results:

• The framework successfully learned orientation-independent and interpretable representations of complex morphologies.
• Benchmarking demonstrated trade-offs in efficiency, generative capability, and representation expressivity.
• The approach facilitated unsupervised discovery of sub-clusters within cellular structures.
• Demonstrated application in phenotypic profiling of nucleoli following drug perturbations.

Conclusions:

• The proposed morphology-appropriate framework offers a robust and generalizable method for analyzing intracellular structures.
• This approach enhances the objective quantification of subcellular organization.
• The framework has potential applications in drug discovery and understanding cellular responses to perturbations.