Integrated Cell 3D

We introduce a framework for end-to-end integrative modeling of 3D single-cell multi-channel fluorescent image data of diverse subcellular structures. We employ stacked conditional β-variational autoencoders to first learn a latent representation of cell morphology, and then learn a latent representation of subcellular structure localization which is conditioned on the learned cell morphology.

Overview

The Integrated Cell 3D (IC3D) project develops deep generative models to understand and predict the spatial organization of subcellular structures in 3D. By learning a hierarchical latent representation of cell shape and organelle localization, we can:

• Generate realistic synthetic cells with biologically plausible organelle arrangements
• Understand the dependencies between cell morphology and organelle positioning
• Predict organelle localization patterns from cell shape alone
• Explore the statistical landscape of cellular organization

Key Features

Hierarchical Architecture: Our model uses stacked conditional variational autoencoders (cVAEs) to capture the hierarchical nature of cellular organization:

1. First level: Learn cell morphology representation
2. Second level: Learn organelle localization conditioned on cell shape

Multi-channel 3D Imaging: The model handles complex 3D fluorescence microscopy data with multiple channels representing different subcellular structures.

Interpretable Latent Space: The learned latent representations are smooth and interpretable, enabling exploration of cellular phenotype space.

Methods

We use β-variational autoencoders (β-VAEs) which balance reconstruction accuracy with latent space disentanglement. The conditioning mechanism ensures that organelle predictions respect the underlying cell morphology.

Key technical components:

• 3D convolutional encoder/decoder networks
• Conditional probabilistic modeling
• Spherical harmonic shape representations
• Multi-scale feature extraction

Results

The model successfully learns to:

• Reconstruct diverse cell morphologies with high fidelity
• Generate realistic organelle localization patterns
• Capture statistical relationships between cell shape and organelle positioning
• Generalize to new cells not seen during training

Publication

Paper: A deep generative model of 3D single-cell organization

PLOS Computational Biology, 2021

Impact

This work provides a computational framework for understanding the principles governing cellular organization in 3D space. It enables researchers to:

• Test hypotheses about structure-function relationships
• Generate synthetic training data for other analyses
• Identify abnormal cellular phenotypes
• Guide experimental design

The approach has been extended to study how cellular organization changes during differentiation and in response to perturbations.