A deep generative model of 3D single-cell organization.

Area of Science:

• Computational Biology
• Biophysics
• Image Analysis

Background:

• Accurate modeling of subcellular structures is crucial for understanding cell biology.
• Existing methods often struggle with 3D, multi-channel fluorescent image data and diverse structures.

Purpose of the Study:

• To introduce a flexible, end-to-end integrative modeling framework for 3D single-cell multi-channel fluorescent image data.
• To enable prediction of subcellular structure localization and quantification of spatial variations.

Main Methods:

• Utilized stacked conditional beta-variational autoencoders (β-VAEs) to learn latent representations of cell morphology and subcellular structure localization.
• Developed a model adaptable to various subcellular structures, sparsity levels, and reconstruction fidelities.
• Trained the model on 3D cell image data and explored design trade-offs in 2D.

Main Results:

• The trained model successfully predicts plausible locations of imaged structures in cells.
• The framework quantifies variations in subcellular structure locations by generating plausible instantiations.
• Demonstrated applicability to new data, including a drug perturbation screen, showing distinct latent representations for drugged vs. unperturbed cells.

Conclusions:

• The proposed framework offers a powerful tool for integrative modeling of complex 3D cell image data.
• The model's ability to predict and quantify subcellular structures has implications for high-throughput screening and drug discovery.
• Latent space analysis reveals expected on-target effects of drug perturbations on cellular structures.