Computational modeling of cellular structures using conditional deep generative networks.

Area of Science:

• Cellular Biology
• Computational Biology
• Biomedical Imaging

Background:

• Simultaneous experimental labeling of all sub-cellular structures within a single cell is challenging.
• Understanding sub-cellular structure localization is crucial for cellular function.
• Computational models are needed to infer unknown localizations from known ones.

Purpose of the Study:

• To develop a computational model for inferring sub-cellular structure localizations.
• To address the challenge of simultaneous experimental labeling limitations.
• To leverage relationships between known and unknown structures for prediction.

Main Methods:

• Formulating the task as a conditional image generation problem.
• Utilizing conditional generative adversarial networks (cGANs) with an encoder-decoder generator.
• Introducing novel skip connections (self-gated, encoder-gated, label-gated) with gating mechanisms to control information flow.

Main Results:

• cGAN-based approaches successfully generate desired sub-cellular structures.
• Proposed methods outperform existing adversarial auto-encoder approaches.
• Novel skip connections enhance performance and ensure biologically consistent localization predictions.

Conclusions:

• The developed cGAN model effectively infers sub-cellular structure localizations.
• The novel gating mechanisms in skip connections improve prediction accuracy.
• The model's predictions align with biological experimental observations.