Conditional generative adversarial network for gene expression inference.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Gene expression profiling is crucial but expensive at the genome-wide level.
• A subset of landmark genes (∼1000) captures significant transcriptome information.
• Linear models struggle with non-linear gene regulatory network associations.

Purpose of the Study:

• To develop a cost-effective deep learning method for inferring target gene expression profiles.
• To address the limitations of linear models in capturing complex gene relationships.

Main Methods:

• Proposed a novel conditional generative adversarial network (cGAN).
• Incorporated adversarial and ℓ1-norm loss terms for improved prediction accuracy.
• Utilized landmark gene expression to infer target gene profiles.

Main Results:

• The deep learning model achieved more accurate and sharper predictions compared to traditional methods.
• Demonstrated consistent and significant improvements in two validation settings.
• Overcame the issue of smooth and blurry predictions common with mean squared error objectives.

Conclusions:

• Deep learning offers a powerful alternative for modeling complex gene interactions.
• The proposed cGAN with coupled loss functions enhances gene expression inference.
• This approach provides a more effective and cost-efficient strategy for gene expression profiling.