GEMDiff: a diffusion workflow bridges between normal and tumor gene expression states: a breast cancer case study

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Breast cancer complexity stems from genetic/epigenetic mutations, challenging traditional machine learning for drug design.
• Existing generative models struggle with scalability and stability for large-scale gene expression data.

Purpose of the Study:

• Introduce GEMDiff, a diffusion model-based workflow to bridge normal and tumor gene expression states.
• Enhance biomarker identification and improve machine learning model performance through data augmentation and perturbation simulation.

Main Methods:

• Utilized a diffusion model (GEMDiff) to augment RNA-seq data and simulate gene expression perturbations.
• Developed a scalable and stable workflow for analyzing large-scale gene expression datasets.
• Avoided task-specific hyper-parameter tuning and specialized loss functions for broad applicability.

Main Results:

• GEMDiff effectively handles large-scale gene expression data without scalability or stability issues.
• Demonstrated utility in a breast cancer case study, identifying 307 core genes in tumor-to-normal state transitions.
• Generated synthetic data to address limitations in biological data availability for machine learning.

Conclusions:

• GEMDiff offers a robust and generalizable tool for gene expression analysis and biomarker discovery.
• The workflow facilitates the development of improved predictive models by augmenting biological data.
• Open-source availability promotes wider adoption and advancement in computational cancer research.