In silico generation of gene expression profiles using diffusion models.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• RNA sequencing (RNA-seq) data is crucial for precision medicine, particularly in cancer research.
• Deep learning models analyze complex gene expression data but require large datasets, which are often scarce in transcriptomics.
• Synthetic data generation is explored to overcome data limitations in transcriptomics analysis.

Purpose of the Study:

• To adapt state-of-the-art diffusion models (DDPM and DDIM) for synthetic transcriptomics data generation.
• To evaluate the performance of diffusion models against existing generative models like VAEs and GANs.
• To investigate reconstruction methods for enhancing the quality and utility of generated transcriptomic data.

Main Methods:

• Implementation and adaptation of two diffusion models (DDPM and DDIM) for transcriptomics data.
• Generation of synthetic L1000 landmark gene expression data.
• Comparison of linear and nonlinear reconstruction methods to recover complete transcriptomes.
• Evaluation of generated data using data quality indicators and predictive performance on TCGA and GTEx datasets.

Main Results:

• Diffusion models demonstrate strong performance in generating synthetic transcriptomics data.
• DM-generated data for L1000 landmark genes exhibit enhanced predictive capabilities compared to TCGA and GTEx datasets.
• Reconstruction methods, both linear and nonlinear, significantly improve the performance of diffusion models and other generative models.

Conclusions:

• Diffusion models are highly effective and promising for generating synthetic transcriptomics data.
• The proposed diffusion model pipeline addresses the challenge of data scarcity in transcriptomics for precision medicine.
• Diffusion models represent a significant advancement in generative modeling for gene expression data analysis.