A Transformer-Based Deep Diffusion Model for Bulk RNA-Seq Deconvolution.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Bulk RNA-seq provides average gene expression but lacks single-cell resolution, limiting cellular heterogeneity insights.
• Computational deconvolution methods aim to estimate cell-type proportions from bulk RNA-seq, but accuracy remains a challenge, particularly in complex tissues.

Purpose of the Study:

• To introduce DiffFormer, a novel deconvolution model integrating conditional diffusion and Transformer architectures.
• To evaluate DiffFormer's performance against existing methods and a baseline diffusion model.

Main Methods:

• Development of DiffFormer, a novel deconvolution model combining conditional diffusion and Transformer architectures.
• Systematic evaluation on four pseudo-bulk datasets and validation on a real-world dataset with FACS-based ground truth.

Main Results:

• DiffFormer demonstrated superior and consistent performance across all tested datasets.
• Outperformed existing deconvolution methods and a baseline MLP-based diffusion model (DiffMLP).
• Achieved a significant reduction in Root Mean Square Error (RMSE) and the highest Pearson Correlation Coefficient (PCC) on real-world data.

Conclusions:

• DiffFormer offers a high-precision, reproducible tool for cellular deconvolution.
• The Transformer architecture is identified as crucial for DiffFormer's success, demonstrating its potential for complex bioinformatics problems.