High-Fidelity Transcriptome Reconstruction of Degraded RNA-Seq Samples Using Denoising Diffusion Models.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• RNA degradation in archived samples causes systematic biases in RNA-sequencing (RNA-seq) data.
• This limits the accuracy of downstream analyses, necessitating robust computational solutions.
• High-fidelity transcriptome restoration is critical for reliable biological interpretation.

Purpose of the Study:

• To develop a computational method for high-fidelity transcriptome restoration from degraded RNA-seq data.
• To introduce DiffRepairer, a deep learning model designed to reverse RNA degradation effects.
• To validate the efficacy of DiffRepairer in restoring biologically meaningful signals.

Main Methods:

• Introduced DiffRepairer, a deep learning model integrating Transformer architecture and a conditional diffusion model.
• Trained the model on simulated "degraded-original" paired data for one-step repair mapping.
• Utilized a comprehensive simulation pipeline to generate training data.

Main Results:

• DiffRepairer demonstrated stable and superior performance across five diverse pseudo-degraded datasets.
• Outperformed traditional statistical methods (e.g., CQN) and standard deep learning models (e.g., VAE).
• Achieved better results on key technical and biological metrics for transcriptome repair.

Conclusions:

• DiffRepairer is a validated, high-precision tool for transcriptome repair.
• Effectively restores biologically meaningful signals from degraded RNA-seq data.
• Highlights the potential of advanced generative models in bioinformatics for data restoration.