DepthDiff: Restoring Low-Depth Single-Cell RNA-Seq Signals via Diffusion Denoising

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Low sequencing depth in single-cell RNA sequencing (scRNA-seq) leads to molecular capture loss and zero inflation.
• This data sparsity compromises the reliability of downstream analyses, hindering biological insights.

Purpose of the Study:

• To develop an effective computational method for enhancing scRNA-seq data with low sequencing depth.
• To improve the accuracy of gene expression reconstruction and preserve biological signals in sparse scRNA-seq datasets.

Main Methods:

• Proposed DepthDiff, a depth-conditional expression enhancement model utilizing diffusion-based denoising.
• Trained the model using low-depth expression profiles and sequencing depth ratio as conditions to learn a supervised residual mapping to high-depth references.
• Employed fixed-UMI downsampling to create benchmarks across three public scRNA-seq datasets.

Main Results:

• DepthDiff demonstrated superior performance compared to supervised baselines, MAGIC, and unenhanced low-depth data in expression reconstruction and biological signal preservation.
• Ablation studies identified x0 prediction and cosine noise scheduling as critical for model stability.
• Cross-dataset transferability and CITE-seq validation confirmed the generalizability and biological relevance of the enhanced signals.

Conclusions:

• DepthDiff provides an efficient supervised framework for enhancing low-depth scRNA-seq data.
• The primary benefits stem from diffusion-based denoising training, not generative reverse sampling, offering a computationally efficient solution.