Optimal tuning of weighted kNN- and diffusion-based methods for denoising single cell genomics data

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell genomics data analysis faces challenges like missing values, sampling issues, and noise.
• Current denoising methods often oversmooth data, losing critical information on cell identity and gene regulatory patterns.
• The k-Nearest Neighbor Graph (kNN-G) is widely used but lacks optimal hyperparameter tuning, leading to information loss.

Purpose of the Study:

• To develop a novel non-stochastic method for optimally preserving biologically relevant variance in single-cell data.
• To introduce a self-supervised framework, Denoising Expression data with a Weighted Affinity Kernel and Self-Supervision (DEWÄKSS), for tuning denoising parameters.
• To address the oversmoothing issue in kNN- and diffusion-based denoising methods.

Main Methods:

• Developed DEWÄKSS, a framework utilizing a self-supervised technique for parameter tuning.
• Investigated the tuning of kNN- and diffusion-based denoising methods.
• Employed a novel objective function to optimally preserve informative variance in single-cell data.

Main Results:

• DEWÄKSS demonstrates robustness to various preprocessing methods across established benchmarks.
• The method effectively disentangles cellular identity and maintains robust clusters across different dimension-reduction techniques.
• DEWÄKSS preserves variance along multiple expression dimensions, avoiding the oversmoothing typical of heuristic methods.
• Denoising primarily utilizes a fixed weighted kNN graph, with minimal reliance on diffusion.

Conclusions:

• DEWÄKSS offers an optimal approach to tuning kNN- and diffusion-based denoising methods for single-cell genomics.
• The framework preserves biologically relevant variance and enhances the accuracy of cell identity and clustering.
• Findings provide new insights into the mechanisms and effectiveness of kNN- and diffusion-based denoising techniques.