Feature selection and dimension reduction for single-cell RNA-Seq based on a multinomial model.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Single-cell RNA sequencing (scRNA-Seq) measures gene expression in individual cells.
• Recent scRNA-Seq datasets utilize unique molecular identifiers (UMIs) for enhanced accuracy.
• Existing normalization methods may introduce artificial variability into scRNA-Seq data.

Purpose of the Study:

• To evaluate the statistical properties of UMI counts in scRNA-Seq data.
• To identify and address sources of false variability in scRNA-Seq analysis.
• To develop and validate improved normalization and feature selection methods for scRNA-Seq data.

Main Methods:

• Analysis of UMI counts using negative controls to determine their distribution.
• Comparison of current normalization techniques (e.g., log of counts per million) with proposed multinomial methods.
• Implementation of generalized principal component analysis (GLM-PCA) for non-normal data distributions.
• Feature selection using deviance for improved dimensionality reduction.

Main Results:

• UMI counts follow multinomial sampling without zero inflation.
• Current normalization and feature selection methods create artificial variability in dimension reduction.
• Proposed multinomial methods, including GLM-PCA and deviance-based feature selection, demonstrate superior performance.
• The new methods outperform current practices in downstream clustering assessments on ground truth datasets.

Conclusions:

• Standard scRNA-Seq normalization methods can introduce misleading variability.
• Multinomial statistical models provide a more accurate framework for UMI count data.
• Generalized principal component analysis and deviance-based feature selection are effective alternatives for scRNA-Seq data analysis.
• The proposed methods enhance the accuracy of dimension reduction and clustering in scRNA-Seq studies.