Compound models and Pearson residuals for single-cell RNA-seq data without UMIs

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Existing methods for single-cell RNA sequencing (scRNA-seq) data normalization often rely on Pearson residuals from Poisson or negative binomial models, primarily for Unique Molecular Identifier (UMI)-based data.
• Normalization of non-UMI scRNA-seq data presents challenges due to amplification biases and complex distributional patterns.

Purpose of the Study:

• To extend Pearson residual-based normalization methods to non-UMI scRNA-seq data.
• To develop a statistical model that accurately captures the technical noise inherent in non-UMI sequencing protocols.
• To improve gene selection and data representation (embeddings) for Smart-seq2 and similar non-UMI datasets.

Main Methods:

• Modeled the RNA amplification step using a compound distribution, combining a negative binomial distribution for captured molecules with an amplification distribution.
• Derived compound Pearson residuals from this novel model.
• Characterized amplification distributions across various sequencing protocols using a broken power law model.

Main Results:

• The compound Pearson residuals effectively enabled meaningful gene selection and generated informative embeddings for Smart-seq2 datasets.
• Demonstrated that amplification distributions in several sequencing protocols follow a broken power law.
• The developed compound model successfully accounts for previously unaddressed overdispersion and zero-inflation in non-UMI scRNA-seq data.

Conclusions:

• The proposed compound distribution model provides a robust framework for normalizing non-UMI scRNA-seq data.
• This approach enhances the accuracy and interpretability of analyses for datasets generated by protocols like Smart-seq2.
• The findings offer a more comprehensive understanding of technical noise in single-cell genomics data.