Compositional data modeling of high-dimensional single cell RNA-seq (CoDA-hd): its advantages over commonly used...

Area of Science:

• Genomics
• Bioinformatics
• Statistical analysis

Background:

• Compositional data analysis (CoDA) is an emerging statistical framework applicable to various biological datasets, including single-cell RNA sequencing (scRNA-seq).
• scRNA-seq data is characterized by high dimensionality (over 20,000 components) and sparsity, posing challenges for traditional statistical methods.
• Standard normalization techniques like log-normalization may introduce artifacts in downstream analyses such as trajectory inference.

Purpose of the Study:

• To investigate the adaptability of CoDA, particularly high-dimensional CoDA (CoDA-hd), to the unique characteristics of scRNA-seq data.
• To evaluate different strategies for handling zero counts in scRNA-seq data within the CoDA framework, including prior-log-normalization, imputation, and specific count addition schemes.
• To explore the transformation of CoDA log-ratio (LR) representations into Euclidean space for compatibility with existing scRNA-seq downstream analyses.

Main Methods:

• Application of CoDA-hd and various log-ratio (LR) transformations to raw count matrices of scRNA-seq data.
• Implementation of innovative count addition schemes (e.g., SGM) to address sparsity and zero-inflation in scRNA-seq data.
• Transformation of CoDA LR representations to Euclidean space for integration with downstream analytical tools.

Main Results:

• Count addition schemes, such as SGM, facilitate the application of CoDA to high-dimensional, sparse scRNA-seq data.
• Log-normalized scRNA-seq data can be successfully transformed into the CoDA LR representation.
• CoDA LR transformations, specifically count-added centered-log-ratio (CLR), demonstrated advantages in dimension reduction visualization, improved clustering distinctness, and enhanced trajectory inference (e.g., Slingshot), while mitigating artifacts from dropouts.

Conclusions:

• CoDA offers a robust, scale-free modeling approach for scRNA-seq data analysis, particularly for downstream tasks like dimension reduction, clustering, and trajectory inference.
• An R package, 'CoDAhd', has been developed to facilitate CoDA LR transformations for high-dimensional scRNA-seq data.
• The 'CoDAhd' package and associated code are publicly available for researchers to apply these methods to their scRNA-seq datasets.