Automatic identification of relevant genes from low-dimensional embeddings of single-cell RNA-seq data

Area of Science:

• Single-cell genomics
• Computational biology
• Bioinformatics

Background:

• Dimensionality reduction is crucial for analyzing single-cell RNA sequencing (scRNA-seq) data, enabling visualization and downstream analysis.
• Nonlinear methods are preferred for complex scRNA-seq data, but identifying genes that drive cell positions in low-dimensional embeddings remains challenging.
• Current methods lack a way to link gene expression to specific cell locations in non-linear embeddings, hindering biological interpretation.

Purpose of the Study:

• To develop a method for identifying genes that drive cell positions in non-linear low-dimensional embeddings of scRNA-seq data.
• To provide tools for characterizing biological processes by understanding gene contributions to cell embeddings.
• To enable the extraction of gene identities crucial for cell positioning in complex datasets.

Main Methods:

• Introduced concepts of local and global gene relevance to compute principal component analysis (PCA) loading equivalents for non-linear embeddings.
• Global gene relevance identifies drivers of the overall embedding structure.
• Local gene relevance pinpoints drivers within specific sub-regions of the embedding.

Main Results:

• Successfully applied the method to diverse scRNA-seq datasets and embedding techniques.
• Demonstrated the versatility of the approach in identifying key genes across various biological processes.
• The method effectively links gene expression patterns to cell positions in low-dimensional space.

Conclusions:

• The developed gene relevance method provides a powerful tool for interpreting non-linear dimensionality reduction in scRNA-seq.
• This approach facilitates the identification of biologically relevant genes driving cellular heterogeneity.
• The method is implemented in the popular R package destiny 3.0 for widespread accessibility and reproducibility.