Structure-preserved dimension reduction using joint triplets sampling for multi-batch integration of single-cell...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Dimension reduction (DR) is crucial for single-cell RNA sequencing (scRNA-seq) data interpretation, visualization, and downstream analysis.
• Existing DR methods often struggle to simultaneously identify cell types, preserve data structure, and handle batch effects.
• Effective batch effect correction is a significant challenge in scRNA-seq analysis.

Purpose of the Study:

• To develop a novel structure-preserved dimension reduction (SPDR) method for scRNA-seq data.
• To create a DR method that simultaneously addresses batch effects and preserves inherent data structures.
• To improve cell type identification and data visualization in scRNA-seq.

Main Methods:

• Developed a novel structure-preserved dimension reduction (SPDR) method utilizing intra- and inter-batch triplets sampling.
• Constructed triplets incorporating mutual nearest neighbors (inter-batch), k-nearest neighbors (intra-batch), and random cells.
• Minimized a robust loss function on chosen triplets to achieve structure preservation and batch correction.

Main Results:

• SPDR demonstrated superior performance compared to existing methods (INSCT, IVIS, Trimap, Scanorama, scVI, UMAP) in batch effect removal and biological variation preservation.
• The 2D embedding generated by SPDR provides clear and authentic expression patterns, aiding in cell type determination.
• SPDR improved clustering accuracy and showed robustness to complex data characteristics and hyperparameter variations.

Conclusions:

• SPDR is a valuable tool for characterizing complex cellular heterogeneity in scRNA-seq data.
• The method effectively integrates structure preservation with batch effect correction.
• SPDR enhances the interpretability and accuracy of scRNA-seq data analysis.