Evaluating the performance of dropout imputation and clustering methods for single-cell RNA sequencing data

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables transcriptome analysis at single-cell resolution.
• Cell clustering is vital for identifying cell subtypes and inferring lineage from scRNA-seq data.
• Technical noise leads to zero counts (dropout), complicating accurate clustering.

Purpose of the Study:

• To evaluate the impact of dropout imputation methods on scRNA-seq clustering performance.
• To identify optimal combinations of imputation and clustering algorithms for scRNA-seq data.
• To assess how dataset size influences the performance of imputation-clustering strategies.

Main Methods:

• Assessed nine dropout imputation methods combined with eight clustering algorithms.
• Utilized 10 well-annotated scRNA-seq datasets of varying sample sizes.
• Evaluated clustering performance and data visualization quality (t-SNE).

Main Results:

• Imputation methods generally enhance clustering performance and data visualization quality.
• Optimal imputation-clustering combinations are dataset-size dependent.
• Single-cell analysis via expression recovery (imputation) + Sparse Subspace Clustering (SSC) performs well on smaller datasets.
• Adaptively-thresholded low-rank approximation (imputation) + single-cell interpretation via multikernel learning (SIMLR) excels on larger datasets.

Conclusions:

• Dropout imputation is beneficial for improving scRNA-seq clustering accuracy.
• Algorithm selection for scRNA-seq analysis should consider dataset size.
• Specific imputation-clustering pairs offer superior performance depending on data scale.