Supervised application of internal validation measures to benchmark dimensionality reduction methods in scRNA-seq...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) generates high-dimensional data, necessitating dimensionality reduction for analysis.
• Existing benchmarks for scRNA-seq dimensionality reduction methods are limited and often rely on downstream analysis accuracy.
• A comprehensive evaluation of these methods is crucial for establishing best practices in scRNA-seq data analysis.

Purpose of the Study:

• To conduct the most comprehensive benchmark of dimensionality reduction methods for scRNA-seq data to date.
• To evaluate method performance using internal validation measures (IVMs) rather than downstream analysis accuracy.
• To assess the preservation of global data structure, computational efficiency, and clustering performance.

Main Methods:

• Utilized over 300,000 compute hours to assess 33 dimensionality reduction methods and 25,000+ embeddings across 55 scRNA-seq datasets.
• Repurposed internal validation measures (IVMs) to assess the quality of biological clusters post-dimensionality reduction.
• Evaluated performance using extensive DBSCAN clustering iterations and assessed global structure preservation and computational resource requirements.

Main Results:

• Internal validation measures (IVMs) effectively assess the formation of biological clusters after dimensionality reduction.
• Hyperparameter optimization using IVMs as an objective function leads to near-optimal clustering performance.
• Latent Dirichlet Allocation (LDA) and Potential of Heat-diffusion for Affinity-based Transition Embedding (PHATE) were identified as high-performing algorithms.

Conclusions:

• This benchmark provides a valuable resource for researchers selecting dimensionality reduction techniques for scRNA-seq data.
• The study guides best practices for dimensionality reduction in scRNA-seq analysis pipelines.
• LDA and PHATE are recommended for their superior performance in preserving biological structures and facilitating accurate clustering.