New algorithms for unsupervised cell clustering from scRNA-seq data.

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for cell type identification.
• Unsupervised clustering of scRNA-seq data faces challenges like high dimensionality, data sparsity, and technical noise.
• Existing methods struggle with the complexity of scRNA-seq datasets.

Purpose of the Study:

• Introduce novel algorithms for robust unsupervised clustering of scRNA-seq data.
• Address the limitations of current methods in handling high-dimensional and sparse scRNA-seq data.
• Evaluate the performance of new algorithms against existing solutions.

Main Methods:

• Developed a k-Minimum Spanning Tree (k-MST) graph algorithm using iterative subgraph analysis without dimensionality reduction.
• Applied the Louvain algorithm for cell clustering on the constructed k-MST graph.
• Explored a neural network-based approach using an autoencoder to learn Gaussian mixture model parameters.

Main Results:

• Benchmark experiments demonstrate competitive accuracy of the proposed algorithms compared to previous methods.
• Algorithm performance is influenced by sequencing depth, cell count, and tissue type.
• The autoencoder model showed superior accuracy on a refractory epilepsy scRNA-seq dataset, while k-MST performed well among graph-based methods.

Conclusions:

• The novel k-MST and autoencoder algorithms provide effective solutions for scRNA-seq data clustering.
• These methods offer viable alternatives for accurate cell type identification in complex biological datasets.
• Further research can optimize these algorithms for specific sequencing depths, cell numbers, and tissue types.