scCCESS

scCCESS implements an autoencoder-based cluster ensemble framework to improve cell type-specific clustering of single-cell RNA-sequencing (scRNA-seq) data.

Key Features:

• Random Subspace Projections: Generates random subspace projections from the original high-dimensional gene expression matrix to reduce dimensionality by selecting subsets of features and mitigate noise and redundancy.
• Autoencoder Compression: Compresses each projected subspace into low-dimensional representations using an autoencoder, an unsupervised artificial neural network that learns efficient data codings.
• Ensemble Clustering: Integrates clustering results across multiple encoded datasets to enhance robustness and accuracy of cell type-specific clusters.
• Compatibility with clustering algorithms: Applies ensemble outputs in conjunction with standard k-means clustering and the kernel-based SIMLR algorithm tailored for scRNA-seq data.
• Quantitative evaluation: Performance is assessed using four distinct evaluation metrics, with reported improvements up to 100% for some metrics.

Scientific Applications:

• Cell type identification: Improves identification of distinct cell types from scRNA-seq transcriptomes.
• Downstream single-cell analyses: Supports downstream analyses that depend on accurate clustering of single-cell transcriptomic data.
• Characterization of cellular heterogeneity: Facilitates detection and characterization of cellular heterogeneity within complex tissues.
• Benchmarking clustering methods: Enables comparative evaluation of clustering approaches for scRNA-seq data using multiple metrics and algorithm comparisons.

Methodology:

The pipeline performs random subspace projections of the gene expression matrix, compresses each projection with an autoencoder to obtain low-dimensional encodings, applies ensemble clustering across encoded datasets, and evaluates results with four distinct metrics including comparisons to k-means and SIMLR.