scDeepCluster

scDeepCluster performs deep generative clustering of single-cell RNA sequencing (scRNA-seq) data to jointly learn low-dimensional latent representations and cluster assignments while explicitly modeling the scRNA-seq data-generation process and accounting for dropout-induced inflation of zero counts.

Key Features:

• Deep generative clustering framework: Implements a deep generative model that integrates clustering into representation learning for scRNA-seq data.
• Joint representation and clustering: Simultaneously learns low-dimensional feature representations and cluster assignments.
• Probabilistic modeling of counts: Explicitly models count distributions to address dropout and zero-inflation in scRNA-seq data.
• Deep embedded clustering integration: Combines probabilistic count modeling with deep embedded clustering techniques.
• Model-based latent representation: Produces latent representations that preserve biologically meaningful structure while mitigating sparse, noisy observations.
• Iterative optimization: Couples representation learning with clustering refinement through iterative optimization.
• Reduced preprocessing dependence: Lowers reliance on ad hoc preprocessing by modeling the data-generation process directly.
• Benchmark evaluation: Evaluated on simulated and empirical datasets generated by four major single-cell sequencing platforms.
• Performance and scalability: Demonstrates improved clustering accuracy, robustness, and computational scalability with runtime scaling linearly with cell count.

Scientific Applications:

• Cell heterogeneity characterization: Characterizing cell heterogeneity in large-scale single-cell transcriptomic studies using scRNA-seq data.
• Cell-type clustering and identification: Clustering cells and identifying cell types from sparse, dropout-affected scRNA-seq measurements.
• Cross-platform benchmarking: Comparative evaluation of clustering performance on simulated and empirical datasets from multiple single-cell sequencing platforms.

Methodology:

Uses a deep generative model with probabilistic modeling of count distributions and deep embedded clustering, and employs iterative optimization to jointly learn latent representations and cluster assignments.