Dimensionality reduction is vital for analyzing large-scale single-cell RNA-seq data. While Gaussian Process Latent Variable Models (GPLVMs) are interpretable, their scalability struggles with effective cell type clustering. To address this, we propose the amortized stochastic variational Bayesian GPLVM (BGPLVM), optimized for single-cell RNA-seq through specialized design strategies. The model illustrates comparable performance to the scVI method across synthetic and real datasets, particularly excelling in accounting for biological influences such as cell-cycle and batch variations, thus revealing clearer latent structures in the data.
Current scalable Gaussian Process Latent Variable Models (GPLVMs) are not effective for clustering cell types, prompting the introduction of the improved BGPLVM model designed specifically for single-cell RNA-seq analysis.
Our results show that the modified BGPLVM model achieves significant improvements over the standard Bayesian GPLVM and performs comparably to the leading scVI approach on various datasets.
Collection
[
|
...
]