Learning discriminative and structural samples for rare cell types with deep generative model.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for identifying cell types as biomarkers for disease diagnosis and therapy.
• Identifying rare cell types in scRNA-seq data presents significant challenges, including the few-shot problem, lack of interpretability, and difficulties in sample generation and cell clustering.

Purpose of the Study:

• To address the challenges in identifying rare cell types within scRNA-seq data.
• To propose a novel deep generative model, scLDS2, for enhanced cell type identification, particularly for rare subpopulations.

Main Methods:

• Developed scLDS2, a deep generative model utilizing adversarial learning to estimate cell distributions and discriminate rare from non-rare cell types.
• Incorporated $\ell _1$-norm for generating sparse samples to enhance interpretability and learn cell relations.
• Employed nuclear-norm to learn block structures in generated samples, grouping similar cells for type identification.
• Integrated sample generation, classification, and feature extraction into a unified generative framework.

Main Results:

• scLDS2 demonstrated superior performance across 20 datasets compared to 17 state-of-the-art methods.
• Achieved an average improvement of 25.12% in Adjusted Rand Index, indicating enhanced accuracy in cell type identification.
• Effectively transformed the rare cell type detection problem into a classification task within a joint learning framework.

Conclusions:

• scLDS2 offers an effective strategy for identifying rare cell types in scRNA-seq data.
• The model's unified framework and generative approach significantly improve the accuracy and interpretability of cell type identification.
• The freely available Python software facilitates broader application in genomic research.