GatorSC: multi-scale cell and gene graphs with mixture-of-experts fusion for single-cell transcriptomics.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals cellular heterogeneity but faces challenges with technical noise and data sparsity.
• Existing methods often fail to fully exploit the rich, multi-scale structural information present in scRNA-seq data.
• Integrating heterogeneous graph-based views of cells and genes is crucial for robust low-dimensional representations.

Purpose of the Study:

• Introduce GatorSC, a unified representation learning framework for scRNA-seq data.
• Leverage multi-scale cell and gene graphs for enhanced information fusion.
• Develop noise-robust and structure-preserving embeddings via self-supervised learning.

Main Methods:

• GatorSC models scRNA-seq data using global cell-cell, global gene-gene, and local gene-gene graphs.
• A mixture-of-experts architecture with a gating network adaptively fuses graph neural network embeddings.
• A unified self-supervised objective couples graph reconstruction and contrastive learning for noise-robust embeddings.

Main Results:

• GatorSC consistently outperformed state-of-the-art methods on 19 diverse scRNA-seq datasets.
• Demonstrated superior performance in cell clustering, gene expression imputation, and cell-type annotation.
• Learned embeddings enabled accurate trajectory inference and recovery of biological insights in an Alzheimer's dataset.

Conclusions:

• GatorSC offers a flexible and powerful framework for comprehensive single-cell transcriptomic analysis.
• The approach effectively integrates multi-scale graph structures for robust representation learning.
• GatorSC can be extended to multi-omic and spatial transcriptomic data.