scHashFormer: a hash-driven graph transformer for scalable scRNA-seq clustering

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables detailed analysis of cellular heterogeneity.
• Clustering scRNA-seq data is crucial for identifying cell types and states.
• Existing Transformer-based methods face challenges with scalable and biologically meaningful tokenization.

Purpose of the Study:

• To introduce scHashFormer, a novel hash-driven tokenization mechanism for scRNA-seq data.
• To address the limitations of existing tokenization methods in terms of scalability and biological relevance.
• To improve the effectiveness of cell clustering and downstream analysis in scRNA-seq data.

Main Methods:

• Developed a novel hash encoder with a learnable hash window size.
• Employed self-supervised learning to group similar cells with identical hash codes.
• Constructed token sequences from hash buckets for information aggregation.
• Utilized Transformer architecture for representation learning and embedding generation.

Main Results:

• scHashFormer demonstrated competitive clustering effectiveness across multiple scRNA-seq datasets.
• The method achieved significant scalability for large-scale scRNA-seq data analysis.
• Generated embeddings from scHashFormer improved performance in downstream tasks like trajectory preservation and differential gene expression analysis.

Conclusions:

• scHashFormer offers an effective and scalable solution for scRNA-seq data analysis.
• The hash-driven tokenization approach enhances cell clustering and biological interpretation.
• This method advances the application of foundation models in single-cell transcriptomics.