scGenoByte: a GenoByte embedding transformer with biological priors for cell type annotation.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Effective cell representation learning is vital for single-cell RNA sequencing (scRNA-seq) analysis, including cell annotation and understanding cellular heterogeneity.
• Current foundation models outperform traditional methods but often discard gene information due to data sparsity and model complexity.
• Modeling the complete transcriptome for cell representation remains a significant computational challenge.

Purpose of the Study:

• To introduce scGenoByte, a unified framework for enhanced cell representation learning via biologically informed full-gene modeling.
• To address the limitations of existing methods that compromise gene information by proposing an efficient full-transcriptome modeling approach.
• To integrate biological priors for improved cell function and representation analysis.

Main Methods:

• Developed GenoBytes, biologically coherent units leveraging protein-protein interaction and gene paralogy networks for efficient full transcriptome modeling.
• Integrated protein representations with GenoByte embeddings to capture critical protein information.
• Employed pathway activity prediction as an auxiliary task for pathway-guided regularization.

Main Results:

• scGenoByte demonstrated superior performance across eight diverse scRNA-seq datasets compared to existing methods.
• The framework effectively models the full transcriptome by incorporating biological priors.
• The integration of biological information significantly enhances cell representation learning.

Conclusions:

• scGenoByte offers an effective solution for computationally challenging full transcriptome modeling in scRNA-seq.
• The study confirms the efficacy of combining full-gene context with biological priors for improved cell representation.
• The framework advances cell annotation and the deciphering of cellular heterogeneity.