Cell type matching across species using protein embeddings and transfer learning.

Area of Science:

• Comparative genomics
• Bioinformatics
• Single-cell RNA sequencing analysis

Background:

• Translating experimental findings between species like mice and humans is vital but challenging due to biological differences.
• Current methods often overlook evolutionary gene information by focusing solely on one-to-one orthologs, limiting cross-species cell type alignment.

Purpose of the Study:

• To present TACTiCS (Transfer and Align Cell Types in Cross-species analysis), a new model for accurate cell type matching and alignment across species.
• To leverage comprehensive gene evolutionary information for improved cross-species comparative analysis.

Main Methods:

• TACTiCS utilizes a natural language processing model to match genes based on protein sequences.
• A neural network classifies cell types within species, followed by transfer learning to propagate labels across species.
• The model was applied to single-cell RNA sequencing (scRNA-seq) data from the primary motor cortex of humans, mice, and marmosets.

Main Results:

• TACTiCS successfully matched and aligned cell types across human, mouse, and marmoset scRNA-seq datasets.
• The model demonstrated superior performance compared to established methods like Seurat and SAMap.
• TACTiCS's gene matching approach, utilizing protein sequences, yielded more accurate cell type correspondences than BLAST.

Conclusions:

• TACTiCS provides a robust framework for cross-species cell type alignment, enhancing the translational potential of biological research.
• The model's ability to utilize broader evolutionary gene information overcomes limitations of previous methods.
• TACTiCS offers a significant advancement in comparative single-cell transcriptomics, with publicly available code and data.