SemiLT: A Multianchor Transfer Learning Method for Cross-Modality Cell Label Annotation from scRNA-seq to scATAC-seq

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Epigenetics

Background:

• Single-cell ATAC sequencing (scATAC-seq) offers deep insights into epigenetic variations but faces challenges in cell type annotation due to data sparsity and high dimensionality.
• Existing transfer learning methods for scATAC-seq annotation often fail to account for temporal differences between single-cell RNA sequencing (scRNA-seq) and scATAC-seq data, leading to exacerbated batch effects.
• Accurate cell type annotation is crucial for interpreting scATAC-seq data and integrating it with other single-cell modalities.

Purpose of the Study:

• To introduce SemiLT, a multi-anchor transfer learning framework designed for robust cell label annotation from scRNA-seq to scATAC-seq data.
• To address and mitigate batch effects arising from temporal discrepancies between scRNA-seq and scATAC-seq modalities.
• To improve the accuracy of cell type annotation, particularly for rare cell populations, and enhance the reliability of downstream analyses in scATAC-seq studies.

Main Methods:

• Development of SemiLT, a novel multi-anchor transfer learning approach specifically tailored for cross-modality cell annotation (scRNA-seq to scATAC-seq).
• Benchmarking SemiLT against existing computational tools using multiple scATAC-seq and scRNA-seq datasets.
• Evaluation of SemiLT's performance in cell type annotation accuracy, rare cell type identification, modality batch correction, and downstream analysis applications like trajectory inference.

Main Results:

• SemiLT demonstrates superior performance compared to existing methods in both cell type annotation and modality batch correction for scATAC-seq data.
• Significant improvement in annotation accuracy, with an average F1 score increase of 18% for rare cell types.
• Accurate reconstruction of hematopoietic stem cell (HSC) trajectory transitions in human bone marrow data and identification of the transcription factor KLF4 in CD8 effector T cells from peripheral blood mononuclear cell (PBMC) datasets.

Conclusions:

• SemiLT effectively overcomes the limitations of existing transfer learning methods by addressing temporal discrepancies between scRNA-seq and scATAC-seq data.
• The method provides high-quality cell type annotations and embeddings, enhancing the reliability and interpretability of scATAC-seq data.
• SemiLT facilitates accurate biological discoveries, such as inferring cell differentiation trajectories and identifying key regulatory factors in specific cell populations.