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

A multianchor transfer learning method, SemiLT, enables accurate cell label annotation in sparse scATAC‐seq data by leveraging well‐annotated scRNA‐seq references. By addressing temporal discrepancies between modalities, SemiLT improves rare cell type identification and batch correction. Applications to hematopoietic and immune datasets highlight its potential to uncover biological signals that are difficult to detect using scRNA‐seq alone. Abstract scATAC‐seq enables the detailed exploration of epigenetic variations across various cell clusters, providing complementary insights to scRNA‐seq. However, its extreme sparsity and high dimensionality pose significant challenges for cell type annotation. Transfer learning can extract key features from well‐annotated data to assist in annotating target data, thereby improving annotation accuracy. However, existing transfer learning methods overlook the temporal discrepancies between scRNA‐seq and scATAC‐seq, which exacerbate batch effects between these two modalities. Therefore, SemiLT, a multi‐anchor transfer learning method, is introduced for cell label annotation from scRNA‐seq to scATAC‐seq. Benchmarking across multiple datasets shows that SemiLT outperforms existing tools in both cell type annotation and modality batch correction. Notably, the F1 score for rare cell types improves by an average of 18%. The high‐quality annotation and embedding provided by SemiLT enhance the reliability of downstream analyses. When applied to the human bone marrow hematopoietic dataset, the trajectory transitions of hematopoietic stem cells (HSCs) are accurately reconstructed. Similarly, when applied to human peripheral blood mononuclear cell (PBMC) datasets, the key low‐abundance transcription factor (TF) KLF4 is identified in CD8 effector T cells through label transfer from scRNA‐seq to scATAC‐seq, a result that is difficult to achieve using scRNA‐seq data alone. A multianchor transfer learning method, SemiLT, enables accurate cell label annotation in sparse scATAC-seq data by leveraging well-annotated scRNA-seq references. By addressing temporal discrepancies between modalities, SemiLT improves rare cell type identification and batch correction. Applications to hematopoietic and immune datasets highlight its potential to uncover biological signals that are difficult to detect using scRNA-seq alone. Abstract scATAC-seq enables the detailed exploration of epigenetic variations across various cell clusters, providing complementary insights to scRNA-seq. However, its extreme sparsity and high dimensionality pose significant challenges for cell type annotation. Transfer learning can extract key features from well-annotated data to assist in annotating target data, thereby improving annotation accuracy. However, existing transfer learning methods overlook the temporal discrepancies between scRNA-seq and scATAC-seq, which exacerbate batch effects between these two modalities. Therefore, SemiLT, a multi-anchor transfer learning method, is introduced for cell label annotation from scRNA-seq to scATAC-seq. Benchmarking across multiple datasets shows that SemiLT outperforms existing tools in both cell type annotation and modality batch correction. Notably, the F1 score for rare cell types improves by an average of 18%. The high-quality annotation and embedding provided by SemiLT enhance the reliability of downstream analyses. When applied to the human bone marrow hematopoietic dataset, the trajectory transitions of hematopoietic stem cells (HSCs) are accurately reconstructed. Similarly, when applied to human peripheral blood mononuclear cell (PBMC) datasets, the key low-abundance transcription factor (TF) KLF4 is identified in CD8 effector T cells through label transfer from scRNA-seq to scATAC-seq, a result that is difficult to achieve using scRNA-seq data alone. Advanced Science, Volume 12, Issue 43, November 20, 2025.