scTECTA: Asymmetric Deep Transfer Learning for Cross-Patient Tumor Microenvironment Single-Cell Annotation.

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Tumor microenvironment cellular heterogeneity drives cancer.
• Single-cell RNA sequencing (scRNA-seq) reveals this heterogeneity.
• Current scRNA-seq annotation methods struggle with data sparsity, biological variation, and batch effects.

Purpose of the Study:

• To develop an advanced computational method for accurate cell-type annotation in the tumor microenvironment.
• To address limitations of existing single-cell annotation techniques, particularly batch effects and data sparsity.
• To improve the analysis of cellular heterogeneity in cancer research.

Main Methods:

• Proposed scTECTA, a graph neural network (GNN)-based method utilizing transfer learning.
• Implemented graph domain adaptation with an asymmetric neural network and domain-adversarial learning.
• Employed graph convolutional networks for distribution shift correction and adversarial training for batch-effect alignment.

Main Results:

• scTECTA demonstrated superior cell-type classification performance compared to 10 benchmark methods.
• The method showed robust correction of batch effects across diverse datasets.
• Evaluated on six cancer types from 34 patients, confirming its broad applicability.

Conclusions:

• scTECTA is an efficient and powerful tool for tumor microenvironment cell-type annotation.
• The transfer learning approach effectively addresses challenges in scRNA-seq data analysis.
• This method enhances the precision and robustness of cancer cell analysis.