Domain generalization enables general cancer cell annotation in single-cell and spatial transcriptomics.

Area of Science:

• Genomics
• Computational Biology
• Oncology

Background:

• Single-cell and spatial transcriptomic sequencing are vital for cancer research.
• Accurate malignant cell annotation is critical but challenging with current methods.
• Existing algorithms struggle with accuracy and generalization across diverse cancer types.

Purpose of the Study:

• To develop a robust algorithm for rapid and accurate malignant cell identification.
• To address limitations in current cell annotation tools for pan-cancer analysis.
• To enable precise malignant cell and spot detection in transcriptomic data.

Main Methods:

• Developed Cancer-Finder, a domain generalization-based deep-learning algorithm.
• Trained and validated the algorithm on single-cell transcriptomic data.
• Adapted Cancer-Finder for spatial transcriptomic data by replacing training datasets.

Main Results:

• Cancer-Finder achieved an average accuracy of 95.16% in identifying malignant cells from single-cell data.
• The algorithm accurately identified malignant spots in spatial transcriptomic slides.
• Applied to clear cell renal cell carcinoma, it identified a 10-gene signature at the tumor-normal interface linked to patient prognosis.

Conclusions:

• Cancer-Finder is an efficient and extensible tool for malignant cell annotation.
• The algorithm demonstrates high accuracy and generalization capabilities in transcriptomic data analysis.
• Identified gene signature provides insights into tumor-normal interactions and patient outcomes.