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

Single-cell and spatial transcriptome sequencing, two recently optimized transcriptome sequencing methods, are increasingly used to study cancer and related diseases. Cell annotation, particularly for malignant cell annotation, is essential and crucial for in-depth analyses in these studies. However, current algorithms lack accuracy and generalization, making it difficult to consistently and rapidly infer malignant cells from pan-cancer data. To address this issue, we present Cancer-Finder, a domain generalization-based deep-learning algorithm that can rapidly identify malignant cells in single-cell data with an average accuracy of 95.16%. More importantly, by replacing the single-cell training data with spatial transcriptomic datasets, Cancer-Finder can accurately identify malignant spots on spatial slides. Applying Cancer-Finder to 5 clear cell renal cell carcinoma spatial transcriptomic samples, Cancer-Finder demonstrates a good ability to identify malignant spots and identifies a gene signature consisting of 10 genes that are significantly co-localized and enriched at the tumor-normal interface and have a strong correlation with the prognosis of clear cell renal cell carcinoma patients. In conclusion, Cancer-Finder is an efficient and extensible tool for malignant cell annotation. Efficient and accurate annotation of malignant cells is crucial for single-cell and spatial transcriptomics in cancer. Here, the authors develop Cancer-Finder, a deep-learning algorithm that can identify malignant cells in cancer single-cell and spatial transcriptomics data with speed and precision.