Deep transfer learning enables lesion tracing of circulating tumor cells

Liquid biopsy offers great promise for noninvasive cancer diagnostics, while the lack of adequate target characterization and analysis hinders its wide application. Single-cell RNA sequencing (scRNA-seq) is a powerful technology for cell characterization. Integrating scRNA-seq into a CTC-focused liquid biopsy study can perhaps classify CTCs by their original lesions. However, the lack of CTC scRNA-seq data accumulation and prior knowledge hinders further development. Therefore, we design CTC-Tracer, a transfer learning-based algorithm, to correct the distributional shift between primary cancer cells and CTCs to transfer lesion labels from the primary cancer cell atlas to CTCs. The robustness and accuracy of CTC-Tracer are validated by 8 individual standard datasets. We apply CTC-Tracer on a complex dataset consisting of RNA-seq profiles of single CTCs, CTC clusters from a BRCA patient, and two xenografts, and demonstrate that CTC-Tracer has potential in knowledge transfer between different types of RNA-seq data of lesions and CTCs. Liquid biopsy offers great promise for noninvasive cancer diagnostics, while the lack of adequate target characterization and analysis hinders its wide application. Here, the authors design a transfer learning-based algorithm to transfer lesion labels from the primary cancer cell atlas to circulating tumor cells.

Automated summary

Liquid biopsy holds great potential for noninvasive cancer diagnostics, but its wide application is hindered by the lack of adequate target characterization and analysis. In this study, a transfer learning-based algorithm is designed to transfer lesion labels from the primary cancer cell atlas to circulating tumor cells.