Deep transfer learning in clear cell renal cell carcinoma multi-omics unveils a novel tyrosine kinase inhibitor-tolerant cell subgroup for precision therapy

Background: Most of clear cell renal cell carcinoma (ccRCC) patients exhibit no response or develop tolerance to first-line drugs, tyrosine kinase inhibitors (TKIs), highlighting the urgent need for breakthroughs in treatment. Methods: This study aimed to employ deep transfer learning to identify cells with potential tolerance to TKIs on a large single-cell atlas consisting of 41 samples and assess their cellular biological function with spatial transcriptomics and spatial metabolomics. The deconvolution methods were utilized to validate the TKI-tolerant cell proportions in real TKI response data. Subsequently, we correlated the tolerance features with patients’ clinical information to construct a patient subtype classifier and a prognostic model, and translated them into an interactive webpage. Results & Conclusion: Utilizing deep transfer learning, the study revealed that TKI-tolerant cells were poorly differentiated endothelial cells and mainly communicated with T/NK cells and fibroblasts, with a high expression of ROCK2, ASAP2, and ARHGAP18. Spatial profiling found that TKI-tolerant cells were close to epithelial and had an association with hypoxia activity and Erythronic acid. In real TKI response data validation, TKI-tolerant cells were present at higher levels in the TKI-tolerant group. (p=0.016). Multi-omics clustering based on tolerance feature molecules resulted in two subtypes, which exhibited significant differences in response to sorafenib, sunitinib, pazopanib, and axitinib. TKI- tolerant genes also showed prognostic significance, with the final model selected through 231 machine learning having strong predictive ability (average C-index: 0.761). This study finally constructed a webpage for work translation, zclab-cnp.shinyapps.io/S22_25_web/, providing valuable references for clinical diagnosis and treatment.