Machine learning algorithm integrates bulk and single-cell transcriptome sequencing to reveal immune-related personalized therapy prediction features for pancreatic cancer

Abstract

Abstract Background: Pancreatic cancer (PC) is a digestive malignancy with worse overall survival and we aimed to detect the TIME-related classifier to facilitate the personalized treatment of PC. Methods: Unsupervised consensus clustering and multiple machine-learning algorithms were implemented to construct the immune-related signature (IRS). scRNA-seq analysis was conducted to explore the regulatory mechanism of IRS on TIME in PC. Finally, pharmacogenomic databases were enrolled to treat high IRS patients. Results: We classified patients into Immune_rich and Immune_desert subgroups. Next, the IRS model was established based on 8 IRGs (SYT12, TNNT1, TRIM46, SMPD3, ANLN, AFF3, CXCL9 and RP1L1) and validated its predictive efficiency in multiple cohorts. RT-qPCR experiments demonstrated the differential expression of 8 IRGs between tumor and normal cell lines. Patients who gained lower IRS score tended to be more sensitive to chemotherapy and immunotherapy, and obtained better overall survival compared to those with higher IRS score. Moreover, scRNA-seq analysis revealed that fibroblast and ductal cells might affect malignant tumor cells via MIF-(CD74+CD44) and SPP1-CD44 axis. Eventually, we identified eight therapeutic targets and one agent for IRS high patients. Conclusion: Our study screened out the specific regulation pattern of TIME in PC, and shed light on the precise treatment of PC.