Association of a deep learning system with progression-free survival (PFS) in patients with renal cell carcinoma (RCC) treated with immune checkpoint blockades (ICBs).

Abstract

e16513 Background: Anti-PD-1 based therapies have been standard-of-care options in both the front-line and treatment refractory settings in RCC, but only 20%~60% of patients respond to ICBs. Current biomarkers often fail to predict benefit of ICBs, so new reliable biomarkers are urgently warranted. Methods: Genomic and transcriptomic data were collected from 3 different ICB-treated RCC cohorts, including Miao et al. (n=53, WES), CheckMate (CM, n=261, WES; n=181, RNA-Seq) and JAVELIN (n=354, RNA-Seq) cohorts. Hallmark pathways downloaded from MSigDB database were used to screened for defective pathways that correlated with PFS using WES data of Miao et al. and CM cohorts. A defective pathway was defined as detecting at least one variant. Enrichment scores were calculated using RNA-seq data of CM cohort based on 28 immune cells gene sets provided by Charoentong et al. Immune cell signatures (ICSs) related to PFS were screened out using cox regression analysis. The selected pathways and ICSs were combined to develop a logistic regression model (10-fold cross-validation) to predict prognosis. Gene expression profiles of CM cohort were randomly split into two groups: 70% as the training set, and the remaining 30% as the test set. The predictive performance was validated in JAVELIN cohort independently. Gene sets expression levels were estimated using ssGSEA and NES methods. ROC curves were used to evaluate the predictive accuracy for 6-month PFS. Survival plots were created using the Kaplan-Meier estimator, and data were analyzed by log-rank test. Results: 7 signatures were selected as input features including 4 pathways (DNA_repair, IFN_alpha, EMT and angiogenesis) identified in Miao et al. and CM cohorts, together with 3 ICSs (Activated.CD4.T.cell, Activated.CD8.T.cell, and Immature.dendritic.cell) identified in CM cohort. Our model showed promising results for predicting 6-month PFS in CM dataset with AUC of 0.984 in the training set and 0.893 in the test set. The prediction of our model outperformed that based on TMB (AUC 0.655 vs. 0.587), GEP (AUC 0.746 vs. 0.823) and PD-L1 (AUC 0.823 vs. 0.768) both in the training and test sets, while combining our model with biomarkers above improved the prediction drastically with AUC of 0.996 in the training set and 0.935 in the test set. Meanwhile, our model maintained preferable performance for differentiating longer PFS from shorter PFS with p vale of 0.021 in JAVELIN cohort. Conclusions: We proposed a risk model to predict prognosis, in which feature selection based on hallmark pathways and immune cell subtypes. Our model showed excellent performance in predicting PFS in RCC patients undergoing immunotherapy, outperforming other ICB-related biomarkers. While combining our model with TMB, GEP, and PD-L1 was correlated more strongly with survival.