2567 Background: No biomarker satisfactorily predict response to anti-PD-L1 therapies. Biomarker studies suffer from small sample size, presence of disease subtypes, and lack of simultaneous measurement of multiple biomarkers. The IMvigor210 dataset (Mariathasan et al., Nature 2018) provides baseline measurements for multiple biomarkers of response to atezolizumab (n range: 105-298) coupled with genomewide RNAseq profiles. We examined predictive performance of individual biomarkers and combined information from multiple biomarkers to measure changes in predictive performance. Methods: We built classification models (PR/CR vs. PD/SD) using genes and gene sets that provide information on pathways (mSigDB), immune components (xCell, Cibersort), and predictors of response (IMPRES, Immunophenoscore, and TIDE). Prognostic features were removed based on survival association in TCGA. All experiments were done with repeated five-fold double cross validation. Predictions from the gene sets model were used as a single biomarker. PD-L1 expression by IHC in tumor core and immune cells, tumor mutation burden(TMB), neo-antigen burden (NB), location of metastatic disease, immune phenotype and genomic subtypes were then systematically merged with the gene set based model. Results: NB was the best predictor of response (AUC 0.77), while a model combining NB, TMB, ECOG and expression signatures was marginally better (AUC 0.81) with a chance of over fitting. Chi-square tests for independence suggested that examined biomarkers do not provide independent information explaining lack of increase in AUC. Signatures for TP53 mutations, M1 macrophages, CD8+ T effector cell and DNA repair, among others, were present frequently in classification using gene expression information (AUC 0.71), suggesting their independent contributions to response. Adding gene expression information to NB didn’t improve AUC for response but provided better survival stratification. Conclusions: Integration of examined biomarkers with machine learning did not improve response prediction significantly. We are now examining sizes of subgroups defined by combination of low NB/TMB with these biomarkers.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.