Integration of RNA Editing with Multiomics Data Improves Machine Learning Models for Predicting Drug Responses in Breast Cancer Patients

Abstract

Background: The integration of conventional omics data such as genomics and transcriptomics data into artificial intelligence models has advanced significantly in recent years; however, their low applicability in clinical contexts, due to the high complexity of models, has been limited in their direct use inpatients. We integrated classic omics, including DNA mutation and RNA gene expression, added a novel focus on promising omics methods based on A>I(G) RNA editing, and developed a drug response prediction model. Methods: We analyzed 104 patients from the Breast Cancer Genome-Guided Therapy Study (NCT02022202). This study was used to train (70%) with 10-fold cross-validation and test (30%) the drug response classification models. We assess the performance of the random forest (RF), generalized linear model (GLM), and support vector machine (SVM) with the Caret package in classifying therapy response via various combinations of clinical data, tumoral and germline mutation data, gene expression data, and RNA editing data via the LASSO and PCA strategies. Results: First, we characterized the cohort on the basis of clinical data, mutation landscapes, differential gene expression, and RNAediting sites in 69 nonresponders and 35 responders to therapy. Second, regarding the prediction models, we demonstrated that RNA editing data improved or maintained the performance of the RF model for predicting drug response across all combinations. To select the final model, we compared the F1 score between models with different data combinations, highlighting an F1 score of 0.96 (95% CI: 0.957--0.961) and an AUC of 0.922, using LASSO for feature selection. Finally, we developed a nonresponse risk score on the basis of features that contributed to the selected model, focusing on three RNA-edited sites in the genes KDM4B, miRNA200/TTLL10-AS1, and BEST1. The score was created to facilitate the clinical translation of our findings, presenting a probability of therapy response according to RNA editing site patterns. Conclusion: Our study highlights the potential of RNA editing as a valuable addition to predictive modeling for drug response in patients with breast cancer. The nonresponse risk score could represent a tool for clinical translation, offering a probability-based assessment of therapy response. These findings suggest that incorporating RNA editing into predictive models could enhance personalized treatment strategies and improve decision-making in oncology.