Graph-Based Immune Checkpoint Inhibitor Response Prediction Model for Metastatic Renal Cell Carcinoma Patients

Immune checkpoint inhibitors (ICIs) have become essential in managing metastatic Renal Cell Carcinoma (mRCC), although selecting the most suitable patients for each specific treatment remains an unmet medical need. In this work, we aimed to create and evaluate a treatment response prediction machine learning model based on tumor gene expression data of patients with mRCC. More specifically, we designed and developed a graph-based prediction model, including information about potential protein-protein interactions (PPI) between the considered expressed genes. For this work, we extracted the expression data collected in two different ICIs treatments (nivolumab and avelumab+axitinib) and organized them into two different graph datasets (considering two initial PPI networks: BioGrid and kidney-specific PPT-Ohmnet). Then, we compressed these four different graph datasets with a graph embedding technique (Graph2Vec) and evaluated the embeddings as input for a supervised classification model (Random Forest) to predict the binarized progression-free survival (PFS). Results showed that the Nivolumab and PPT-Ohmnet graph dataset obtained the highest classification performance, with an AUC of 74%. These results suggested that our pipeline could accurately model the response to the drugs, especially those modeling a short-term response (in this case, nivolumab), and that this graph-based approach benefits from tissue-specific networks (kidney-specific PPT-Ohmnet PPI). This work is presented as a data-driven solution towards improving drug selection and identifying different responses to immunotherapy treatments, including poor responses that should be offered alternative therapeutic options.