Abstract Glioblastoma (GBM) resistance to chemo-, radiation, precision, and immuno-therapies is due to the blood brain barrier, tumor heterogeneity, local and systemic immunosuppression, and diffuse infiltration. Oncolytic virotherapy uniquely overcomes these defense mechanisms. Zika virus (ZIKV) is an ideal GBM oncolytic virus candidate as it is neurotropic and elicits both anti-tumor and antiviral immune responses. While ZIKV utilizes the AXL receptor for GBM cell entry, that does not guarantee productive ZIKV infection. Identifying transcriptional features underlying ZIKV susceptibility or resistance will inform patient stratification and direct targeted oncolytic therapy development. To develop a model that predicts ZIKV response, RNA sequencing data from 10 patient-derived GBM cell lines and 5 commercially available GBM cell lines (LN229, LN18, U251, U87, T98G) were obtained, encompassing 60 samples. Each cell line was treated and classified as sensitive or resistant to ZIKV infection. We compared two feature selection approaches: unsupervised selection via principal component analysis (PCA) (Method 1) and supervised selection combining PCA with differentially expressed genes (DEGs) selected by adjusted p-value (Method 2). Random forest (RF) machine learning classifiers were evaluated using stratified, leave-one-out (LOO), leave-one-cell-line-out (LOCO), and external cross-validation (CV) strategies. LOCO-CV and external validation framework inclusion ensures predictive performance reflects generalization across biologically distinct GBM models rather than reflecting cell line-specific transcriptional noise, limiting overfitting. The two-step feature selection approach yields perfect discrimination under stratified CV and LOO-CV (AUC = 1. 00). Importantly, strong performance persists under stringent validation strategies, with LOCO-CV achieving AUCs of 0. 51 (Method 1) and 0. 97 (Method 2), and external validation achieving AUCs of 0. 66 (Method 1) and 1. 00 (Method 2). These results indicate that biologically informed feature selection enables generalizable ZIKV susceptibility prediction across heterogeneous GBM cell lines with high accuracy. DEGs driving classifier performance were enriched for neuronal lineage and glial differentiation gene sets, consistent with ZIKV neurotropism. Immune-related pathway enrichment suggests anti-viral and inflammatory signaling contribute to heterogeneous ZIKV susceptibility across GBM cell lines. Combining multi-level feature selection with strict validation frameworks supports biologically interpretable models for predicting ZIKV susceptibility. These findings highlight the importance of building precision-medicine frameworks for identifying GBM patients most likely to benefit from ZIKV therapies, improving targeted oncolytic virus treatment. Citation Format: Anna Lundeen, Steven Markwell, Dimitri Nwankwo, Parvez Akhtar, Richard Rovin. Transcriptomic prediction for evaluating Zika virus susceptibility in glioblastoma using feature selection and machine learning approaches abstract. In: Proceedings of the AACR Special Conference in Cancer Research: Brain Cancer; 2026 Mar 23-25; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Res 2026;86 (6Suppl): Abstract nr B009.
Lundeen et al. (Mon,) studied this question.