Due to their recurrence and complex biology, brain tumors remain among the most challenging cancers and significantly contribute to global cancer mortality. With continuous development in precision oncology, accurately predicting patient-specific drug responses is essential for effective treatment and drug design. In this study, we propose a novel multimodal framework utilizing a Light Gradient Boosting Machine (LightGBM) regressor for brain tumors drug response prediction. The model integrates both biological and chemical data features using three modality-specific Variational Autoencoders (VAEs) to encode gene expression, gene mutation, and drug molecular fingerprint features respectively. The integrated feature representations are used by the LightGBM regressor to predict the half-maximal inhibitory concentration IC₅₀ of drugs. Reliable results are obtained using subject-level stratified nested cross-validation. Our model has yielded improved RMSE and correlation R² values 1.12 and 0.76, respectively. These results are statistically significant (p<0.05) as compared to several existing models. Furthermore, using proposed model five FDA-approved drugs with the most accurately predicted IC₅₀ values were identified. Using the statistical analysis of Glioblastoma (GBM) cell lines, we explored several over-expressed genes: EGFR, MKI67, BIRC5, TOP2A, AURKA and under-expressed genes: GFAP, MBP, NEFL, SLC1A2, PLP1., highlighting their biological roles in tumor progression and suppression. For clinical perspective, we have carried out the Survival analysis that showed that highly expressed tumor genes did not significantly affect normal patient survival (p > 0.05). It is anticipated that this study would be useful in precision oncology for anticancer drug development.