RNA modifications play a crucial role in regulating gene expression and various biological processes. However, their detection remains challenging due to the limitations of experimental methods, which are often costly and time-consuming. Existing computational approaches are typically restricted to single modification types and suffer from overfitting or inadequate feature representation. To address these challenges, we propose RNA-ModCaller, a deep learning-based multi-feature fusion computational framework for efficient RNA modification prediction using only nucleotide base sequences. Our approach integrates four distinct feature extraction strategies—physicochemical, sequence composition, Word2Vec-based, and atomic-level features, combined with deep learning models, including BiGRU, BiLSTM, TransformerEncoder, and 1D-ResNet-SE. Through stacking ensemble learning, we construct a meta learner model that leverages the predictions from multiple feature-model combinations to optimize predictive performance. We evaluate RNA-ModCaller using a multi-type RNA modification dataset and demonstrate its superior performance compared to state-of-the-art methods. Additionally, RNA-ModCaller achieves higher or comparable performances in three datasets of single-type modifications, including m6A, Am, and Um. By integrating multi-perspective features with diverse deep learning architectures, RNA-ModCaller provides a robust solution for high-performance RNA modification prediction, offering a powerful computational tool for exploring biological mechanisms and advancing the study of epitranscriptomics.