Protein-RNA interactions play critical roles in numerous biological processes such as posttranscriptional regulation and protein synthesis. However, experimental screening of protein-RNA interactions is usually laborious and time-consuming. It is therefore desirable to develop efficient bioinformatics methods to predict protein-RNA interactions, which can provide valuable hints for future experimental design and advance our understanding of the interaction mechanisms. In this study, we propose a novel method for predicting protein-RNA interactions based on both sequence and structure descriptors of protein and RNA (e.g., the sequence-based physicochemical features, the secondary and three-dimensional structure-based features). We train and compare several classifiers using these descriptors on several benchmark datasets, and the random forest method is selected to build an efficient predictor of protein-RNA interactions. We conduct further cross-validation and case studies, and the results clearly suggest the efficacy of the proposed method. A novel computational method is proposed for predicting protein-RNA interactions.The efficiency and advantage are shown in multiple benchmarks and comparison studies.Case studies in protein-miRNA/lncRNA interactions demonstrate its powerful prediction ability.