Selective recognition of viral RNA by capsid proteins is essential for genome packaging during the assembly of single-stranded RNA (ssRNA) viruses. However, identification of capsid protein binding sites in RNA genome remains challenging because current experimental techniques are labor-intensive and low-throughput, motivating the development of computational approaches. Here, we present a sequence-based framework that integrates RNA tertiary structural modeling, local geometric feature extraction, and machine learning to predict capsid protein binding sites. Using the Qβ bacteriophage as a proof-of-concept system, we constructed a benchmark dataset of experimentally identified binding and non-binding RNA fragments. We designed a set of geometric descriptors to characterize the local structural feature of RNA backbone. A supervised neural network trained on these features achieved an area under the receiver operating characteristic curve (AUC) of 0.88 in a 5-fold cross-validation. We further applied the trained model to local geometric features extracted from AlphaFold-predicted RNA structures. Despite substantial structural differences between predicted and experimentally determined models, the classifier retained moderate predictive performance (AUC = 0.75), indicating that approximate RNA tertiary structures preserve biologically meaningful information for capsid recognition. These results also suggest that viral genome packaging is not only governed by both intrinsic RNA structural features, but also by additional dynamic factors beyond static RNA conformations. In summary, our findings provide new mechanistic insights into RNA–capsid interactions and establish a foundation for extending this approach to diverse ssRNA viruses.