Representing crystal structures for machine learning property prediction traditionally relies on either composition-based methods or structure-based graph neural networks (GNNs). While these methods have been successful in predicting certain properties, they fall short in accurately capturing the periodicity of crystal structures, particularly long-range information. In this work, we show that topological features derived from labeled quotient graphs (LQGs)--finite graph representations that encode bond topology without relying on real-space geometric information--can effectively predict non-local properties, i.e., properties that are not solely determined by individual local atomic environments. Using a dataset of 25,000 silica zeolite structures, we demonstrate that XGBoost models trained on LQG-derived topological features (XGB-LQG) outperform conventional GNNs (CGCNN, MEGNet) in predicting non-local properties. Furthermore, hybrid architectures that combine GNN embeddings with LQG features achieve intermediate performance, highlighting the complementary nature of geometric and topological representations. Our results establish LQGs as a powerful representation for incorporating bond topology into crystal property prediction.