In single-cell sequencing analysis, several computational methods have been developed to map the cellular state space, but little has been done to map the gene space. Here, we present a graph signal processing approach to learning rich gene representations from single-cell data using a dictionary of diffusion wavelets on the cell-cell graph. This allows for the characterization of genes based on their patterning on the cellular manifold. It also enables us to understand how localized or diffuse the expression of a gene is, for which we present a score called thegene localization score. We find that highly localized genes can be used to better characterize the cellular space, especially for trajectory-like structure. We formulate the gene embedding problem setup, design tasks with simulated single-cell data to evaluate representations, and establish eight relevant baselines. We also motivate and demonstrate the efficacy of this method for a range of biological datasets and questions, such as identifying gene coexpression modules and perturbation-specific gene-gene interactions, learning active gene signaling networks from single-cell and spatial data, and classifying therapeutic response from patient-specific gene signatures.