Quantitative relationship between the activity/property and the structure of compound is critical in chemical applications. To learn this quantitative relationship, hundreds of molecular descriptors have been designed to describe the structure, mainly based on the properties of vertices and edges of molecular graph. However, many descriptors degenerate to the same values for different compounds with the same molecular graph, resulting in model failure. In this paper, we design a multidimensional signal for each vertex of the molecular graph to derive new descriptors with higher discriminability. We treat the new and traditional descriptors as the signals on the descriptor graph learned from the descriptor data, and enhance descriptor dissimilarity using the Laplacian filter derived from the descriptor graph. Combining these with model learning techniques, we propose a graph signal processing based approach to obtain reliable new models for learning the quantitative relationship and predicting the properties of compounds. We also provide insights from chemistry for the boiling point model. Several experiments are presented to demonstrate the validity, effectiveness and advantages of the proposed approach.