The integration of deep neural networks with molecular representation and property investigation has been pivotal in advancing pharmaceutical research, particularly in the field of bio-property prediction and chemical reaction prediction. However, most existing methods for prediction of molecular properties focus on the view of global molecular graphs with potential unrelated to the labels, neglecting the analysis of significant ingredient patterns. In this paper, a novel model termed Structure-aware Scoring Graph Neural Network is proposed to discover prominent substructures of molecules for representation learning. Specifically, a Vital Substructure Picker is designed to identify substructural graph for predictive feature extraction and jointly generate overall sketch with prominent substructure according to their assignment scores by self-attention-based scoring layer. Furthermore, an Atomic Type-aware Classifier is adopted to maintain the robustness and effectiveness of representation learning in a self-supervised manner. In this way, the representative pattern is excavated from a fine-grained view. Extensive experiments prove that the proposed model is competitive with a significant improvement of both accuracy and stability. These results offer a fresh viewpoint on reconstructing the overall pattern by focusing on prominent local structures to obtain effective molecular representations for downstream tasks.