<p>Graph neural networks (GNNs) have played an increasingly important role in molecular property prediction. However, GNN models are prone to face the oversmoothing problem. By reviewing existing works on addressing oversmoothing, we noticed that those methods are designed for graph data and often break the original topological structure. However, it is not acceptable in molecule data which the real physicochemical meanings are lost. Motivated by this, to address the problem of oversmoothing and fill the gap in molecule property prediction, we proposed AdapGNN, a novel model-agnostic framework that designed for molecule property prediction specially. This is achieved through the integration of original node feature into the message-passing step of GNN models. Besides, to emphasize the crucial part of a molecule during predicting and further enhance the explainability of our model, we proposed a weight projection module to generate node-specific weight when merging node features. Furthermore, to validate the efficiency of our method and addressing the problem that existing benchmark data set lacks of the ground-truth of atom importance. We proposed MolExplain, a new benchmark data set for quantitative explainability evaluation in molecule property prediction. Experimental results show that the AdapGNN significantly improves the explainability of GNN models while maintaining high predictive accuracy.</p>