Abstract
A key step in developing high-performance industrial products lies in the design of their constituent molecules. Computer-aided molecular design (CAMD) has garnered significant attention for its potential to accelerate and improve the design process. The mainstream method involves using property prediction models to predict the properties of potential molecules and selecting the best candidates based on these predictions. However, prediction errors are inevitable, introducing unreliability into the design. To address this issue, this paper proposes a novel component property modeling framework based on a molecular similarity coefficient. By calculating the similarity between a target molecule and those in an existing database, the framework selects the most similar molecules to form a tailored training dataset. The similarity coefficient also quantifies the reliability of the property predictions. In tests across various properties, this framework not only provides a quantifiable evaluation of prediction reliability but also improves the prediction accuracy of molecules with high reliability, which has the potential to enhance the integrity of molecular design.