Accurate property predictions and reliability quantification in molecular design based on molecular similarity

Abstract

A crucial step in developing high-performance chemical products is the design of their constituent molecules. Computer-aided molecular design (CAMD) has gained significant attention for its potential to accelerate and enhance this design process. The typical approach involves using machine learning models trained on existing molecular databases to predict the properties of potential molecules. From these predictions, the best candidates are selected. However, prediction errors can occur, leading to unreliability in the design and limiting the effectiveness of molecular discovery. To tackle this issue, this paper presents a novel framework for modeling molecular properties based on a similarity coefficient. This framework introduces a new formula for assessing molecular similarity. By calculating the similarity between a target molecule and those in an existing database, the framework selects the most similar molecules, creating a tailored dataset for model training. This significantly enhances the accuracy of property predictions. Additionally, a quantitative reliability index is proposed based on the similarity coefficient, which allows for more informed decision-making during the molecular selection process.