In molecular designs using machine learning, molecules with good predicted values are searched by a model constructed with numerical structural information of the explanatory variable X, and the physical properties and activity values of the molecule as the objective variable Y. Molecular structures are mainly searched inside the applicability domain (AD) of the model, which defines the domain of reliable X values for predicted Y values. If no molecular structure with a good predicted value of y is found, the outer edges of the AD are also searched. However, if the attempt to synthesize the proposed molecular structure fails because it is unstable, the synthesis is inefficient because of the low prediction reliability due to the outer edges of AD. Here, we develop a method to define a chemical stability domain (CSD), where only stable molecular structures exist in chemical space, as a domain where the stability of the proposed molecular structure can be identified. CSD is a domain that can be judged stable if a molecular structure of unknown stability is inside it and unstable if outside it. Application of CSD to molecular design using machine learning makes it possible to filter out unstable molecules in advance. The domain defined by the k-nearest-neighbor method from the descriptor space occupied by real molecules was able to filter out unstable molecular structures, with 99.5% of the prepared stable compounds and 8.66% of the unstable molecular structures in the domain.
Kosakai et al. (Sun,) studied this question.