A Graph-based Molecular Structure Identification Method via Feature Extraction for Three-dimensional Electron Diffraction Data

Three-dimensional electron diffraction is a growing technique and allows researchers to obtain detailed molecular structures from their small crystals using transmission electron microscopy. The molecular structure is represented by three-dimensional atomic coordinates of the constituent atoms of a target molecule. However, a significant amount of molecular structure data is often generated from numerous diffraction images due to their relatively poorer signal-to-noise ratios. Among these data, some inaccurately represent the molecule. Consequently, researchers are required to check all the data one by one using visualization tools to identify correct structure data. It takes an enormous amount of time and is a burden for the researchers. To address this challenge, this paper proposes a method to identify candidate data that may correctly represent a target molecule. The method first extracts molecular features, such as atomic connection information, by generating graphs from three-dimensional coordinate values. Next, using the features of the graphs, distances in the feature space are calculated among all the molecular data. The distance between two molecular data represents their similarity and tends to be zero if the two data are the same from the viewpoint of the molecular structure. The proposed method then divides the molecular data into several groups where the distance between data is zero. These groups are suggested as the candidate data to researchers who judge the correctness of data. As the number of the data is greatly reduced compared to that of the original entire datasets, human efforts can also be reduced significantly. The evaluation demonstrates that the proposed method reduces the number of structure data to be checked by researchers by 97.2% on average.