| 초록 |
When it comes to material design or drug design, there exist numerous chemical candidates to consider. Importantly, experimental validation for such a large volume of chemicals is prohibitive in terms of time and efforts. In addition, side effect and toxicity should also be taken into account when designing drugs. Here, a simple, but powerful machine learning method, named dendrite perceptron (DP), is proposed to predict chemical properties and relevant substructures of a chemical structure given as an input. In contrast to a deep neural network having multiple layers, DP has a single layer architecture, but generates more precise predictions for several intended applications. DP was examined on three different datasets from DrugBank, Tox21 database and dye-sensitive solar cell database (DSSCDB). As a result, the DP shows higher accuracies and predictions than existing state-of-the-art methods, and further suggested substructures that are important for the observed chemical properties. The DP is expected to help our better understanding and more efficient processing of a large number of chemical structures that are important in chemical and biological systems. |
