Estimation of Pharmacological Activity by Combining Molecular Fingerprints and Skeletal Formula Images

In this paper, we propose a method to predict pharmacological activity using skeletal formula images, in addition to molecular fingerprinting. With machine learning, molecular fingerprints that express the presence of specific molecular structures are used to predict the pharmacological activity of compounds. As molecular fingerprints focus on the substructure of a compound, information on the entire molecular structure may be overlooked. Molecular graphs and molecular images can be used to represent the entire molecular structure. In the case of molecular graphs, prediction performance of machine-learning models utilizing molecular graphs is not necessarily superior to that using molecular fingerprints. For molecular images, there are none reported when machine learning is used to predict pharmacological activity utilizing skeletal formula images. Therefore, we hypothesized that the prediction performance of pharmacological activity could be improved by combining molecular fingerprints that represent partial structures and skeletal formula images that represent the entire molecular structure. Two experiments were conducted: (1) we evaluated machine-learning models utilizing skeletal formula images and (2) we integrated inference results using molecular fingerprints and skeletal formula images. In the evaluation of machine-learning models using skeletal structural formula images, the model proposed by AutoKeras showed relatively good classification performance. The integration of inference results improved recall compared with that using only molecular fingerprints. Therefore, prediction using the representation of the entire molecular structure may reduce the possibility of missing candidate compounds in the search for pharmacologically active compounds.