VLA-SMILES: Variable-Length-Array SMILES Descriptors in Neural Network-Based QSAR Modeling

Antonina L. Nazarova, Aiichiro Nakano

Machine learning represents a milestone in data-driven research, including material informatics, robotics, and computer-aided drug discovery. With the continuously growing virtual and synthetically available chemical space, efficient and robust quantitative structure–activity relationship (QSAR) methods are required to uncover molecules with desired properties. Herein, we propose variable-length-array SMILES-based (VLA-SMILES) structural descriptors that expand conventional SMILES descriptors widely used in machine learning. This structural representation extends the family of numerically coded SMILES, particularly binary SMILES, to expedite the discovery of new deep learning QSAR models with high predictive ability. VLA-SMILES descriptors were shown to speed up the training of QSAR models based on multilayer perceptron (MLP) with optimized backpropagation (ATransformedBP), resilient propagation (iRPROP‒), and Adam optimization learning algorithms featuring rational train–test splitting, while improving the predictive ability toward the more compute-intensive binary SMILES representation format. All the tested MLPs under the same length-array-based SMILES descriptors showed similar predictive ability and convergence rate of training in combination with the considered learning procedures. Validation with the Kennard–Stone train–test splitting based on the structural descriptor similarity metrics was found more effective than the partitioning with the ranking by activity based on biological activity values metrics for the entire set of VLA-SMILES featured QSAR. Robustness and the predictive ability of MLP models based on VLA-SMILES were assessed via the method of QSAR parametric model validation. In addition, the method of the statistical H0 hypothesis testing of the linear regression between real and observed activities based on the F2,n−2 -criteria was used for predictability estimation among VLA-SMILES featured QSAR-MLPs (with n being the volume of the testing set). Both approaches of QSAR parametric model validation and statistical hypothesis testing were found to correlate when used for the quantitative evaluation of predictabilities of the designed QSAR models with VLA-SMILES descriptors.