Deep drug molecule generation model based on skeleton structure guidance and hybrid attention mechanism

In this paper, we propose a deep generative model based on skeleton and considering stereochemical information of molecules, which characterizes molecules in the form of molecular diagrams, and generates new molecules by adding atoms and bonds sequentially to the skeleton diagrams using molecular skeleton and physicochemical properties as constraints. The model is capable of controlling multiple physicochemical properties of the molecule when generating the molecule, and has a multi-objective optimization function. In terms of model architecture: we use CVAE containing GRU to build the model, which consists of three parts: encoder (inference network), a priori network and decoder (generation network). In addition, in order to improve the efficiency and accuracy of the model, we combine a hybrid attention mechanism in the model: self-attention mechanism and external attention mechanism. The attention mechanism can automatically find the most informative feature points in the feature graph. Specifically, the attention mechanism can learn a task-oriented weighted feature graph and use it for subsequent representation learning. The validation results of the model performance show that: in terms of validity, uniqueness and novelty, the model has better validity and novelty than the pair-group model and good generalization ability; in terms of diversity, the diversity of the new molecules generated by the model is better than that of the molecules in the test set.