The architecture consists of an encoder q ϕ and a decoder p θ, parametrized by ϕ and θ, respectively. We realized our model as a variational autoencoder (VAE), 27 with the architecture depicted in Figure 1. Also, when clear distinction is necessary, we will call a molecule a “whole-molecule” to distinguish it from a scaffold. On the other hand, the molecules that our model generates according to p ( G S ) always include S as a substructure.īefore we proceed further, we refer the reader to Table 1 for the notations we will use in what follows. In such case, the distribution can have nonzero probabilities on graphs that are not supergraphs of S. Often in other works of molecule generation 25, 20, a substructure moiety is imposed as a condition, hence defining a conditional distribution p ( G | S ) Molecular properties are introduced as a condition, by which the model can define conditional distributions p ( G S | y, y S ), where y and y S are the vectors containing the property values of a molecule and its scaffold, respectively. We also emphasize that p ( G S ) is a distribution of G alone S acts as a parametric argument, explicitly confining the domain of the distribution. Our notation here intends to manifest the particular relation, i.e., the supergraph–subgraph relation, between G and S. The underlying distribution of G can be expressed as p ( G S ). To this end, we set our generative model to be such that accepts a graph representation S of a molecular scaffold and generates a graph G that is a supergraph of S. Our purpose is to generate molecules with target properties while retaining a given scaffold as a substructure. Table 1: Notations used throughout the paper. The vector of molecular properties of a scaffold The vector of molecular properties of a whole-molecule 2 Method NotationĪn arbitrary or whole-molecule graph, depending on the context Moreover, the model can control multiple properties simultaneously with comparable performance to the single-property result. Nevertheless, our model can efficiently control molecular properties with high validity and success rate of molecule generation. However, property control becomes more challenging in scaffold-based molecule generation because fixing a scaffold confines chemical space, decreasing the possibility of finding desirable molecules. We tested whether our model can generate molecules with desirable properties under the constraint of fixing a scaffold.Ĭonditional molecule generation has been already reported in other molecular generative models. Multiple molecular properties when generating new molecules. Restraint of fixing core structures, our model could simultaneously control Mapping from scaffolds to molecules during learning. This confirms that the model can generalize the learnedĬhemical rules of adding atoms and bonds rather than simply memorizing the Validity, uniqueness, and novelty of generated molecules as high as the case Our evaluation of the model using unseen scaffolds showed the Model guarantees the generated molecules to retain the given scaffold withĬertainty. In contrast to previous related models, our Molecular graphs by extending the graph of a scaffold through sequentialĪdditions of vertices and edges. Proposes a scaffold-based molecular generative model. The way as such has called for a strategy of designing molecules retaining a Searching new molecules in areas like drug discovery often starts from theĬore structures of candidate molecules to optimize the properties of interest.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |