MolGrow: A Graph Normalizing Flow for Hierarchical Molecular Generation

Abstract

We propose a hierarchical normalizing flow model for generating molecular graphs. The model produces new molecular structures from a single-node graph by recursively splitting every node into two. All operations are invertible and can be used as plug-and-play modules. The hierarchical nature of the latent codes allows for precise changes in the resulting graph: perturbations in the first layer cause global structural changes, while perturbations in the consequent layers change the resulting molecule only marginally. Proposed model outperforms existing generative graph models on the distribution learning task. We also show successful experiments on global and constrained optimization of chemical properties using latent codes of the model.

Application of Computational Biology and Artificial Intelligence in Drug Design

“…This approach leverages deep learning strategy to learn the probability distribution of molecular data and produces continuous or discrete latent representation for molecules with property optimization. It finally maps learned probability distribution and molecule representation into novel molecules while optimizing molecular properties through the tuning of parameters of latent codes [ 230 , 231 , 244 , 245 , 246 ]. The generative approach is effective in property-based design, LBDD, and SBDD by generating both 2D and 3D molecules [ 239 , 247 ].…”

“…The properties of chemical ligands are also optimized during generation, such as ADMET, binding affinity, logP, QED, solubility, easy to synthesize, and clearance. Properties can be optimized in two ways: one is property-based generation, wherein models would learn the chemical space of molecules with desirable properties, and then, the novel molecules are generated within a desired property space [ 245 , 257 ]. Autoencoder is a typical artificial neural network for property-based generation, which encodes molecular data along with corresponding properties into latent space.…”

“…In these methods, SMILES was firstly tokenized into a sequence of tokens, and each token was embedded into a space vector by different encoding methods such as one-hot encoding, pre-trained models [ 243 , 276 ]. Thus, grammar information of SMILES strings is learned for participating in the subsequent construction of generative models such as the transformer model, language model, variational autoencoder model, and generative adversarial networks [ 245 , 253 ]. There are recent studies used SMILES strings directly for molecule generation.…”

“…Composing invertible functions, the normalizing flow model learns a series of invertible transformations from the prior distribution of molecular data to simple distribution such as Gaussian. It finally converts the simple distribution into high-dimensional molecular data for de novo drug design [ 245 , 246 ]. Figure 7 illustrates the architecture of the normalizing flow model in molecule generation for drug design.…”

“…The hierarchical normalizing flow model, called MolGrow, generates a molecule from a single node and recursively splits every node into two; these operations are invertible and use graphical representation for node and edge attributes and feed them into an L invertible level architecture, wherein the generated latent codes fit into Gaussian distribution. Each level contains multiple blocks and linear transformations for noise separation and node merging; inside each block, three channel-wise transformations and two RealNVP layers are present [ 245 ]. GraphDF is a discrete latent variable model that applies normalizing flow for molecule generation.…”