De Novo Molecule Design Through the Molecular Generative Model Conditioned by 3D Information of Protein Binding Sites

Abstract

De novo molecule design through molecular generative model is gaining increasing attention in recent years. Here a novel generative model was proposed by integrating the 3D structural information of the protein binding pocket into the conditional RNN (cRNN) model to control the generation of drug-like molecules. In this model, the composition of protein binding pocket is effectively characterized through a coarse-grain strategy and the threedimensional information of the pocket can be represented by the sorted eigenvalues of the coulomb matrix (EGCM) of the coarse-grained atoms composing the binding pocket. In current work, we used our EGCM method and a previously reported binding pocket descriptor DeeplyTough to train cRNN models and compared their performance. It has been shown that the molecules generated with the control of protein environment information have a clear tendency on generating compounds with higher similarity to the original X-ray bound ligand than normal RNN model and also achieving better performance in terms of docking scores. Our results demonstrate the potential application of EGCM controlled generative model for the targeted molecule generation and guided exploration on the drug-like chemical space.

Pocket2Drug: An Encoder-Decoder Deep Neural Network for the Target-Based Drug Design

“…Pocket2Drug can be improved by incorporating reinforcement learning imposing additional restraints on the synthetic accessibility, solubility, and toxicity of generated molecules, depending on a specific application. Additional improvements can also be achieved by applying a framework similar to the conditional recurrent neural network (cRNN), utilizing the RNN with the prior information ( Xu et al, 2021 ), to the heterogeneous input data. In contrast to cRNN, in which the pre-computed information is used as the prior condition for RNN, Pocket2Drug is an end-to-end DNN, therefore the encoder is updated during training.…”

Pocket2Mol: Efficient Molecular Sampling Based on 3D Protein Pockets

“…Skalic et al (2019) employed a modified GAN model, named BicycleGAN (Zhu et al, 2017), to represent molecules in the hidden space from protein pockets and decode the representations to SMILES strings using a captioning network. Another representative work by Xu et al (2021) designed two structure descriptors to encode the pocket and generated SMILES using conditional RNN. However, these methods merely generate 1D SMILES strings or 2D molecular graphs.…”

“…Early approaches modify the pocket-free models by integrating evaluation functions like docking scores between sampled molecules and pockets to guide the candidate searching (Li et al, 2021). Another types of models transform the 3D pocket structures to molecular SMILES strings or 2D molecular graph (Skalic et al, 2019;Xu et al, 2021) without modeling the interactions between the small molecular structures and 3D pockets explicitly. Conditional generative models are developed to model the 3D atomic density distributions within the 3D pocket structures (Masuda et al, 2020;Luo et al, 2021).…”

A pocket-based 3D molecule generative model fueled by experimental electron density

“…Therefore, researchers pursue solutions in two directions: (1) reducing the constraints by generating non-3D molecules, in which the molecules are represented as strings (i.e. SMILES) or graphs 8 – 10 ; (2) expanding the dataset, for example, by generating ligand–protein complex data using docking approaches 11 . Although these attempts are inspiring, a groundbreaking strategy that can pour additional experimental data in the models is still absent.…”