Drug molecule generation method based on regularization variation automatic encoder

Abstract

The invention discloses a drug molecule generation method based on a regularization variation automatic encoder, and the method comprises the following steps: expressing drug molecules as graph data,and establishing a drug molecule generation model comprising an encoder and a decoder by using a variation automatic encoder framework; enabling the encoder to directly encode an input drug moleculardiagram by using a diagram neural network; wherein the decoder adopts a multi-layer perceptron, the optimization target comprises reconstruction loss, KL loss and property regularization loss, and theproperty regularization loss is estimated by utilizing Monte Carlo sampling. Candidate drug molecules generated by adopting the technology disclosed by the invention are higher in effectiveness and more excellent in molecule property.

Description

technical field [0001] The present invention relates to the cross technical field of computer artificial intelligence and new drug molecule design, in particular to a drug molecule generation method based on regularized variational autoencoder, which is a method based on graph neural network, deep generation model and property target regularization The method for designing new drug molecules is applicable to the design and generation of candidate drug molecules in the process of new drug discovery. Background technique [0002] The development of new drugs is costly, takes a long time and has a low success rate. The screening of candidate drug molecules is the key link in the early stage. The introduction of computer-aided design and the latest artificial intelligence technology has greatly improved the efficiency of molecular screening. However, traditional computer screening methods mostly target existing compounds, or screen them based on structure or properties. The new...

AI Technical Summary

Problems solved by technology

However, the method based on reinforcement learning needs to introduce an additional network, which increases the difficulty of calculation and convergence, and it is difficult to design a reasonable immediate reward; the method based on Bayesian optimization due to its two-stage characteristics makes the optimization of the property target very large. The degree depends on the smoothness of the latent vector space of the model learned in the first stage

Images