Brand-new molecular generation method based on multi-task capsule auto-encoder neural network

Abstract

The invention discloses a brand-new molecule generation method based on a multi-task capsule auto-encoder neural network. The method comprises steps of building a brand-new molecule generation model which comprises an encoder, a multi-task capsule classifier and a decoder through an auto-encoder frame; expressing drug molecules as SMILES (simplified molecular linear input specification), marking atarget property label, and learning characteristics of known drug molecules through a training stage to obtain a training model; reconstructing molecules by utilizing the training model through a reconstruction stage; in the generation stage, molecules being generated by using a training model, the generated molecules having multiple set target properties at the same time, and meanwhile, the generated molecules having a large number of new molecules and new skeletons. The method can be used for generating various molecules such as medicines or compounds, and the characteristics and propertiesof the known medicines can be learned through one-time training, so the molecules meeting the required physical, chemical and biological properties at the same time are generated; the molecules generated by the method are higher in effectiveness and more excellent in property.

Description

technical field [0001] The invention relates to the technical field of the intersection of computer artificial intelligence and new molecular design, in particular to a new molecular generation method of a multi-task capsule self-encoder neural network, which is a new method based on the self-encoder framework and the multi-task capsule classifier framework. Molecular design methods that are suitable for generating molecules with multiple physical, chemical and biological properties simultaneously. Background technique [0002] Small molecule drug design approaches play a key role in the active drug discovery process. Traditional drug design methods such as virtual screening and pharmacophore modeling are mainly used to search known virtual compound libraries. Due to the large number of potentially synthesizable molecules in chemical space (10 23 -10 60 ) and the limitation of current computer computing performance, it is difficult to search the entire chemical space glob...

AI Technical Summary

Problems solved by technology

However, the existing molecular generation methods only consider the generation of molecules with one target property, and it is difficult to learn other characteristics other than this property, and it is impossible to simultaneously optimize multiple properties of molecules, which affects the final generation effect and cannot meet the requirements of molecular design of new drugs.

In the process of molecular generation, a key difficulty lies in the selection of classifiers. Commonly used support vector machines cannot be directly trained in conjunction with deep neural networks. Convolutional neural networks do not perform well in classification, and are difficult to apply to molecular classification of various target properties. and generate

Images