A Molecular Design Approach Based on Small-Scale Datasets and Generative Models

Abstract

The invention discloses a molecular design method based on a small-scale data set and a generative model, based on an initial data set D o Build an extended dataset D a ; using the extended dataset D a Train the generative model to adjust the model parameters; introduce the information of the trained generative model to the scoring model, using the initial data set D o Train the scoring model to adjust the scoring model parameters to obtain the optimized scoring model; use the initial data set D o Adjust the parameters of the trained generation model to obtain the final generation model; use the final generation model to generate new molecular structures; use the optimized scoring model to evaluate and screen new molecular structures to obtain candidate molecules. The large data set is constructed from the initial data set without reference to additional data and does not contain symbols that do not exist in the large data set. Compared with the method of using predefined atoms or fragments to form molecules, the molecules generated by this method will have Better properties, more natural, easier to synthesize.

Description

technical field [0001] The invention relates to the technical field of computer-aided molecular design, in particular to a molecular design method based on small-scale data sets and generation models. Background technique [0002] The goal of molecular design of materials or drugs is to identify molecules with desirable properties. The molecular design task consists of two core steps: the first step, creating the molecule; and the second step, scoring and filtering the created molecule. The first step is to create a rational, efficient and novel molecular structure. In the past, there were mainly two strategies: (1) using pre-defined atoms or fragments to form molecules, the disadvantage of which is that the resulting molecules are often difficult to synthesize; (2) based on expert coding The disadvantage of regular virtual chemical reactions is that they can fail and unnecessarily limit the chemical space. [0003] In recent years, models using deep learning for molecular ...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to provide a molecular design method based on small-scale data sets and generative models, which is used to solve the problem that the lack of data sets in the prior art limits the application of molecular design methods based on deep learning in a wider and more detailed material field. Application, the amount of data is very small, it is impossible to train an effective production model, and there is a problem of separation between the generation model and the scoring model

Images