Self-supervised pre-training method and system for molecular attribute predictive map network

Abstract

The invention discloses a self-supervised pre-training method and system for a molecular attribute prediction graph network, functional groups in molecules are divided by using knowledge in the chemical field, a corresponding functional group tree is constructed, and then a self-supervision task generated by the functional groups is designed based on the functional group tree, so that the graph network learns the structure and semantic information of the functional groups, and better molecular expression is achieved; moreover, multi-task learning is carried out in combination with an atomic-level masking prediction task, and compared with the prior art, the accuracy of downstream molecular attribute prediction is greatly improved.

Description

technical field [0001] The invention relates to the fields of machine learning and data mining, in particular to a self-supervised pre-training method and system for molecular attribute prediction graph networks. Background technique [0002] The prediction of molecular properties is of great significance for drug synthesis and screening, such as the screening of specific drugs for the new coronavirus. The molecular properties that usually need to be predicted include the chemical energy of the molecule, drug activity and toxicity, etc. Traditional molecular property prediction methods such as density functional theory (DFT) are time-consuming and expensive, usually requiring several hours to predict the relevant properties of a molecule. At present, data-driven molecular attribute prediction methods can greatly reduce the prediction time, and a representative method is graph network (GNN). Usually, for molecular property prediction tasks, the input molecules can be modele...

AI Technical Summary

Problems solved by technology

However, the current self-supervised pre-training of the molecular attribute prediction graph network does not take into account the information of molecular functional groups, only considers the related self-supervised tasks at the molecular level or atomic level, resulting in the inability to effectively use chemical domain knowledge to help the molecular attribute prediction graph network. Self-supervised pre-training

Images