Molecular Property Prediction Method Based on Molecular Spatial Position Encoding Attention Neural Network Model

Abstract

The invention discloses a molecular property prediction method based on molecular spatial position encoding attention neural network model. This method encodes the spatial position in the 3D conformation of the molecule through machine learning technology, better expresses the impact of different positions in the molecule on the substructure, and better characterizes the molecule. At the same time, it uses the neural network structure and gating loop of the attention mechanism. Network (GRU) is used to predict its chemical properties. The invention makes full use of the topological relationship of the molecular substructure, conducts experiments on public data sets, effectively improves the prediction effect of molecular properties such as water solubility, toxicity and hydrophilicity, and provides a new method for prediction of drug molecules.

Description

technical field [0001] The present invention relates to the fields of cheminformatics and machine learning / artificial intelligence, and specifically relates to a theoretical method based on spatial position coding and using cheminformatics, machine learning / artificial intelligence to predict properties of unknown molecules. Background technique [0002] The drug research and development process has the characteristics of high capital density, high risk and long cycle, and requires a lot of investment in capital, manpower and material resources. In the field of drug development, although many potential drug molecules have been studied in animal models, there are still more than 30% of candidate drug molecules that fail in practical applications due to various other intrinsic properties of the molecule that do not meet the requirements. Property prediction is of great value and can be used to better predict early molecular properties, which can greatly reduce the load of late ...

AI Technical Summary

Problems solved by technology

Secondly, using atoms directly as the basic unit of the graph is not conducive to maintaining molecular chemical information. Most existing graph networks use molecules as nodes for training, which often ignores the internal information between molecular substructures. Currently, based on sequences, chemical Molecules are regarded as sentence vectors. Through methods similar to natural language processing, representations of different molecules or the same substructure of the same molecule in different positions cannot be well distinguished, which reduces the performance of the model.

Images