Drug small molecule-protein target reaction prediction method based on multi-dimensional information

Abstract

The invention discloses a drug small molecule-protein target reaction prediction method based on multi-dimensional information. The method comprises the following steps: (1), obtaining feature representation of a protein target, including an amino acid vector, a protein three-dimensional space structure feature and a protein chemical feature; (2), obtaining feature representations of the drug small molecules, including semantic features of the drug molecules, general network features of BERT and chemical features of the drug molecules; (3), fusing the features of protein targets and drug smallmolecules; (4), using the fused features as input of a classifier, and training the whole network in combination with labels in a training set. The fusion of three different types of features can greatly mine the potential features of drugs and proteins, and is convenient for more accurate DTI prediction. The method can improve the DTI prediction speed and accuracy at the same time, and is higherin practical value.

Description

technical field [0001] This application relates to artificial intelligence drug research and development, specifically to the field of drug small molecule-protein target interaction. Background technique [0002] Prediction of drug target interactions (drug target interactions, DTIs) is the basis for studying the therapeutic effect of drugs and the side effects of most drugs, and plays a vital role in the process of drug development. Although some experiments have verified that DTI (drug-target interaction) is a relatively reliable method, the combination space of drug small molecule-protein targets is too large, and it will consume huge manpower and computing resources to verify them one by one through experiments. Generally, DTI prediction methods can be divided into two categories: physics-based methods and machine learning-based methods. Among them, physics-based methods, such as molecular docking, utilize predetermined scoring functions to evaluate DTI at the atomic le...

AI Technical Summary

Problems solved by technology

Nevertheless, in the process of DTI prediction, there are still the following difficulties: 1) The size of the distance matrix of different proteins is inconsistent, since each element in the matrix represents the distance between a pair of amino acids, so different distance matrices It is unreasonable to directly down-sample and scale to the same size; 2) The SMILES sequence lengths of different drug molecules are inconsistent, and an appropriate model needs to be used to obtain semantic features; 3) The training data set for DTI tasks is relatively small, so Designing an appropriate network structure is critical

Images