Method and device for classifying protein-ligand complexes

Abstract

The invention discloses a method for classifying protein-ligand complexes, which aims at negative sample data leading in the practical application of classification of the protein-ligand complexes, and introduces information of non-binding protein-ligand complexes as a negative sample, training of a residue neural network dichotomy model DeepBindBC developed and constructed based on a deep learning neural network is carried out, and classification is realized through the model. The model constructed by combining the negative sample can learn data more comprehensively, and training data distribution is closer to practical application, so that classification can be completed accurately, recognition of natural protein-small molecule complexes on real drug virtual screening is facilitated, and proceeding of early steps in drug development is facilitated. The model has sufficient depth; the types of proteins and ligands in the data are sufficient, and the reliability is high; the method comprises spatial information and can be matched with other models not based on the structure; physical and chemical key information can be effectively reserved by using a complex atom type representation method.

Description

technical field [0001] The invention relates to the technical field of protein-ligand complexes, in particular to a classification method and a classification device for protein-ligand complexes. Background technique [0002] Small-molecule drugs have become one of the important means of treating diseases by binding to disease-related proteins with high strength and specificity. Therefore, in computational molecular biology, the identification of natural protein-ligand complexes (protein-small molecule complexes) It is an important step in structure-based drug design. Some existing methods mainly rely on experimental binding data. Experimental methods to determine whether proteins interact with small molecule ligands are expensive and time-consuming. [0003] Later, people developed a series of computer-aided methods to accelerate drug screening. In particular, with the increase of experimental data of protein-small molecule complexes and the development of machine learning...

AI Technical Summary

Problems solved by technology

[0004] However, the current machine learning or deep learning models for the interaction between proteins and small molecules directly rely on the existing experimental structures and experimental combinations to train the models, often lacking negative information (lack of non-binding protein-ligand complexes, or lack of nonfunctional protein-small molecule complexes), resulting in suboptimal performance in real-world applications

Because, in real drug virtual screening, most of the complexes are inactive protein-small molecule complexes (non-binding protein-small molecule complexes), so the difference between the training data and the actual application data leads to the performance of many current methods Poor, especially small molecule combinations of binding proteins account for the vast majority of practical applications of large-scale virtual screening

Images