Compound image molecular structural formula extraction method based on adversarial learning

Abstract

The invention relates to a compound image molecular structural formula extraction method based on adversarial learning, and belongs to the field of deep learning, image recognition and compound molecular formula extraction, and the method comprises the following steps: S1, constructing a data set of data pairs composed of compound images and SMILES codes; S2, constructing an adversarial network composed of an SMILES code generator and an SMILES code determiner, and carrying out the network weight initialization; S3, alternately training the adversarial network, and performing testing; S4, inputting the compound image of which the molecular structural formula needs to be extracted into an SMILES code generator to generate an SMILES code. According to the method, the adaptivity and generalization performance of compound image feature extraction are improved, the judgment of a compound generation rule is fused, and the recognition rate, precision and robustness of molecular structural formula extraction are improved.

Description

technical field [0001] The invention relates to a method for extracting molecular structural formulas of compound images based on adversarial learning, which belongs to deep learning, image recognition and compound molecular formula extraction, and is especially suitable for extracting molecular structural formulas of compound images. Background technique [0002] In drug research and development, it is often necessary to read a large number of documents, such as articles published in various journals and magazines, patents, etc. These documents contain structural information for many compounds and are often described in the form of pictures. Although this graphical chemical structure is convenient for everyone to browse, it cannot be edited directly. Of course, we can draw in a chemical editor according to the chemical structure in the picture, but this manual extraction method is too time-consuming and labor-intensive, and it is also prone to errors, especially the large ...

AI Technical Summary

Problems solved by technology

[0003] Currently, most publications on molecular data do not provide computer-readable formats of molecular structures, such as Simplified Molecular Input Line Entry System (SMILES), Connection table, etc.

However, the existing extraction methods rely on artificially customized rules and artificially designed recognition features. These recognition rules and features can maintain a high recognition rate when extracting conventional simple compound structures, but the recognition rate and accuracy in practical application scenarios All are relatively low, for example: complex chemical structure patterns in the processed images, different styles of publications, various types of noise, low resolution that cannot meet the needs of recognition, etc.

Artificially designed rules and features are difficult to obtain high adaptability and generalization capabilities. At the same time, various rules and features also have interdependence, such as compound molecular formula segmentation, chemical bonds and chemical symbol features are interdependent, and the chemical bond segmentation effect is not good. Good usually leads to missed identification or wrong identification

Images