As the volume of chemical structures published in the literature continues to grow at an ever-increasing pace, the need to systematically capture this data in a machine-readable manner (e.g., for the purposes of information retrieval and knowledge extraction) has never been more urgent. While there have been ongoing efforts to encourage publishers to require that chemical structures be made available in a machine-readable format (e.g., InChI, SMILES), much of the relevant chemical structures remain embedded within journal articles as digital images. There are a number of tools that can convert chemical structure images into machine-readable formats with high accuracy. However, the availability of these tools is rather limited in that they are either commercial, lack source code, or require complex package dependencies to build. Herein we report on our recent effort in developing a robust and fast Java library for converting any chemical structure image into a variety of 2D molecular vector (MOLVEC) formats (e.g., MDL, InChI with coordinates, etc.). The library is (i) compact with minimal dependencies (a self-contained executable JAR file can be as small as 10MB), (ii) relatively fast (3-4 images per second), and (iii) accurate over a wide range of image quality and resolution (achieved over 90% accuracy on the USPTO dataset). MOLVEC can be used standalone or embedded within a web service. It is also currently being distributed as part of the G-SRS substance registration software and is heavily used by FDA staff to register APIs, impurities and other chemical entities found in IND applications and other regulatory submissions. The complete source code and test benchmarks for MOLVEC are available at https://spotlite.nih.gov/ncats/molvec.

MOLVEC: Open source library for chemical structure recognition

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.