Automated annotation of chemical names in the literature with tunable accuracy

BACKGROUND: A significant portion of the biomedical and chemical literature refers to small molecules. The accurate identification and annotation of compound name that are relevant to the topic of the given literature can establish links between scientific publications and various chemical and life science databases. Manual annotation is the preferred method for these works because well-trained indexers can understand the paper topics as well as recognize key terms. However, considering the hundreds of thousands of new papers published annually, an automatic annotation system with high precision and relevance can be a useful complement to manual annotation. RESULTS: An automated chemical name annotation system, MeSH Automated Annotations (MAA), was developed to annotate small molecule names in scientific abstracts with tunable accuracy. This system aims to reproduce the MeSH term annotations on biomedical and chemical literature that would be created by indexers. When comparing automated free text matching to those indexed manually of 26 thousand MEDLINE abstracts, more than 40% of the annotations were false-positive (FP) cases. To reduce the FP rate, MAA incorporated several filters to remove "incorrect" annotations caused by nonspecific, partial, and low relevance chemical names. In part, relevance was measured by the position of the chemical name in the text. Tunable accuracy was obtained by adding or restricting the sections of the text scanned for chemical names. The best precision obtained was 96% with a 28% recall rate. The best performance of MAA, as measured with the F statistic was 66%, which favorably compares to other chemical name annotation systems. CONCLUSIONS: Accurate chemical name annotation can help researchers not only identify important chemical names in abstracts, but also match unindexed and unstructured abstracts to chemical records. The current work is tested against MEDLINE, but the algorithm is not specific to this corpus and it is possible that the algorithm can be applied to papers from chemical physics, material, polymer and environmental science, as well as patents, biological assay descriptions and other textual data.