Biomedical articles share annotations with their citation neighbors

Abstract

Background: Numerous efforts have been poured into annotating the wealth of knowledge contained in biomedical articles. Thanks to such efforts, it is now possible to quantitatively explore relations between these annotations and the citation network at large scale.

Results: With the aid of several large and small annotation databases, this study shows that articles share annotations with their citation neighborhood to the point that the neighborhood's most common annotations are likely to be those appearing in the article.

Conclusions: These findings posit that an article's citation neighborhood defines to a large extent the article's annotated content. Thus, citations should be considered as a foundation for future knowledge management and annotation of biomedical articles.

Keywords: Biomedical database; Citation network; Document annotation.

Fig. 1

Recall, precision and mean average precision (MAP) for the gene2pubmed (first row), dbSNP (second row), MeSH (third row) and UniProt (fourth row) datasets for first-degree neighbors with annotations (first column), including second-degree neighbors with annotations (second column), and randomized network (third column)

Fig. 2

a Recall for BC2GN annotations based on the number of annotated neighbors. A sigmoidal function fit (R² = 0.61) with 95% confidence intervals is shown alongside (dotted lines). The sigmoidal function was chosen because it is a monotonic function that is bounded within a range. b Including second-degree neighbors (sigmoidal function fit with R² = 0.85). c For the randomized network

Fig. 3

Recall, precision and MAP for the BC2GN dataset: a for first-degree neighbors with annotations, b including second-degree neighbors with annotations, and c for the randomized network. Values shown are noisy due to the low number of samples associated to each data point

Fig. 4

Recall, precision and MAP for MeSH term overlap analysis on the a NLM2007 and b L1000 datasets using first-degree neighbors with annotations. Values are noisy for small datasets, such as these, due to the low number of samples associated to each data point. Thus, values associated to NLM2007, with 200 articles, are noisier than for L1000, with 1000 articles

Fig. 5

Mean average precision (MAP) for the gene2pubmed datasets for first-degree neighbors with annotations: a considering citing and cited references. b Considering citing references alone, and c cited references alone