Using link prediction methods to examine networks of co-occurring MESH terms in Zika and CRISPR research

Abstract

This research applied the Literature-based Discovery approach and supervised link prediction methods to predict previously unknown research links between medical subject headings (MeSH) terms in Zika and CRISPR research. Both Zika and CRISPR research was extracted from the PubMed dataset and analyzed respectively. For Zika research, the timeframe for the data extraction was between 1952 and 2017, containing 1,939 research articles and 2,546 distinct MeSH terms. For CRISPR research, the data were collected from 2002 to 2016, including 4,572 research articles and 4,203 distinct MeSH terms. The link prediction measures, Common Neighbor, Jaccard’s Coefficient, Adamic/Adr, Preferential Attachment, and Resource Allocation Index, were generated as input variables to predict whether a non-linkage between two MeSH terms is formed in the future. This research applied the Logistic Regression, Naïve Bayes, Decision Tree, and Random Forests algorithms to build classification models. Because the outcome variables are highly unbalanced, the stratified sampling and under/over-sampling methods were used to generate representative training and testing sets. The results indicate that the Logistic Regression has better performance for predicting a MeSH link in Zika research. In contrast, the Naïve Bayes has better performance for predicting a MeSH link in CRISPR research. Thus, the methods proposed by this research can be used to discover possible research areas of MeSH terms and new research directions. For biomedical policymakers, the results can be considered as an evidence-based source for the decisions of public fund allocation.