Enhancing Automatic Construction of Gene Subnetworks by Integrating Multiple Sources of Information

We present an approach to extracting information from textual documents of biological knowledge and demonstrate how cellular gene pathways may be inferred. Natural language processing techniques are used to represent title and abstract fields of publications to derive a gene similarity vectors which are subject to cluster analysis. Gene interactions are derived by parsing sentences in the abstracts to infer causal relationships. We show how high throughput transcriptome data may then be used to enhance the construction of gene pathways from information derived from text. Subnetworks constructed by integrating information automatically derived from literature with gene expression data is validated by comparing biological processes defined in the Gene Ontology 2(GO) database. We find that precision increases in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$58\%$$\end{document} of the clusters when enhanced in this manner while a decrease in precision is observed in a relatively small number of clusters. These results are compared to similar attempts at the same problem and appear to be better in terms of precision of network construction. We also show an example of a subnetwork found by this analysis that overlaps a known gene pathway in KEGG and MIPS databases.