Nonnegative Tensor Factorization of Biomedical Literature for Analysis of Genomic Data

Rapid growth of the biomedical literature related to genes and molecular pathways presents a serious challenge for interpretation of genomic data. Previous work has focused on using singular value decomposition (SVD) and nonnegative matrix factorization (NMF) to extract gene relationships from Medline abstracts. However, these methods work for two dimensional data. Here, we explore the utility of nonnegative tensor factorization to extract semantic relationships between genes and the transcription factors (TFs) that regulate them, using a previously published microarray dataset. A tensor was generated for a group of 86 interferon stimulated genes, 409 TFs, and 2325 terms extracted from shared Medline abstracts. Clusters of terms, genes and TFs were evaluated at various k. For this dataset, certain genes (Il6 and Jak2) and TFs (Stat3, Stat2 and Irf3) were top ranking across most ks along with terms such as activation, interferon, cell and signaling. Further examination of several clusters, using gene pathway databases as well as natural language processing tools, revealed that nonnegative tensor factorization accurately identified genes and TFs in well established signaling pathways. For example, the method identified genes and TFs in the interferon/Toll receptor pathway with high average precision (0.695–0.938) across multiple ks. In addition, the method revealed gene-TF clusters that were not well documented, perhaps pointing to new discoveries. Taken together, this work provides proof-of-concept that nonnegative tensor factorization could be useful in interpretation of genomic data.