DISCUSSION Using a Literature-based NMF Model for Discovering Gene Functional Relationships Using a Literature-based NMF Model for Discovering Gene Functional Relationships Elina Tjioe, * Michael Berry, § Ramin Homayouni, ‡ and Kevin Heinrich Θ * Genome Science and Technology Graduate School, § Department of Electrical Engineering and Computer Science, University of Tennessee, Knoxville; ‡ Bioinformatics Program, Department of Biology, University of Memphis, TN; and Θ Computable Genomix, LLC ABSTRACT The rapid growth of the biomedical literature and genomic information present a major challenge for determining the functional relationships among genes. Several bioinformatics tools have been developed to extract and identify gene relationships from various biological databases. In this study, we develop a Web-based bioinformatics tool called Feature Annotation Using Nonnegative matrix factorization (FAUN) to facilitate both the discovery and classification of functional relationships among genes. Both the computational complexity and parameterization of nonnegative matrix factorization (NMF) for processing gene sets is discussed. FAUN is first tested on a small manually constructed 50 gene collection that we, as well as others, have previously used. We then apply FAUN to analyze several microarray- derived gene sets obtained from studies of the developing cerebellum in normal and mutant mice. FAUN provides utilities for collaborative knowledge discovery and identification of new gene relationships from text streams and repositories (e.g., MEDLINE). It is particularly useful for the validation and analysis of gene associations suggested by microarray experimentation. FUTURE WORK GENE DOCUMENT TEST SET Gene List PMID Citations in Entrez Gene PubMed Titles & Abstracts Gene A Documen t Gene Document Collection Gene B Document Gene B Document Gene C Documen t Nonnegative Matrix Factorization (NMF) Feature x Gene (H) Matrix Term x Feature (W) Matrix Dominant Terms Dominant Terms Feature Annotation Dominant Genes Dominant Genes Gene-to-Gene Correlation Gene-to-Gene Correlation Term x Gene Doc Matrix General Text Parser (GTP) New Gene Document New Gene Document Feature Classifier Dominant Features Dominant Features FAUN site: Development Alzheimer Cancer 5 F: 7, 13, 16, F: 1, 3, 4, 5, 8, 9, 12, 14, 15 17, 18, F: 2, 6, 10, 11 REFERENCES 1.Heinrich, K.E., Berry, M.W., Homayouni, R. (2008) Gene Tree Labeling Using Nonnegative Matrix Factorization on Biomedical Literature. Journal of Computational Intelligence and Neuroscience, to appear. 2.Berry, M.W., Browne, M., Langville, A.N., Pauca, V.P., Plemmons, R.J. (2007) Algorithms and Applications for Approximate Nonnegative Matrix Factorization. Computational Statistics & Data Analysis 52(1): Shahnaz, F., Berry, M.W., Pauca, V.P., Plemmons, R.J. (2006) Document Clustering Using Nonnegative Matrix Factorization. Information Processing & Management 42(2), Homayouni, R., Heinrich, K., Wei, L., Berry, M.W. (2005) Gene Clustering by Latent Semantic Indexing of MEDLINE Abstracts. Bioinformatics 21(1),  Capture annotation of feature across different resolutions (ranks of nonnegative matrix factorization).  Improve FAUN model by optimizing the NMF rank; consider use of smoothness constraints.  Build a FAUN-based classifier with annotated features for the real-time classification of new gene document streams. Fig. 3. Classification of genes and features. Classification of genes in the 50-test-gene document collection based on manual examination of the biomedical literature. Classification of annotated features based on manual examination of the dominant terms in each feature. For a preliminary assessment of FAUN feature classification, each gene in the 50-test-gene collection was classified based on its most dominant annotated feature or based on some feature weight threshold. The FAUN classification using the strongest feature (per gene) yielded 90% accuracy. A FAUN-based analysis of a new cerebellum gene set has revealed new knowledge – the gene set contains a large component of transcription factors. Fig. 1. FAUN screenshot showing use of dominant terms across genes highly associated with the user-selected feature. Fig. 2. FAUN screenshot of gene-to-gene correlation and feature strength. ActualSFT3T5T6T2T1 Alzheimer Cancer Development Alz & Dev Can & Dev Total genes Table. 1. The number of genes that are classified correctly using the strongest feature (SF), feature weight threshold 3, 5, 6, 2, and 1 for T3, T5, T6, T2, and T1 respectively. The actual number of genes that are classified based on manual examination of the biomedical literature is shown in the first column.